context: ignore zero scale LoRAs when checking sameness (#20166)

* Enhance /clear command to include system prompt

Add system prompt to messages when clearing chat history.

* Use lambda

CONTRIBUTING.md

@@ -159,7 +159,7 @@ Maintainers reserve the right to decline review or close pull requests for any r
 
 # Code maintenance
 
-- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file reponsible for:
+- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file responsible for:
   - Reviewing and merging related PRs
   - Fixing related bugs
   - Providing developer guidance/support

README.md

@@ -287,7 +287,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
-| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
+| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
 | [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
 
 ## Obtaining and quantizing models

common/arg.cpp

@@ -1279,13 +1279,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
-        {"--ctx-checkpoints", "--swa-checkpoints"}, "N",
+        {"-ctxcp", "--ctx-checkpoints", "--swa-checkpoints"}, "N",
         string_format("max number of context checkpoints to create per slot (default: %d)"
             "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_ctx_checkpoints),
         [](common_params & params, int value) {
             params.n_ctx_checkpoints = value;
         }
     ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+    add_opt(common_arg(
+        {"-cpent", "--checkpoint-every-n-tokens"}, "N",
+        string_format("create a checkpoint every n tokens during prefill (processing), -1 to disable (default: %d)", params.checkpoint_every_nt),
+        [](common_params & params, int value) {
+            params.checkpoint_every_nt = value;
+        }
+    ).set_env("LLAMA_ARG_CHECKPOINT_EVERY_NT").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-cram", "--cache-ram"}, "N",
         string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)"
@@ -2399,7 +2406,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.fit_params = false;
             } else {
                 throw std::runtime_error(
-                    string_format("error: unkown value for --fit: '%s'\n", value.c_str()));
+                    string_format("error: unknown value for --fit: '%s'\n", value.c_str()));
             }
         }
     ).set_env("LLAMA_ARG_FIT"));
@@ -2827,6 +2834,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.webui_config_json = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG_FILE"));
+    add_opt(common_arg(
+        {"--webui-mcp-proxy"},
+        {"--no-webui-mcp-proxy"},
+        string_format("experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: %s)", params.webui_mcp_proxy ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.webui_mcp_proxy = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_MCP_PROXY"));
     add_opt(common_arg(
         {"--webui"},
         {"--no-webui"},

common/common.h

@@ -516,14 +516,15 @@ struct common_params {
     std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
-    int32_t port              = 8080;         // server listens on this network port
-    int32_t timeout_read      = 600;          // http read timeout in seconds
-    int32_t timeout_write     = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
-    bool    cache_prompt      = true;         // whether to enable prompt caching
-    int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
-    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
+    int32_t port                = 8080;          // server listens on this network port
+    int32_t timeout_read        = 600;           // http read timeout in seconds
+    int32_t timeout_write       = timeout_read;  // http write timeout in seconds
+    int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
+    bool    cache_prompt        = true;  // whether to enable prompt caching
+    int32_t n_ctx_checkpoints   = 32;     // max number of context checkpoints per slot
+    int32_t checkpoint_every_nt = 8192;   // make a checkpoint every n tokens during prefill
+    int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -545,6 +546,7 @@ struct common_params {
 
     // webui configs
     bool webui = true;
+    bool webui_mcp_proxy = false;
     std::string webui_config_json;
 
     // "advanced" endpoints are disabled by default for better security
@@ -869,7 +871,7 @@ std::string common_detokenize(
 // Embedding utils
 //
 
-// TODO: repace embd_norm with an enum
+// TODO: replace embd_norm with an enum
 void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
 
 float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);

common/console.cpp

@@ -80,6 +80,8 @@ namespace console {
     static termios      initial_state;
 #endif
 
+    static completion_callback completion_cb = nullptr;
+
     //
     // Init and cleanup
     //
@@ -493,7 +495,7 @@ namespace console {
     }
 
     static void set_line_contents(std::string new_line, std::string & line, std::vector<int> & widths, size_t & char_pos,
-                                  size_t & byte_pos) {
+                                  size_t & byte_pos, int cursor_byte_pos = -1) {
         move_to_line_start(char_pos, byte_pos, widths);
         clear_current_line(widths);
 
@@ -503,6 +505,7 @@ namespace console {
         char_pos = 0;
 
         size_t idx = 0;
+        int back_width = 0;
         while (idx < line.size()) {
             size_t advance = 0;
             char32_t cp = decode_utf8(line, idx, advance);
@@ -511,8 +514,15 @@ namespace console {
             if (real_width < 0) real_width = 0;
             widths.push_back(real_width);
             idx += advance;
-            ++char_pos;
-            byte_pos = idx;
+            if (cursor_byte_pos >= 0 && static_cast<size_t>(cursor_byte_pos) < idx) {
+                back_width += real_width;
+            } else {
+                ++char_pos;
+                byte_pos = idx;
+            }
+        }
+        if (cursor_byte_pos >= 0) {
+            move_cursor(-back_width);
         }
     }
 
@@ -784,6 +794,20 @@ namespace console {
                 break;
             }
 
+            if (completion_cb && input_char == '\t') {
+                auto candidates = completion_cb(line, byte_pos);
+
+                if (!candidates.empty()) {
+                    if (candidates.size() > 1 || candidates[0].first != line) {
+                        // TODO?: Display all candidates
+                        set_line_contents(candidates[0].first, line, widths, char_pos, byte_pos, candidates[0].second);
+                    } else {
+                        // TODO: Move cursor to new byte_pos
+                    }
+                    continue;
+                }
+            }
+
             if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D */) {
                 end_of_stream = true;
                 break;
@@ -1062,6 +1086,10 @@ namespace console {
         return readline_advanced(line, multiline_input);
     }
 
+    void set_completion_callback(completion_callback cb) {
+        completion_cb = cb;
+    }
+
     namespace spinner {
         static const char LOADING_CHARS[] = {'|', '/', '-', '\\'};
         static std::condition_variable cv_stop;

common/console.h

@@ -4,7 +4,9 @@
 
 #include "common.h"
 
+#include <functional>
 #include <string>
+#include <vector>
 
 enum display_type {
     DISPLAY_TYPE_RESET = 0,
@@ -21,6 +23,9 @@ namespace console {
     void set_display(display_type display);
     bool readline(std::string & line, bool multiline_input);
 
+    using completion_callback = std::function<std::vector<std::pair<std::string, size_t>>(std::string_view, size_t)>;
+    void set_completion_callback(completion_callback cb);
+
     namespace spinner {
         void start();
         void stop();

common/debug.h

@@ -18,7 +18,7 @@ template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml
 // prints tensors that are processed in the computation graph
 // by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
 // non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
-// The template parameter determins whether an error should be thrown whenever a NaN is encountered
+// The template parameter determines whether an error should be thrown whenever a NaN is encountered
 // in a tensor (useful for stopping debug sessions on first erroneous tensor)
 // The callback data will be passed as the third parameter (user_data)
 template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);

common/jinja/README.md

@@ -63,7 +63,7 @@ The llama.cpp Jinja engine introduces `jinja::string` (see `jinja/string.h`), wh
   - **One-to-many** (e.g., split): result is marked `is_input` **only if ALL** input parts are marked `is_input`
   - **Many-to-one** (e.g., join): same as one-to-many
 
-For string concatenation, string parts will be appended to the new string as-is, while perserving the `is_input` flag.
+For string concatenation, string parts will be appended to the new string as-is, while preserving the `is_input` flag.
 
 **Enabling Input Marking:**
 

convert_hf_to_gguf.py

@@ -4031,7 +4031,7 @@ class Qwen2VLVisionModel(MmprojModel):
                 # split Conv3D into Conv2Ds
                 c1, c2, kt, kh, kw = data_torch.shape
                 del c1, c2, kh, kw  # unused
-                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+                assert kt == 2, "Current implementation only support temporal_patch_size of 2"
                 yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...])
                 yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...])
             else:
@@ -4842,12 +4842,12 @@ class _LinearAttentionVReorderBase(Qwen3NextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Qwen3_5ForConditionalGeneration")
+@ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
 class Qwen3_5TextModel(_LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35
 
 
-@ModelBase.register("Qwen3_5MoeForConditionalGeneration")
+@ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
 class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
@@ -5404,7 +5404,7 @@ class KimiLinearModel(TextModel):
         # Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
         linear_attn_config = self.hparams["linear_attn_config"]
         # n_head == 0 for KDA layers, n_head > 0 for MLA layers
-        # full_attention_layers list will be used to distingush layer type
+        # full_attention_layers list will be used to distinguish layer type
         _num_kv_heads = list()
         _full_attn_layers = linear_attn_config["full_attn_layers"]
         for il in range(self.hparams["num_hidden_layers"]):
@@ -6505,7 +6505,7 @@ class Gemma3VisionModel(MmprojModel):
         super().set_gguf_parameters()
         hparams = self.hparams
         self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GEMMA3)
-        # default values below are taken from HF tranformers code
+        # default values below are taken from HF transformers code
         self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
         self.gguf_writer.add_vision_use_gelu(True)
         # calculate proj_scale_factor (used by tinygemma3 test model)
@@ -7097,7 +7097,7 @@ class Rwkv7Model(TextModel):
 
             if bid == 0 and "time_mix_a" in new_name:
                 # dummy v0/v1/v2 on first layer
-                # easist way to make llama happy
+                # easiest way to make llama happy
                 yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch)
 
             yield (new_name, data_torch)
@@ -9596,7 +9596,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         # NOTE: Explicitly include hparam prefix prefix for d_model to
         #   disambiguate with top-level head_dim
         # NOTE 2: If needed for future models, this can be isolated in a method
-        #   to separate the prefix setting and teh keys used
+        #   to separate the prefix setting and the keys used
         self.d_model = self.find_hparam([f"{self.hparam_prefixes[0]}_head_dim", "hidden_size", "d_model"])
         self.n_group = self.find_hparam(["n_groups", "num_groups"])
         self.d_inner = self.find_hparam(["expand", "num_heads"]) * self.d_model
@@ -9743,7 +9743,7 @@ class NemotronHModel(GraniteHybridModel):
         self.gguf_writer.add_value_length(self.head_dim)
 
         # Set feed_forward_length
-        # NOTE: This will trigger an override warning. This is preferrable to
+        # NOTE: This will trigger an override warning. This is preferable to
         #   duplicating all the parent logic
         if not self.is_moe:
             n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])

docs/backend/CANN.md

@@ -20,7 +20,7 @@
 
 **Llama.cpp + CANN**
 
-The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
+The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are integrated to CANN Toolkit and kernels to using Ascend NPU directly.
 
 ## News
 
@@ -210,7 +210,7 @@ docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager
     # and install driver.
     sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
     ```
-    If the following messaage appers, firmware is installed successfully.
+    If the following message appears, firmware is installed successfully.
     ```sh
     Firmware package installed successfully!
     ```

docs/backend/SYCL.md

@@ -708,7 +708,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   - Remove **build** folder or try a clean-build.
 
-- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
+- I can **not** see `[ext_oneapi_level_zero:gpu]` after installing the GPU driver on Linux.
 
   Please double-check with `sudo sycl-ls`.
 

docs/backend/snapdragon/README.md

@@ -116,7 +116,7 @@ Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920
 ### Windows
 
 All artifacts are already installed in the `pkg-snapdragon` folder.
-To run, adapt below instructions to use Powershell scrits in `scripts/snapdragon/windows`.
+To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/windows`.
 
 ## How to Run
 

docs/backend/snapdragon/windows.md

@@ -144,7 +144,7 @@ Once the build is complete HTP ops libraries will be installed like this
 -a----         1/22/2026   6:01 PM           4139 libggml-htp.cat
 ```
 
-The .cat file, the signature and proper certicate installation can be verified with
+The .cat file, the signature and proper certificate installation can be verified with
 
 ```
 > signtool.exe verify /v /pa .\pkg-snapdragon\lib\libggml-htp.cat

docs/build.md

@@ -108,7 +108,7 @@ Building through oneAPI compilers will make avx_vnni instruction set available f
 - Using oneAPI docker image:
   If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
 
-Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
+Check [Optimizing and Running LLaMA2 on Intel® CPU](https://builders.intel.com/solutionslibrary/optimizing-and-running-llama2-on-intel-cpu) for more information.
 
 ### Other BLAS libraries
 
@@ -595,7 +595,7 @@ You can verify that KleidiAI is being used by running
 ```bash
 ./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
 ```
-If KleidiAI is enabled, the ouput will contain a line similar to:
+If KleidiAI is enabled, the output will contain a line similar to:
 ```
 load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
 ```
@@ -699,7 +699,7 @@ To read documentation for how to build on Android, [click here](./android.md)
 
 ## WebGPU [In Progress]
 
-The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `bed1a61`.
 
 In the llama.cpp directory, build with CMake:
 

docs/multimodal/MobileVLM.md

@@ -281,7 +281,7 @@ llama_print_timings:       total time =    5990.25 ms /   202 tokens
 
 Just the same as above.
 
-**ouput**
+**output**
 ```sh
 encode_image_with_clip: image embedding created: 144 tokens
 
@@ -305,7 +305,7 @@ llama_print_timings:       total time =   15513.95 ms /   412 tokens
 ## Run on Intel(R) Core(TM) Ultra7 115H
 ### operation system
 Windows11
-### comiple
+### compile
 ```sh
 make -j32
 ```

docs/ops.md

@@ -24,7 +24,7 @@ Legend:
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |

docs/ops/WebGPU.csv

@@ -9535,38 +9535,38 @@
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0,inplace=1","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0,inplace=1","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0,inplace=1","support","1","yes","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","WebGPU"

examples/batched/batched.cpp

@@ -5,6 +5,7 @@
 #include "sampling.h"
 
 #include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -16,6 +17,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.prompt = "Hello my name is";

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -5,14 +5,16 @@
 #include "common.h"
 #include "log.h"
 
+#include <algorithm>
+#include <cassert>
+#include <cinttypes>
+#include <climits>
+#include <clocale>
+#include <cstdarg>
+#include <cstring>
+#include <ctime>
 #include <unordered_map>
 #include <vector>
-#include <cassert>
-#include <climits>
-#include <cstring>
-#include <cstdarg>
-#include <cinttypes>
-#include <ctime>
 #include <random>
 #include <stdexcept>
 #include <sstream>
@@ -874,6 +876,8 @@ static std::string basename(const std::string &path) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_init();
 
     struct train_params params = get_default_train_params();

examples/debug/README.md

@@ -2,7 +2,7 @@
 
 This is a utility intended to help debug a model by registering a callback that
 logs GGML operations and tensor data. It can also store the generated logits or
-embeddings as well as the prompt and token ids for comparision with the original
+embeddings as well as the prompt and token ids for comparison with the original
 model.
 
 ### Usage

examples/deprecation-warning/deprecation-warning.cpp

@@ -1,11 +1,14 @@
 // Warns users that this filename was deprecated, and provides a link for more information.
 
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <unordered_map>
 
 // Main
 int main(int argc, char** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     std::string filename = "main";
     if (argc >= 1) {
         filename = argv[0];

examples/diffusion/README.md

@@ -43,12 +43,12 @@ Choose one of the following scheduling methods:
 - `-b`: Batch size
 
 ### Examples
-#### Dream architechture:
+#### Dream architecture:
 ```
 llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual
 ```
 
-#### LLaDA architechture:
+#### LLaDA architecture:
 ```
 llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual
 ```

examples/diffusion/diffusion-cli.cpp

@@ -7,6 +7,7 @@
 #include <limits.h>
 
 #include <algorithm>
+#include <clocale>
 #include <cmath>
 #include <cstring>
 #include <limits>
@@ -538,6 +539,8 @@ static std::string format_input_text(const std::string & prompt, const std::stri
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     ggml_time_init();
 
     common_params params;

examples/embedding/embedding.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <ctime>
 #include <algorithm>
 
@@ -94,6 +95,8 @@ static void print_raw_embeddings(const float * emb,
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {

examples/eval-callback/eval-callback.cpp

@@ -4,6 +4,8 @@
 #include "log.h"
 #include "llama.h"
 #include "llama-cpp.h"
+
+#include <clocale>
 #include <string>
 #include <vector>
 
@@ -29,6 +31,8 @@ static bool run(llama_context * ctx, const common_params & params) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     base_callback_data cb_data;
 
     common_params params;

examples/gen-docs/gen-docs.cpp

@@ -1,6 +1,7 @@
 #include "arg.h"
 #include "common.h"
 
+#include <clocale>
 #include <fstream>
 #include <sstream>
 #include <string>
@@ -100,6 +101,8 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
 }
 
 int main(int, char **) {
+    std::setlocale(LC_NUMERIC, "C");
+
     for (const auto & md : md_files) {
         std::ifstream infile(md.fname);
         if (!infile.is_open()) {

examples/gguf-hash/gguf-hash.cpp

@@ -1,13 +1,14 @@
 #include "ggml.h"
 #include "gguf.h"
 
-#include <cstdlib>   /* abort() */
+#include <algorithm>
+#include <clocale>
 #include <cstddef>
 #include <cstdio>
-#include <string>
-#include <stdexcept>
-#include <algorithm>
+#include <cstdlib>   /* abort() */
 #include <cstring>
+#include <stdexcept>
+#include <string>
 
 #include <sstream>
 #include <fstream>
@@ -626,6 +627,8 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 }
 
 int main(int argc, const char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     hash_params params;
     manifest_check_params manifest_check;
     hash_params_parse(argc, argv, params);

examples/gguf/gguf.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "gguf.h"
 
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <sstream>
@@ -240,6 +241,8 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     if (argc < 3) {
         printf("usage: %s data.gguf r|w [n]\n", argv[0]);
         printf("r: read data.gguf file\n");

examples/llama.vim

@@ -52,8 +52,8 @@ highlight llama_hl_info guifg=#77ff2f ctermfg=119
 "   n_prefix:         number of lines before the cursor location to include in the local prefix
 "   n_suffix:         number of lines after  the cursor location to include in the local suffix
 "   n_predict:        max number of tokens to predict
-"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
-"   t_max_predict_ms: max alloted time for the prediction
+"   t_max_prompt_ms:  max allotted time for the prompt processing (TODO: not yet supported)
+"   t_max_predict_ms: max allotted time for the prediction
 "   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
 "   auto_fim:         trigger FIM completion automatically on cursor movement
 "   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor

examples/lookahead/lookahead.cpp

@@ -4,10 +4,11 @@
 #include "log.h"
 #include "llama.h"
 
+#include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <algorithm>
 
 struct ngram_data {
     bool active = false;
@@ -38,6 +39,8 @@ struct ngram_container {
 };
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {

examples/lookup/lookup-create.cpp

@@ -3,10 +3,13 @@
 #include "ngram-cache.h"
 #include "llama.h"
 
+#include <clocale>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/lookup/lookup-merge.cpp

@@ -3,6 +3,7 @@
 #include "common.h"
 #include "ngram-cache.h"
 
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
@@ -17,6 +18,8 @@ static void print_usage(char* argv0) {
 }
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     if (argc < 3) {
         print_usage(argv[0]);
         exit(1);

examples/lookup/lookup-stats.cpp

@@ -5,14 +5,17 @@
 #include "llama.h"
 #include "ggml.h"
 
+#include <cinttypes>
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
-#include <cinttypes>
 #include <fstream>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/lookup/lookup.cpp

@@ -6,6 +6,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
@@ -13,6 +14,8 @@
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/model-conversion/README.md

@@ -69,7 +69,7 @@ Command line arguments take precedence over environment variables when both are
 
 In cases where the transformer implementation for the model has not been released
 yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
-will then cause the transformer implementation to be loaded explicitely and not
+will then cause the transformer implementation to be loaded explicitly and not
 use AutoModelForCausalLM:
 ```
 export UNRELEASED_MODEL_NAME=SomeNewModel
@@ -120,7 +120,7 @@ The converted model can be inspected using the following command:
 (venv) $ make causal-run-converted-model
 ```
 
-### Model logits verfication
+### Model logits verification
 The following target will run the original model and the converted model and
 compare the logits:
 ```console
@@ -235,7 +235,7 @@ new model the model can be converted to GGUF format using the following command:
 (venv) $ make embedding-run-converted-model
 ```
 
-### Model logits verfication
+### Model logits verification
 The following target will run the original model and the converted model (which
 was done manually in the previous steps) and compare the logits:
 ```console
@@ -335,7 +335,7 @@ $ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LO
 
 ## HuggingFace utilities
 The following targets are useful for creating collections and model repositories
-on Hugging Face in the the ggml-org. These can be used when preparing a relase
+on Hugging Face in the the ggml-org. These can be used when preparing a release
 to script the process for new model releases.
 
 For the following targets a `HF_TOKEN` environment variable is required.
@@ -347,7 +347,7 @@ For the following targets a `HF_TOKEN` environment variable is required.
 > $ unset HF_TOKEN
 
 ### Create a new Hugging Face Model (model repository)
-This will create a new model repsository on Hugging Face with the specified
+This will create a new model repository on Hugging Face with the specified
 model name.
 ```console
 (venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"

examples/parallel/parallel.cpp

@@ -7,12 +7,13 @@
 #include "log.h"
 #include "llama.h"
 
+#include <algorithm>
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 #include <ctime>
-#include <algorithm>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {
@@ -153,6 +154,8 @@ static std::vector<std::string> split_string(const std::string& input, char deli
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     srand(1234);
 
     common_params params;

examples/passkey/passkey.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <string>
@@ -16,6 +17,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.n_junk = 250;

examples/retrieval/retrieval.cpp

@@ -4,6 +4,7 @@
 #include "llama.h"
 
 #include <algorithm>
+#include <clocale>
 #include <fstream>
 #include <iostream> // TODO: remove me
 
@@ -112,6 +113,8 @@ static void batch_process(llama_context * ctx, llama_batch & batch, float * outp
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {

examples/save-load-state/save-load-state.cpp

@@ -2,11 +2,14 @@
 #include "common.h"
 #include "llama.h"
 
+#include <clocale>
 #include <vector>
 #include <cstdio>
 
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.prompt = "The quick brown fox";

examples/simple-chat/simple-chat.cpp

@@ -1,4 +1,5 @@
 #include "llama.h"
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <iostream>
@@ -12,6 +13,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     std::string model_path;
     int ngl = 99;
     int n_ctx = 2048;

examples/simple/simple.cpp

@@ -1,4 +1,5 @@
 #include "llama.h"
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <string>
@@ -11,6 +12,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     // path to the model gguf file
     std::string model_path;
     // prompt to generate text from

examples/speculative-simple/speculative-simple.cpp

@@ -5,12 +5,15 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {

examples/speculative/speculative.cpp

@@ -5,6 +5,7 @@
 #include "llama.h"
 
 #include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <random>
@@ -30,6 +31,8 @@ struct seq_draft {
 };
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     // needed to get candidate probs even for temp <= 0.0

examples/sycl/README.md

@@ -6,11 +6,11 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 |Tool Name| Function|Status|
 |-|-|-|
-|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
+|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, etc.|Support|
 
 ### llama-ls-sycl-device
 
-List all SYCL devices with ID, compute capability, max work group size, ect.
+List all SYCL devices with ID, compute capability, max work group size, etc.
 
 1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 

examples/sycl/ls-sycl-device.cpp

@@ -6,8 +6,10 @@
 
 
 #include "ggml-sycl.h"
+#include <clocale>
 
 int main() {
+    std::setlocale(LC_NUMERIC, "C");
     ggml_backend_sycl_print_sycl_devices();
     return 0;
 }

examples/training/finetune.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
@@ -14,6 +15,8 @@
 #endif
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
     params.escape = false;
 

ggml/include/ggml-backend.h

@@ -259,7 +259,7 @@ extern "C" {
       Example usage:
 
         // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
+        // preferably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
         sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);

ggml/include/ggml-opt.h

@@ -138,7 +138,7 @@ extern "C" {
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
 
-    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    // set gradients to zero, initialize loss, and optionally reset the optimizer
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
     GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically

ggml/include/ggml.h

@@ -2575,7 +2575,7 @@ extern "C" {
         struct ggml_tensor *  grad,
         struct ggml_tensor *  sgd_params); // alpha, weight decay
 
-    // build forward mutiple tensors and select one of them for computing
+    // build forward multiple tensors and select one of them for computing
     // this is useful for creating graphs that have constant topology but compute different things based on the input
     // ref: https://github.com/ggml-org/llama.cpp/pull/18550
     //

ggml/src/ggml-backend.cpp

@@ -1455,6 +1455,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
+        if (sched->events[split_backend_id][sched->cur_copy] == NULL) {
+            ggml_backend_synchronize(split_backend);
+        }
+
         // copy the input tensors to the split backend
         for (int input_id = 0; input_id < split->n_inputs; input_id++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1465,16 +1469,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else {
-                    ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy(input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
             } else {
                 // wait for the split backend to finish using the input before overwriting it
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
-                } else {
-                    ggml_backend_synchronize(split_backend);
                 }
 
                 // when offloading MoE weights, we can reduce the amount of data copied by copying only the experts that are used
@@ -1578,6 +1578,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             }
         }
 
+        if (sched->events[split_backend_id][sched->cur_copy] == NULL) {
+            ggml_backend_synchronize(split_backend);
+        }
+
         if (!sched->callback_eval) {
             enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
             if (ec != GGML_STATUS_SUCCESS) {

ggml/src/ggml-cpu/amx/common.h

@@ -9,6 +9,8 @@
 
 #if defined(GGML_USE_OPENMP)
 #include <omp.h>
+#else
+#include <thread>
 #endif
 
 #define TILE_M 16
@@ -56,18 +58,40 @@ inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
 }
 
 template <typename func_t>
-inline void parallel_for(int n, const func_t& f) {
+inline void parallel_for(int n, const func_t & f) {
+    if (n <= 0) {
+        return;
+    }
 #if defined(GGML_USE_OPENMP)
-#pragma omp parallel
-{
-    int nth = omp_get_num_threads();
-    int ith = omp_get_thread_num();
-    int tbegin, tend;
-    balance211(n, nth, ith, tbegin, tend);
-    f(tbegin, tend);
-}
+    #pragma omp parallel
+    {
+        int nth = omp_get_num_threads();
+        int ith = omp_get_thread_num();
+        int tbegin, tend;
+        balance211(n, nth, ith, tbegin, tend);
+        f(tbegin, tend);
+    }
 #else
-    f(0, n);
+    int nth = std::thread::hardware_concurrency();
+    if (nth <= 1) {
+        f(0, n);
+        return;
+    }
+    if (nth > n) {
+        nth = n;
+    }
+    std::vector<std::thread> threads;
+    threads.reserve(nth);
+    for (int ith = 0; ith < nth; ++ith) {
+        threads.emplace_back([&f, n, ith, nth] {
+            int tbegin, tend;
+            balance211(n, nth, ith, tbegin, tend);
+            f(tbegin, tend);
+        });
+    }
+    for (auto & t : threads) {
+        t.join();
+    }
 #endif
 }
 

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -195,7 +195,7 @@ struct tile_config_t{
 // will be needed.
 //
 // Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16;
-// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`.
+// and the single batch gemm (m=1) has a special fast path with `avx512-vnni`.
 //
 // ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/
 //    advanced-matrix-extensions-intrinsics-functions.html
@@ -1379,8 +1379,8 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         // sum of offsets, shared across COLS
         //
         // avx512-vnni does not have `_mm512_dpbssd_epi32`,
-        // need to transfrom ss to us:
-        //   a * (b - 8) is equavilent to b * a - 8 * a
+        // need to transform ss to us:
+        //   a * (b - 8) is equivalent to b * a - 8 * a
         //   s    u   u                   u   s   u   s
         //
         __m512i vcomp;

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -968,7 +968,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int vector_length = ggml_cpu_get_sve_cnt()*8;
 
-    //VLA Implemenation for SVE
+    //VLA Implementation for SVE
     switch (vector_length) {
         case 128:
             {

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -781,7 +781,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
 
                 const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
 
-                // Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
+                // Load the 64 quants from q8K duplicated to use vecdots with the interleaved columns
                 // but still need the qs to use the low and hi bits from q4
                 const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
                 int8x16_t      q8_qs[8];
@@ -3796,7 +3796,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
 
                 for (int b = 0; b < nb; b++) {
                     // bsums pairs belongs to the same q8_k subblock
-                    // 64 elemnts loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
+                    // 64 elements loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
                     const int16x8_t bsums[4]{
                         vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
                         vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -423,7 +423,7 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             quants_interleaved[j] = i0;
         }
 
-        // Masks to shuffle the quants of corresonding sub blocks for rearraning quants for vectorized bsums computation
+        // Masks to shuffle the quants of corresponding sub blocks for rearranging quants for vectorized bsums computation
         __m256i shuffle_mask_sb2 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 0, 1, 4, 5, 6, 7, 8, 9, 8, 9, 12, 13, 14, 15));
         shuffle_mask_sb2 = _mm256_permute2f128_si256(shuffle_mask_sb2, shuffle_mask_sb2, 0);
         __m256i shuffle_mask_sb3 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 0, 1, 6, 7, 8, 9, 10, 11, 8, 9, 14, 15));
@@ -625,7 +625,7 @@ static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                 iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170));
                 iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255));
 
-                // Accumulated values multipled with appropriate scales
+                // Accumulated values multiplied with appropriate scales
                 acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
             }
 
@@ -868,7 +868,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                     const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
                     const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
 
-                    // Multiply with appropiate scales and accumulate
+                    // Multiply with appropriate scales and accumulate
                     acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
                     acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
                     acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -1076,7 +1076,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                 const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
                 const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
 
-                // Multiply with appropiate scales and accumulate
+                // Multiply with appropriate scales and accumulate
                 acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
                 acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
                 acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
@@ -1257,7 +1257,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                     // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
                     const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
 
-                    // Multiply with appropiate scales and accumulate
+                    // Multiply with appropriate scales and accumulate
                     acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
                     acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
                     acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -1428,7 +1428,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                 // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
                 const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
 
-                // Multiply with appropiate scales and accumulate
+                // Multiply with appropriate scales and accumulate
                 acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
                 acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
                 acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
@@ -1612,7 +1612,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
 
                     // Dot product done within 32 bit lanes and accumulated in the same vector
-                    // First done for first sub block and thenn for second sub block in each sb
+                    // First done for first sub block and then for second sub block in each sb
                     // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
                     // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
                     // ...........................................................................
@@ -2422,7 +2422,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
                         const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
 
-                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
                         acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
                         acc_rows[rp * 4  + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
                         acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -2785,7 +2785,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
                     const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
 
-                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
                     acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
                     acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
                     acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
@@ -2802,7 +2802,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
                 }
             }
-            // Store accumlated values
+            // Store accumulated values
             for (int i = 0; i < 4; i++) {
                 _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
             }
@@ -3130,7 +3130,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
                         const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);//GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
 
-                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
                         acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
                         acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
                         acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -3460,7 +3460,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
                     const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); //GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
 
-                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
                     acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
                     acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
                     acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
@@ -4268,7 +4268,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
                         const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
 
-                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
                         acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
                         acc_rows[rp * 4  + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
                         acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -5035,7 +5035,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
                 }
             }
-            // Store accumlated values
+            // Store accumulated values
             for (int i = 0; i < 4; i++) {
                 _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
             }
@@ -5677,7 +5677,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
                         const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
 
-                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
                         acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
                         acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
                         acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
@@ -6349,7 +6349,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
                     const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
 
-                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
                     acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
                     acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
                     acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2477,7 +2477,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
 
     if (prio != GGML_SCHED_PRIO_LOW) {
         // Tell Windows that this thread should not be throttled (needs its own CPU core).
-        // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
+        // Newer Windows 11 versions aggressively park (offline) CPU cores and often place
         // all our threads onto the first 4 cores which results in terrible performance with
         // n_threads > 4
         #if _WIN32_WINNT >= 0x0602

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -533,7 +533,7 @@ class tinyBLAS {
         if constexpr (RN > 1) {
             return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
         } else {
-            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
             GGML_ASSERT(false); // we have miss something.
         }
     }
@@ -711,7 +711,7 @@ class tinyBLAS_RVV {
         if constexpr (RN > 1) {
             return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
         } else {
-            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
             GGML_ASSERT(false); // we have miss something.
         }
     }

ggml/src/ggml-cpu/ops.cpp

@@ -375,7 +375,7 @@ static void ggml_compute_forward_dup_bytes(
         const size_t rs = ne00 * type_size;
 
         if (nb00 == type_size) {
-            // src0 is contigous on first dimension, copy by rows
+            // src0 is contiguous on first dimension, copy by rows
             for (int64_t i03 = 0; i03 < ne03; i03++) {
                 for (int64_t i02 = 0; i02 < ne02; i02++) {
                     id += rs * ir0;
@@ -1795,7 +1795,7 @@ void ggml_compute_forward_repeat(
             {
                 ggml_compute_forward_repeat_f32(params, dst);
             } break;
-        // TODO: templateify the implemenation and support for I64
+        // TODO: templateify the implementation and support for I64
         //       ref https://github.com/ggml-org/llama.cpp/pull/14274#discussion_r2169492225
         //case GGML_TYPE_I64:
         //    {
@@ -2129,12 +2129,12 @@ static void ggml_compute_forward_gelu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2176,13 +2176,13 @@ static void ggml_compute_forward_gelu_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2325,12 +2325,12 @@ static void ggml_compute_forward_gelu_erf_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2372,13 +2372,13 @@ static void ggml_compute_forward_gelu_erf_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2444,12 +2444,12 @@ static void ggml_compute_forward_gelu_quick_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2491,13 +2491,13 @@ static void ggml_compute_forward_gelu_quick_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2563,12 +2563,12 @@ static void ggml_compute_forward_silu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2610,13 +2610,13 @@ static void ggml_compute_forward_silu_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2766,7 +2766,7 @@ static void ggml_compute_forward_silu_back_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2802,7 +2802,7 @@ static void ggml_compute_forward_silu_back_f16(
                 (ggml_fp16_t *) ((char *) src1->data + i1*(src1->nb[1])),
                 (ggml_fp16_t *) ((char *) grad->data + i1*(grad->nb[1])));
 
-    #ifndef NDEBUG
+#ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
@@ -2810,7 +2810,7 @@ static void ggml_compute_forward_silu_back_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-    #endif
+#endif // NDEBUG
     }
 }
 
@@ -2893,7 +2893,7 @@ static void ggml_compute_forward_reglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2953,7 +2953,7 @@ static void ggml_compute_forward_reglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3036,7 +3036,7 @@ static void ggml_compute_forward_geglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3096,7 +3096,7 @@ static void ggml_compute_forward_geglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3179,7 +3179,7 @@ static void ggml_compute_forward_swiglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3239,7 +3239,7 @@ static void ggml_compute_forward_swiglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3330,7 +3330,7 @@ static void ggml_compute_forward_swiglu_oai_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3409,7 +3409,7 @@ static void ggml_compute_forward_geglu_erf_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3469,7 +3469,7 @@ static void ggml_compute_forward_geglu_erf_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3552,7 +3552,7 @@ static void ggml_compute_forward_geglu_quick_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3612,7 +3612,7 @@ static void ggml_compute_forward_geglu_quick_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -5303,7 +5303,7 @@ static void ggml_compute_forward_soft_max_f32(
                     //printf("p[%d] = %f\n", i, p[i]);
                     assert(!isnan(wp[i]));
                 }
-#endif
+#endif // NDEBUG
 
                 float max = -INFINITY;
                 ggml_vec_max_f32(ne00, &max, wp);
@@ -5328,7 +5328,7 @@ static void ggml_compute_forward_soft_max_f32(
                     assert(!isnan(dp[i]));
                     assert(!isinf(dp[i]));
                 }
-#endif
+#endif // NDEBUG
             }
         }
     }
@@ -5402,7 +5402,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dy[i]));
             assert(!isnan(y[i]));
         }
-#endif
+#endif // NDEBUG
         // Jii = yi - yi*yi
         // Jij = -yi*yj
         // J = diag(y)-y.T*y
@@ -5435,7 +5435,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dx[i]));
             assert(!isinf(dx[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -10700,7 +10700,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         float max = -INFINITY;
         ggml_vec_max_f32(nc, &max, s0);
@@ -10719,7 +10719,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(st[i]));
             assert(!isinf(st[i]));
         }
-#endif
+#endif // NDEBUG
     }
     sums[ith] = sum_thread;
     ggml_barrier(params->threadpool);
@@ -10792,7 +10792,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         // soft_max
         float max = -INFINITY;
@@ -10810,7 +10810,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(ds0[i]));
             assert(!isinf(ds0[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 

ggml/src/ggml-cpu/repack.cpp

@@ -3032,7 +3032,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             case GGML_OP_MUL_MAT_ID:
                 {
                     size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
+                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next block.
 
                     const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
                     const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
@@ -3297,7 +3297,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         auto * wdata          = (char *)params->wdata;
         auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
 
-        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
+        // total of [n_as][ne12 + 1] elements of type mmid_row_mapping (2*int32_t = int64_t)
         auto * matrix_row_counts = (int64_t *) (wdata_src1_end);                                        // [n_as]
         struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
 

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1215,7 +1215,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     }
 
     // If attention sinks are used, potentially re-scale if KQ_max is small.
-    // Also add the sink as a value to KQ_rowsum, this is done after synchonization of KQ_rowsum
+    // Also add the sink as a value to KQ_rowsum, this is done after synchronization of KQ_rowsum
     //     so it's being done unconditionally for every thread.
     if (!is_fixup && (np == 1 || threadIdx.y % np == 0) && sinks_f) {
         float KQ_max_scale[cols_per_thread];

ggml/src/ggml-cuda/fattn-vec.cuh

@@ -10,7 +10,7 @@ static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
     return 128;
 }
 
-// Currenlty llvm with the amdgcn target does not support unrolling loops
+// Currently llvm with the amdgcn target does not support unrolling loops
 // that contain a break that can not be resolved at compile time.
 #ifdef __clang__
 #pragma clang diagnostic push

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -18,7 +18,7 @@
 #if defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
 #define GGML_USE_WMMA_FATTN
 #elif defined(RDNA4)
-#warning "rocwmma fattn is not suported on RDNA4 on rocwmma < v2.0.0, expect degraded performance"
+#warning "rocwmma fattn is not supported on RDNA4 on rocwmma < v2.0.0, expect degraded performance"
 #endif // defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
 #endif // defined(GGML_HIP_ROCWMMA_FATTN)
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2803,11 +2803,14 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+    //enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
+    bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+
+    if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
@@ -2818,14 +2821,17 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
     ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
 
-    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+    if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
+        !copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
 #endif
         return false;
     }
 
-    if (backend_src != backend_dst) {
+    if (copy_from_host) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
+    } else if (backend_src != backend_dst) {
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
             CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
@@ -3330,7 +3336,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
             return false;
         }
 
-        //rms_norm kernel assumes contigous rows
+        //rms_norm kernel assumes contiguous rows
         if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
             return false;
         }

ggml/src/ggml-cuda/quantize.cu

@@ -235,7 +235,7 @@ static __global__ void quantize_mmq_q8_1(
     q.z = roundf(xi.z*d_inv);
     q.w = roundf(xi.w*d_inv);
 
-    // Write back 4 int8 values as a single 32 bit value for better memroy bandwidth:
+    // Write back 4 int8 values as a single 32 bit value for better memory bandwidth:
     char4 * yqs4 = (char4 *) y[ib].qs;
     yqs4[iqs/4] = q;
 

ggml/src/ggml-cuda/softmax.cu

@@ -46,7 +46,7 @@ struct soft_max_params {
 };
 
 // When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
-// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
+// As we want to keep pragma unroll for all other cases we suppress the clang transformation warning here.
 #ifdef __clang__
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wpass-failed"

ggml/src/ggml-cuda/solve_tri.cu

@@ -83,7 +83,7 @@ static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
 // ======================
 // When ncols_template == 0 the bounds for the loops in this function are not
 // known and can't be unrolled. As we want to keep pragma unroll for all other
-// cases we supress the clang transformation warning here.
+// cases we suppress the clang transformation warning here.
 #ifdef __clang__
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wpass-failed"

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -139,7 +139,7 @@ struct ggml_hexagon_session {
 };
 
 void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_buffer *bufs, uint32_t n_bufs, bool sync) {
-    // Bump pending flag (cleared in the session::flush once we get the responce)
+    // Bump pending flag (cleared in the session::flush once we get the response)
     this->op_pending++;  // atomic inc
 
     int err = dspqueue_write(this->queue,
@@ -443,7 +443,7 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Repack the scales
         ggml_half * d = (ggml_half *) (y_d + i * dblk_size);
@@ -503,7 +503,7 @@ static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         const ggml_half * d = (const ggml_half *) (y_d + i * dblk_size);
@@ -552,7 +552,7 @@ static void init_row_q4x4x2(block_q4_0 * x, int64_t k) {
 
     // Init the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         x[i * 8 + 0].d = 0;
@@ -770,7 +770,7 @@ static void repack_row_q8x4x2(uint8_t * y, const block_q8_0 * x, int64_t k) {
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Repack the scales
         ggml_half * d = (ggml_half *) (y_d + i * dblk_size);
@@ -829,7 +829,7 @@ static void unpack_row_q8x4x2(block_q8_0 * x, const uint8_t * y, int64_t k) {
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         const ggml_half * d = (const ggml_half *) (y_d + i * dblk_size);
@@ -878,7 +878,7 @@ static void init_row_q8x4x2(block_q8_0 * x, int64_t k) {
 
     // Init the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q8_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         x[i * 8 + 0].d = 0;
@@ -1120,7 +1120,7 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Repack the scales
         uint8_t * e = (uint8_t *) (y_e + i * eblk_size);
@@ -1180,7 +1180,7 @@ static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k)
 
     // Repack the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4_0x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         const uint8_t * e = (const uint8_t *) (y_e + i * eblk_size);
@@ -1229,7 +1229,7 @@ static void init_row_mxfp4x4x2(block_mxfp4 * x, int64_t k) {
 
     // Init the scales
     // Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4x4x2)
-    // the last block is truncated and overriden by the scales.
+    // the last block is truncated and overridden by the scales.
     for (int i = 0; i < nb; i++) {
         // Unpack the scales
         x[i * 8 + 0].e = 0;
@@ -1865,15 +1865,26 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
     const struct ggml_tensor * src1 = op->src[1];
     const struct ggml_tensor * dst  = op;
 
-    if (src0->type != GGML_TYPE_F32) {
-        return false;
-    }
-    if (src1->type != GGML_TYPE_F32) {
-        return false;
-    }
-    if (dst->type != GGML_TYPE_F32) {
+    if (src0->type == GGML_TYPE_F32) {
+        if (src1->type != GGML_TYPE_F32) {
+            return false;
+        }
+        if (dst->type != GGML_TYPE_F32) {
+            return false;
+        }
+    }
+    else if (src0->type == GGML_TYPE_F16) {
+        if (src1->type != GGML_TYPE_F16) {
+            return false;
+        }
+        if (dst->type != GGML_TYPE_F16) {
+            return false;
+        }
+    }
+    else {
         return false;
     }
+
     if (!ggml_are_same_shape(src0, dst)) {
         return false;
     }
@@ -2670,7 +2681,7 @@ static std::vector<int> ggml_hexagon_graph_optimize_reorder(const std::vector<no
     // The main goal here is to stack the MUL_MAT ops with the same src1 input.
     // This allows use to reuse dynamically quantized src1 in VTCM.
 
-    // TODO: the current version might do incorrect reodering in cases where quantized src0
+    // TODO: the current version might do incorrect reordering in cases where quantized src0
     //       input is an output of another Op.
 
     for (int i0 = 0; i0 < n; i0++) {

ggml/src/ggml-hexagon/htp-drv.cpp

@@ -282,7 +282,7 @@ static std::string get_driver_path() {
     // Replace \SystemRoot with an absolute path from system ENV windir
     const std::wstring systemRootEnv = L"windir";
 
-    // Query the number of wide charactors this variable requires
+    // Query the number of wide characters this variable requires
     DWORD numWords = GetEnvironmentVariableW(systemRootEnv.c_str(), NULL, 0);
     if (numWords == 0) {
         GGML_LOG_ERROR("ggml-hex: Failed get systemRoot environment variable\n");

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -693,8 +693,8 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     size_t src0_row_size = src0->nb[1];
     size_t src1_row_size = src1->nb[1]; // zero bytes if src1 is not used
@@ -748,13 +748,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
     // Prepare context
     struct htp_act_context actx;
     actx.octx = octx;
 
-    actx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    actx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     actx.src0_row_size = src0_row_size;
     actx.src1_row_size = src1_row_size;
@@ -794,7 +792,7 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     actx.data_src1 = data_src1;
     actx.data_dst  = (uint8_t *) dst->data;
 
-    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_threads);
     return HTP_STATUS_OK;
 }
 

ggml/src/ggml-hexagon/htp/argsort-ops.c

@@ -241,6 +241,9 @@ int op_argsort(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
+    const uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    const uint32_t n_threads = MIN(total_rows, octx->n_threads);
+
     // Allocate scratchpad
     // We need 1 row of float + 1 row of int32 per thread.
     uint32_t ne00 = octx->src0.ne[0];
@@ -251,7 +254,7 @@ int op_argsort(struct htp_ops_context * octx) {
     // Make sure we round up to 256 for alignment requirements
     spad_per_thread = hex_round_up(spad_per_thread, 256);
 
-    size_t total_spad_size = spad_per_thread * octx->n_threads;
+    size_t total_spad_size = spad_per_thread * n_threads;
 
     if (octx->ctx->vtcm_size < total_spad_size) {
         FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
@@ -267,15 +270,12 @@ int op_argsort(struct htp_ops_context * octx) {
          octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
          octx->src0.data, octx->dst.data);
 
-    uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
-    uint32_t n_jobs = MIN(total_rows, octx->n_threads);
-
     struct htp_argsort_context actx;
     actx.octx = octx;
-    actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
+    actx.nrows_per_thread = (total_rows + n_threads - 1) / n_threads;
 
     // Run jobs
-    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/binary-ops.c

@@ -95,43 +95,87 @@ static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_
 }
 
 // Macro for scalar op switch
-#define COMPUTE_SCALAR_OP(DST, SRC, VAL, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_DIV: hvx_mul_scalar_f32_aa(DST, SRC, 1.0f / (VAL), N); break; \
-        default: break; \
+#define COMPUTE_SCALAR_OP(DST, SRC, VAL, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_DIV: hvx_mul_scalar_f32_aa(DST, SRC, 1.0f / (*(float *)VAL), N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_SUB: hvx_sub_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_MUL: hvx_mul_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_DIV: hvx_div_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (All Aligned)
-#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_aaa(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_aaa(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_aaa(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (Dst Aligned, Src0 Aligned, Src1 Unaligned)
-#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_aau(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_aau(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_aau(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (All Unaligned - generic loop used in element repeat)
-#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_uuu(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_uuu(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_uuu(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // 1. Scalar src1 (ne10 == 1)
@@ -140,6 +184,8 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -170,7 +216,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -199,13 +245,12 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
-            float val = *(float *)src1_ptr;
+            COMPUTE_SCALAR_OP(r_dst, r_src0, src1_ptr, src0_type, ne00);
             src1_ptr += s1_stride;
-            COMPUTE_SCALAR_OP(r_dst, r_src0, val, ne00);
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -216,7 +261,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
 
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -230,6 +275,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -268,8 +315,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
-        dma_queue_push(q, dma_make_ptr(s1_spad, src1_base), bctx->src1_row_size_aligned, bctx->src1_dma_stride, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
+        dma_queue_push(q, dma_make_ptr(s1_spad, src1_base), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -284,7 +331,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = s1_spad + r * bctx->src1_row_size_aligned;
             uint8_t * r_dst  = d_spad  + r * bctx->dst_row_size_aligned;
-            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint32_t i03, i02, i01, rem;
@@ -293,7 +340,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         i02 = fastdiv(rem, &bctx->dim1_div);
         i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -310,8 +357,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              uint8_t * s1_next = (uint8_t *)src1->data + p13 * nb13 + p12 * nb12 + p11 * nb11;
 
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
-             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, bctx->src1_dma_stride, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
+             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, next_block_size);
 
              ir_prefetch += next_block_size;
         }
@@ -326,6 +373,8 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -359,7 +408,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -373,7 +422,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = (uint8_t *)s1_ptr; // Constant
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
-            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint32_t i03, i02, i01, rem;
@@ -382,7 +431,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         i02 = fastdiv(rem, &bctx->dim1_div);
         i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -392,7 +441,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -406,6 +455,8 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -435,7 +486,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -462,11 +513,11 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
 
             // Read src1 from DDR (unaligned)
-            COMPUTE_VECTOR_OP_AAU(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAU(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -476,7 +527,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -490,6 +541,9 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t elem_size_bytes = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
+    const uint32_t row_size_bytes = ne00 * elem_size_bytes;;
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -519,7 +573,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -549,12 +603,12 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
             for (uint32_t c = 0; c < ne00; c += ne10) {
                 uint32_t len = MIN(ne10, ne00 - c);
                 // Use UUU for speed and simplicity
-                COMPUTE_VECTOR_OP_UUU(r_dst + c * sizeof(float), r_src0 + c * sizeof(float), r_src1_row, len);
+                COMPUTE_VECTOR_OP_UUU(r_dst + c * elem_size_bytes, r_src0 + c * elem_size_bytes, r_src1_row, src0_type, len);
             }
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -564,7 +618,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -672,18 +726,20 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
     dma_queue_flush(q);
 }
 
-static int execute_op_binary_f32(struct htp_ops_context * octx) {
+static int execute_op_binary(struct htp_ops_context * octx) {
     const struct htp_tensor * src0 = &octx->src0;
     const struct htp_tensor * src1 = &octx->src1;
     struct htp_tensor *       dst  = &octx->dst;
 
-    const uint32_t n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     // Use packed row sizes for VTCM allocation
-    const size_t src0_row_size = src0->ne[0] * sizeof(float);
-    const size_t src1_row_size = src1->ne[0] * sizeof(float);
-    const size_t dst_row_size  = dst->ne[0] * sizeof(float);
+    const uint32_t src0_type = octx->src0.type;
+    const size_t elem_size = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
+    const size_t src0_row_size = src0->ne[0] * elem_size;
+    const size_t src1_row_size = src1->ne[0] * elem_size;
+    const size_t dst_row_size  = dst->ne[0] * elem_size;
 
     // Align to VLEN
     const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
@@ -694,7 +750,7 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
     bool is_scalar = !is_add_id && (src1->ne[0] == 1);
 
     // Determine which kernel we will use to alloc memory and dispatch
-    bool use_vector_same = !is_add_id && !is_scalar && src1->ne[0] == src0->ne[0] &&
+    bool use_vector_same = !is_add_id && !is_scalar && ((src0->nb[1] % VLEN) == 0) && (src1->ne[0] == src0->ne[0]) &&
                (src1->ne[1] == src0->ne[1] || src1->ne[1] == 1) &&
                (src1->ne[2] == src0->ne[2] || src1->ne[2] == 1) &&
                (src1->ne[3] == src0->ne[3] || src1->ne[3] == 1);
@@ -726,7 +782,7 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
     }
 
     if (rows_per_buffer < 1) {
-         FARF(ERROR, "binary-f32: VTCM too small\n");
+         FARF(ERROR, "binary: VTCM too small\n");
          return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
@@ -761,16 +817,14 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
     dma_queue * q = octx->ctx->dma[0];
     if (is_row_bcast) {
-        dma_queue_push(q, dma_make_ptr(octx->src1_spad.data, (const void *) src1->data), src1_row_size_aligned, 0, src1->ne[0] * sizeof(float), 1);
+        dma_queue_push(q, dma_make_ptr(octx->src1_spad.data, (const void *) src1->data), src1_row_size_aligned, 0, src1->ne[0] * elem_size, 1);
     }
 
     struct htp_binary_context bctx;
     bctx.octx = octx;
-    bctx.nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    bctx.nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
     bctx.block_max = rows_per_buffer;
     bctx.src0_row_size_aligned = src0_row_size_aligned;
     bctx.src1_row_size_aligned = src1_row_size_aligned;
@@ -814,14 +868,24 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
         dma_queue_pop(q);
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &bctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &bctx, n_threads);
 
     return HTP_STATUS_OK;
 }
 
 int op_binary(struct htp_ops_context * octx) {
-    if (octx->src0.type == HTP_TYPE_F32) {
-        return execute_op_binary_f32(octx);
+
+    // Does not support permutations of src1
+    const struct htp_tensor * src1 = &octx->src1;
+    if (src1->nb[1] < src1->nb[0]) {
+        return HTP_STATUS_NO_SUPPORT;
     }
+
+    const uint32_t src0_type = octx->src0.type;
+    if ((src0_type == HTP_TYPE_F32) || (src0_type == HTP_TYPE_F16)) {
+        return execute_op_binary(octx);
+    }
+
     return HTP_STATUS_NO_SUPPORT;
 }
+

ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -202,6 +202,8 @@ static void cpy_work_func(unsigned int n, unsigned int i, void *data) {
 int op_cpy(struct htp_ops_context * octx) {
     cpy_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     struct htp_copy_context ct;
     ct.octx = octx;
 
@@ -227,8 +229,7 @@ int op_cpy(struct htp_ops_context * octx) {
     const bool transposed = (nb00 > nb01) || (nb0 > nb1);
     const bool sameshape  = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    ct.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
     if (sametype && sameshape) {
         ct.copy = cpy_thread_sametype_sameshape;
@@ -245,7 +246,7 @@ int op_cpy(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -10,6 +10,7 @@
 
 #include "hex-dma.h"
 #include "hvx-utils.h"
+#include "hvx-dump.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -17,6 +18,16 @@
 #include "htp-msg.h"
 #include "htp-ops.h"
 
+// Must be multiple of 32
+#define FLASH_ATTN_BLOCK_SIZE (32 * 2)
+
+// This is a bit of a hack because the compiler is strugling to properly inline
+// the default hvx_vec_f32_to_f16 with output into the local array.
+static void __attribute__((noinline)) hvx_vec_f32_to_f16_a(void *ptr, HVX_Vector v0, HVX_Vector v1)
+{
+    *(HVX_Vector *) ptr = hvx_vec_f32_to_f16(v0, v1);
+}
+
 // Dot product of two F16 vectors, accumulating to float
 static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict x, const void * restrict y, unsigned int n, float s) {
     const HVX_Vector * restrict vx = (const HVX_Vector * restrict) x; // fp16
@@ -25,175 +36,184 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16; // leftover elements
 
-    HVX_Vector rsum = Q6_V_vsplat_R(0);
+    HVX_VectorPair rsum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
 
     uint32_t i = 0;
 
     #pragma unroll(4)
     for (i = 0; i < nvec; i++) {
-        HVX_Vector y_hf = vy[i];
-        HVX_Vector x_hf = vx[i];
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
+        rsum_p = hvx_vec_mpyacc_f32_f16(rsum_p, vx[i], vy[i]);
     }
 
     if (nloe) {
-        // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
         HVX_Vector y_hf = Q6_V_vand_QV(bmask, vy[i]);
         HVX_Vector x_hf = Q6_V_vand_QV(bmask, vx[i]);
 
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
+        rsum_p = hvx_vec_mpyacc_f32_f16(rsum_p, x_hf, y_hf);
     }
 
-    rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
-    hvx_vec_store_u(r, 4, Q6_Vsf_equals_Vqf32(rsum));
+    HVX_Vector rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum_p), Q6_V_hi_W(rsum_p)));
+    rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum)));
+    hvx_vec_store_u(r, 4, rsum);
 }
 
-static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
-                                          const void * restrict y,
-                                          const void * restrict x0,
-                                          const void * restrict x1,
-                                          unsigned int n,
-                                          float        s) {
-    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x0;  // fp16
-    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
-    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp16
+static inline HVX_Vector hvx_dot_f16_f16_aa_rx4(const void * restrict y,
+                                                const uint8_t * restrict x,
+                                                const size_t stride_x,
+                                                const size_t nvec,
+                                                const size_t nloe) {
+    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x;                   // fp16
+    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) (x + stride_x);      // fp16
+    const HVX_Vector * restrict vx2 = (const HVX_Vector * restrict) (x + stride_x * 2);  // fp16
+    const HVX_Vector * restrict vx3 = (const HVX_Vector * restrict) (x + stride_x * 3);  // fp16
+    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;                   // fp16
 
-    uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16;  // leftover elements
-
-    HVX_Vector rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector rsum1 = Q6_V_vsplat_R(0);
+    HVX_VectorPair rsum0_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair rsum1_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair rsum2_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair rsum3_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
 
     uint32_t i = 0;
 
-    #pragma unroll(4)
     for (i = 0; i < nvec; i++) {
         HVX_Vector y_hf  = vy[i];
         HVX_Vector x0_hf = vx0[i];
         HVX_Vector x1_hf = vx1[i];
+        HVX_Vector x2_hf = vx2[i];
+        HVX_Vector x3_hf = vx3[i];
 
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+        rsum0_p = hvx_vec_mpyacc_f32_f16(rsum0_p, x0_hf, y_hf);
+        rsum1_p = hvx_vec_mpyacc_f32_f16(rsum1_p, x1_hf, y_hf);
+        rsum2_p = hvx_vec_mpyacc_f32_f16(rsum2_p, x2_hf, y_hf);
+        rsum3_p = hvx_vec_mpyacc_f32_f16(rsum3_p, x3_hf, y_hf);
     }
 
     if (nloe) {
         // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        HVX_Vector x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
-        HVX_Vector x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
-        HVX_Vector y_hf  = Q6_V_vand_QV(bmask, vy[i]);
+        HVX_Vector     y_hf  = Q6_V_vand_QV(bmask, vy[i]);
+        HVX_Vector     x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
+        HVX_Vector     x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+        HVX_Vector     x2_hf = Q6_V_vand_QV(bmask, vx2[i]);
+        HVX_Vector     x3_hf = Q6_V_vand_QV(bmask, vx3[i]);
 
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+        rsum0_p = hvx_vec_mpyacc_f32_f16(rsum0_p, x0_hf, y_hf);
+        rsum1_p = hvx_vec_mpyacc_f32_f16(rsum1_p, x1_hf, y_hf);
+        rsum2_p = hvx_vec_mpyacc_f32_f16(rsum2_p, x2_hf, y_hf);
+        rsum3_p = hvx_vec_mpyacc_f32_f16(rsum3_p, x3_hf, y_hf);
     }
 
-    HVX_Vector rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32x2(rsum0, rsum1));
-    hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
+    HVX_Vector rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum0_p), Q6_V_hi_W(rsum0_p)));
+    HVX_Vector rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum1_p), Q6_V_hi_W(rsum1_p)));
+    HVX_Vector rsum2 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum2_p), Q6_V_hi_W(rsum2_p)));
+    HVX_Vector rsum3 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum3_p), Q6_V_hi_W(rsum3_p)));
+
+    HVX_Vector_x4 rsum0123 = { .v = { rsum0, rsum1, rsum2, rsum3 } };
+    return hvx_vec_reduce_sum_f32x4(rsum0123);
 }
 
-// MAD: y (F32) += x (F16) * s (F32)
-static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, int n, float s) {
-    const HVX_Vector * restrict ptr_x = (const HVX_Vector *) x;
-    HVX_Vector * restrict ptr_y = (HVX_Vector *) y;
+static inline HVX_Vector hvx_dot_f16_f16_aa_rx32(const void * restrict y,
+                                                 const uint8_t * restrict x,
+                                                 const size_t stride_x,
+                                                 const size_t n,
+                                                 float        s) {
+
+    const size_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
+    const size_t nloe = n % VLEN_FP16; // leftover elements
+
+    HVX_Vector   sums;  // initialize at j = 0
+    const size_t stride_x_4 = stride_x * 4;
+    for (uint32_t j = 0; j < VLEN_FP32; j += 4) {
+        HVX_Vector     sums_x4 = hvx_dot_f16_f16_aa_rx4(y, x, stride_x, nvec, nloe);
+        HVX_VectorPred pred    = Q6_Q_vsetq_R(j * SIZEOF_FP32);
+        sums                   = Q6_V_vmux_QVV(pred, sums, sums_x4);
+        x += stride_x_4;
+    }
+
+    sums = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), sums);
+    return Q6_Vsf_equals_Vqf32(sums);
+}
+
+// MAD: y (F32) += x (F16) * s (F16)
+static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, const __fp16 * restrict s, int n) {
+    const HVX_Vector * restrict vx0 = (const HVX_Vector *) x;
+
+    HVX_VectorPair * restrict vy_p = (HVX_VectorPair *) y;
+    HVX_Vector * restrict vy = (HVX_Vector *) y;
 
     uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16; // leftover elements
 
-    HVX_Vector S = hvx_vec_splat_f16(s);
+    HVX_Vector S0 = hvx_vec_splat_f16(*s);
 
     uint32_t i = 0;
-    #pragma unroll(4)
+
+    #pragma unroll(2)
     for (i = 0; i < nvec; ++i) {
-        // Multiply x * s -> pair of F32 vectors
-        HVX_VectorPair xs_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x[i]), S);
-        ptr_y[i*2]   = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_lo_W(xs_p), ptr_y[i*2]));
-        ptr_y[i*2+1] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_hi_W(xs_p), ptr_y[i*2+1]));
+        vy_p[i] = hvx_vec_mpyacc_f32_f16(vy_p[i], Q6_Vh_vshuff_Vh(vx0[i]), S0);
     }
 
     if (nloe) {
-        HVX_VectorPair xs_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x[i]), S);
+        HVX_VectorPair xy_p = vy_p[i];
+        xy_p = hvx_vec_mpyacc_f32_f16(xy_p, Q6_Vh_vshuff_Vh(vx0[i]), S0);
 
-        HVX_Vector xs = Q6_V_lo_W(xs_p);
-        i = 2 * i; // index for ptr_y
+        HVX_Vector xy = Q6_V_lo_W(xy_p);
+        i = 2 * i;  // index for vy
 
-        if (nloe >= 32) {
-            ptr_y[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
-            nloe -= 32; ++i; xs = Q6_V_hi_W(xs_p);
+        if (nloe >= VLEN_FP32) {
+            vy[i] = xy;
+            nloe -= VLEN_FP32; ++i; xy = Q6_V_hi_W(xy_p);
         }
 
         if (nloe) {
-            HVX_Vector xy = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
-            hvx_vec_store_a(&ptr_y[i], nloe * 4, xy);
+            hvx_vec_store_a(&vy[i], nloe * 4, xy);
         }
     }
 }
 
-// MAD: y (F32) += x0 (F16) * s0 (F32) + x1 (F16) * s1 (F32)
-static inline void hvx_mad_f32_f16_aa_rx2(float * restrict y,
-                                          const void * restrict x0,
-                                          const void * restrict x1,
-                                          float s0,
-                                          float s1,
-                                          int   n) {
-    const HVX_Vector * restrict ptr_x0 = (const HVX_Vector *) x0;
-    const HVX_Vector * restrict ptr_x1 = (const HVX_Vector *) x1;
-    HVX_Vector * restrict ptr_y        = (HVX_Vector *) y;
+// MAD: y (F32) += x0 (F16) * s0 (F16) + x1 (F16) * s1 (F16)
+static inline void hvx_mad_f32_f16_aa_rx2(float * restrict y, const void * restrict x0, const void * restrict x1,
+                                          const __fp16 * restrict s0, const __fp16 * restrict s1, int n) {
+    const HVX_Vector * restrict vx0 = (const HVX_Vector *) x0;
+    const HVX_Vector * restrict vx1 = (const HVX_Vector *) x1;
+
+    HVX_VectorPair * restrict vy_p  = (HVX_VectorPair *) y;
+    HVX_Vector * restrict vy        = (HVX_Vector *) y;
 
     uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16;  // leftover elements
 
-    HVX_Vector S0 = hvx_vec_splat_f16(s0);
-    HVX_Vector S1 = hvx_vec_splat_f16(s1);
+    HVX_Vector S0 = hvx_vec_splat_f16(*s0);
+    HVX_Vector S1 = hvx_vec_splat_f16(*s1);
 
     uint32_t i = 0;
+
     #pragma unroll(2)
     for (i = 0; i < nvec; ++i) {
-        // Multiply x * s -> pair of F32 vectors
-        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
-        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
-
-        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
-        HVX_Vector xs_p_hi = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
-
-        ptr_y[i * 2]     = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_lo, ptr_y[i * 2]));
-        ptr_y[i * 2 + 1] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_hi, ptr_y[i * 2 + 1]));
+        vy_p[i] = hvx_vec_mpyacc_f32_f16(vy_p[i], Q6_Vh_vshuff_Vh(vx0[i]), S0);
+        vy_p[i] = hvx_vec_mpyacc_f32_f16(vy_p[i], Q6_Vh_vshuff_Vh(vx1[i]), S1);
     }
 
     if (nloe) {
-        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
-        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
+        HVX_VectorPair xy_p = vy_p[i];
+        xy_p = hvx_vec_mpyacc_f32_f16(xy_p, Q6_Vh_vshuff_Vh(vx0[i]), S0);
+        xy_p = hvx_vec_mpyacc_f32_f16(xy_p, Q6_Vh_vshuff_Vh(vx1[i]), S1);
 
-        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
-        HVX_Vector xs      = xs_p_lo;
-        i = 2 * i;  // index for ptr_y
+        HVX_Vector xy = Q6_V_lo_W(xy_p);
+        i = 2 * i;  // index for vy
 
-        if (nloe >= 32) {
-            ptr_y[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
-            nloe -= 32; ++i;
-            xs = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
+        if (nloe >= VLEN_FP32) {
+            vy[i] = xy;
+            nloe -= VLEN_FP32; ++i; xy = Q6_V_hi_W(xy_p);
         }
 
         if (nloe) {
-            HVX_Vector xy = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
-            hvx_vec_store_a(&ptr_y[i], nloe * 4, xy);
+            hvx_vec_store_a(&vy[i], nloe * 4, xy);
         }
     }
 }
 
-#define FLASH_ATTN_BLOCK_SIZE 128
-
 struct htp_fa_context {
     const struct htp_ops_context * octx;
 
@@ -226,7 +246,12 @@ struct htp_fa_context {
     size_t size_v_block;
     size_t size_m_block;
 
+    uint32_t qrows;
+    uint32_t qrows_per_thread;
+
     bool is_q_fp32;
+
+    uint64_t t_start;
 };
 
 static inline void hvx_scale_vec_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, HVX_Vector vs) {
@@ -296,9 +321,8 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
     const uint32_t nb3 = dst->nb[3];
 
     // total rows in q
-    const uint32_t nr = neq1*neq2*neq3;
-
-    const uint32_t dr = (nr + nth - 1) / nth;
+    const uint32_t nr = factx->qrows;
+    const uint32_t dr = factx->qrows_per_thread;
     const uint32_t ir0 = dr * ith;
     const uint32_t ir1 = MIN(ir0 + dr, nr);
 
@@ -337,15 +361,8 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
         const uint8_t * q_row_ptr = (const uint8_t *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3);
         dma_queue_push(dma, dma_make_ptr(spad_q, q_row_ptr), factx->size_q_row_padded, nbq1, size_q_row, 1);
 
-        const uint32_t h = iq2; // head index
-        const float slope = (factx->max_bias > 0.0f) ? (h < factx->n_head_log2 ? powf(factx->m0, h + 1) : powf(factx->m1, 2*(h - factx->n_head_log2) + 1)) : 1.0f;
-
-        HVX_Vector S_vec = hvx_vec_splat_f32(0.0f);
-        HVX_Vector M_vec = hvx_vec_splat_f32(-INFINITY);
-
-        // Clear accumulator
-        hvx_splat_f32_a(spad_a, 0, DV);
-        float * VKQ32 = (float *) spad_a;
+        // FARF(HIGH, "fa %u: prefetch Q: ir %u iq1 %u iq2 %u iq3 %u q_row_ptr %p size %u : usec %u", ith, ir, iq1, iq2, iq3, q_row_ptr, size_q_row,
+        //                 (unsigned)HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - factx->t_start));
 
         const __fp16 * mp_base = NULL;
         if (mask) {
@@ -376,8 +393,23 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                 // Mask is 1D contiguous for this row
                 dma_queue_push(dma, dma_make_ptr(m_dst, m_src), current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
             }
+
+            // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
+            //             ith, ir, ib, iq1, iq2, iq3,
+            //             size_k_row, size_v_row, current_block_size,
+            //             (unsigned)HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - factx->t_start));
         }
 
+        const uint32_t h = iq2; // head index
+        const float slope = (factx->max_bias > 0.0f) ? (h < factx->n_head_log2 ? powf(factx->m0, h + 1) : powf(factx->m1, 2*(h - factx->n_head_log2) + 1)) : 1.0f;
+
+        HVX_Vector S_vec = hvx_vec_splat_f32(0.0f);
+        HVX_Vector M_vec = hvx_vec_splat_f32(-INFINITY);
+
+        // Clear accumulator
+        hvx_splat_f32_a(spad_a, 0, DV);
+        float * VKQ32 = (float *) (spad_a + 0);
+
         uint8_t * q_ptr_vtcm = dma_queue_pop(dma).dst;
         if (factx->is_q_fp32) {
             hvx_copy_f16_f32_aa(q_ptr_vtcm, q_ptr_vtcm, DK);  // inplace convert f32 to f16
@@ -393,23 +425,19 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             uint8_t * v_base = dma_queue_pop(dma).dst; // V
             __fp16  * m_base = mask ? dma_queue_pop(dma).dst : NULL; // M
 
+            // FARF(HIGH, "fa %u: process: ir %u ib %u : iq1 %u iq2 %u iq3 %u q_ptr_vtcm %p : usec %u",
+            //              ith, ir, ib, iq1, iq2, iq3, q_ptr_vtcm,
+            //             (unsigned)HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - factx->t_start));
+
             // Inner loop processing the block from VTCM
             uint32_t ic = 0;
 
-            // Process in blocks of 32 (VLEN_FP32)
-            static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 <= 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
-            HVX_Vector_x4 scores_x4;
+            // Process in sub-blocks of 32 (VLEN_FP32)
+            HVX_Vector sb_scores[FLASH_ATTN_BLOCK_SIZE / VLEN_FP32];
             HVX_Vector v_max = hvx_vec_splat_f32(-INFINITY);
             for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
-                float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
-                for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
-                    const uint32_t cur_ic = ic + j;
-                    const uint8_t * k_ptr = k_base + cur_ic * factx->size_k_row_padded;
-                    hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + factx->size_k_row_padded, DK, factx->scale);
-                }
-
-                HVX_Vector scores = *(HVX_Vector *) scores_arr;
+                HVX_Vector scores = hvx_dot_f16_f16_aa_rx32(q_ptr_vtcm, k_base + ic * factx->size_k_row_padded, factx->size_k_row_padded, DK, factx->scale);
 
                 // 2. Softcap
                 if (factx->logit_softcap != 0.0f) {
@@ -428,35 +456,35 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
-                scores_x4.v[iv] = scores;
+                sb_scores[iv] = scores;
                 v_max = hvx_vec_reduce_max2_f32(scores, v_max); // All lanes have block max
             }
 
             {
                 // 4. Online Softmax Update
                 HVX_Vector M_new_vec = Q6_Vsf_vmax_VsfVsf(v_max, M_vec);
-                HVX_Vector diff_vec  = Q6_Vqf32_vsub_VsfVsf(M_vec, M_new_vec);
-                HVX_Vector ms_vec    = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(diff_vec));
+                HVX_Vector diff_vec  = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(M_vec, M_new_vec));
+                HVX_Vector ms_vec    = hvx_vec_exp_f32(diff_vec);
                 M_vec = M_new_vec;
 
                 hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
                 HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
                 for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
-                    HVX_Vector scores = scores_x4.v[iv];
+                    HVX_Vector scores = sb_scores[iv];
                     HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_vec);
                     HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
 
                     p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
 
                     // 5. Accumulate V
-                    float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
-                    *(HVX_Vector *) p_arr = P;
+                    __fp16 __attribute__((aligned(VLEN))) p_arr[VLEN_FP16];
+                    hvx_vec_f32_to_f16_a(p_arr, P, hvx_vec_splat_f32(0));
 
                     for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
                         const uint32_t  cur_ic = ic2 + j;
                         const uint8_t * v_ptr  = v_base + cur_ic * factx->size_v_row_padded;
-                        hvx_mad_f32_f16_aa_rx2(VKQ32, v_ptr, v_ptr + factx->size_v_row_padded, p_arr[j], p_arr[j + 1], DV);
+                        hvx_mad_f32_f16_aa_rx2(VKQ32, v_ptr, v_ptr + factx->size_v_row_padded, (p_arr + j), (p_arr + j + 1), DV);
                     }
                 }
 
@@ -464,47 +492,50 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                 S_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(S_vec, ms_vec)), p_sum_vec));
             }
 
-            // Sync scalars for leftover/next block if needed
-            float M = hvx_vec_get_f32(M_vec);
-            float S = hvx_vec_get_f32(S_vec);
+            if (ic < current_block_size) {
+                // Sync scalars for leftover/next block if needed
+                float M = hvx_vec_get_f32(M_vec);
+                float S = hvx_vec_get_f32(S_vec);
 
-            // Leftover
-            for (; ic < current_block_size; ++ic) {
-                float s_val;
-                const uint8_t * k_ptr = k_base + ic * factx->size_k_row_padded;
-                hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, factx->scale);
-                if (factx->logit_softcap != 0.0f) {
-                    s_val = factx->logit_softcap * tanhf(s_val);
+                // Leftover
+                for (; ic < current_block_size; ++ic) {
+                    float s_val;
+                    const uint8_t * k_ptr = k_base + ic * factx->size_k_row_padded;
+                    hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, factx->scale);
+                    if (factx->logit_softcap != 0.0f) {
+                        s_val = factx->logit_softcap * tanhf(s_val);
+                    }
+
+                    if (mask) {
+                        const float m_val = m_base[ic];
+                        s_val += slope * m_val;
+                    }
+
+                    const float Mold = M;
+                    __fp16 vs = 1.0f;
+
+                    if (s_val > M) {
+                        M = s_val;
+                        HVX_Vector diff_vec = hvx_vec_splat_f32(Mold - M);
+                        HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
+                        hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
+
+                        float ms = hvx_vec_get_f32(ms_vec);
+                        S = S * ms + vs;
+                    } else {
+                        HVX_Vector diff_vec = hvx_vec_splat_f32(s_val - M);
+                        vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
+                        S += vs;
+                    }
+
+                    const uint8_t * v_ptr = v_base + ic * factx->size_v_row_padded;
+
+                    hvx_mad_f32_f16_aa(VKQ32, v_ptr, &vs, DV);
                 }
 
-                if (mask) {
-                    const float m_val = m_base[ic];
-                    s_val += slope * m_val;
-                }
-
-                const float Mold = M;
-                float vs = 1.0f;
-
-                if (s_val > M) {
-                    M = s_val;
-                    HVX_Vector diff_vec = hvx_vec_splat_f32(Mold - M);
-                    HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
-                    hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
-
-                    float ms = hvx_vec_get_f32(ms_vec);
-                    S = S * ms + vs;
-                } else {
-                    HVX_Vector diff_vec = hvx_vec_splat_f32(s_val - M);
-                    vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
-                    S += vs;
-                }
-
-                const uint8_t * v_ptr = v_base + ic * factx->size_v_row_padded;
-
-                hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, vs);
+                M_vec = hvx_vec_splat_f32(M);
+                S_vec = hvx_vec_splat_f32(S);
             }
-            M_vec = hvx_vec_splat_f32(M);
-            S_vec = hvx_vec_splat_f32(S);
 
             // Issue DMA for next+1 block (if exists)
             if (ib + 2 < factx->n_blocks) {
@@ -525,6 +556,11 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                     const uint8_t * m_src = (const uint8_t *) (mp_base + next_ic_start);
                     dma_queue_push(dma, dma_make_ptr(m_base, m_src), next_block_size * 2, next_block_size * 2, next_block_size * 2, 1);
                 }
+
+                // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u : iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
+                //         ith, ir, next_ib, iq1, iq2, iq3,
+                //         size_k_row, size_v_row, next_block_size,
+                //         (unsigned)HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - factx->t_start));
             }
         }
 
@@ -586,6 +622,8 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     struct htp_fa_context factx;
     factx.octx = octx;
 
+    factx.t_start = HAP_perf_get_qtimer_count();
+
     factx.src0_div21 = init_fastdiv_values(q->ne[2] * q->ne[1]);
     factx.src0_div1  = init_fastdiv_values(q->ne[1]);
 
@@ -632,6 +670,15 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     factx.m0 = powf(2.0f, -(max_bias       ) / factx.n_head_log2);
     factx.m1 = powf(2.0f, -(max_bias / 2.0f) / factx.n_head_log2);
 
+    // total rows in q
+    const uint32_t neq0 = q->ne[0];
+    const uint32_t neq1 = q->ne[1];
+    const uint32_t neq2 = q->ne[2];
+    const uint32_t neq3 = q->ne[3];
+
+    factx.qrows = neq1*neq2*neq3;
+    factx.qrows_per_thread = (factx.qrows + octx->n_threads - 1) / octx->n_threads;
+
     size_t size_vkq_acc = hex_round_up(v->ne[0] * sizeof(float), 128); // VKQ32
 
     octx->src0_spad.size_per_thread = size_q_block * 1;

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -82,6 +82,8 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
 int op_get_rows(struct htp_ops_context * octx) {
     get_rows_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     if (octx->src0.type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
@@ -103,9 +105,8 @@ int op_get_rows(struct htp_ops_context * octx) {
     grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src1.ne[0]);
     grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src1.ne[0] * octx->src1.ne[1]);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    grctx.src1_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    grctx.src1_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
-    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_threads);
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -13,14 +13,15 @@
 // Binary operations (add, mul, sub)
 //
 
-#define hvx_arith_loop_body(dst_type, src0_type, src1_type, vec_store, vec_op) \
+#define UNUSED(x) (void)(x)
+
+#define hvx_arith_loop_body(dst_type, src0_type, src1_type, elem_size, vec_store, vec_op) \
     do {                                                                       \
         dst_type * restrict vdst  = (dst_type *) dst;                          \
         src0_type * restrict vsrc0 = (src0_type *) src0;                       \
         src1_type * restrict vsrc1 = (src1_type *) src1;                       \
                                                                                \
-        const uint32_t elem_size = sizeof(float);                              \
-        const uint32_t epv  = 128 / elem_size;                                 \
+        const uint32_t epv  = 128 / (elem_size);                               \
         const uint32_t nvec = n / epv;                                         \
         const uint32_t nloe = n % epv;                                         \
                                                                                \
@@ -32,62 +33,74 @@
         }                                                                      \
         if (nloe) {                                                            \
             HVX_Vector v = vec_op(vsrc0[i], vsrc1[i]);                         \
-            vec_store((void *) &vdst[i], nloe * elem_size, v);                 \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
         }                                                                      \
     } while(0)
 
 #if __HVX_ARCH__ < 79
-#define HVX_OP_ADD(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b))
-#define HVX_OP_SUB(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b))
-#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b))
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b))
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+
 #else
-#define HVX_OP_ADD(a, b) Q6_Vsf_vadd_VsfVsf(a, b)
-#define HVX_OP_SUB(a, b) Q6_Vsf_vsub_VsfVsf(a, b)
-#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_vadd_VsfVsf(a, b)
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_vsub_VsfVsf(a, b)
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+
 #endif
 
+#define HVX_OP_ADD_F16(a, b) hvx_vec_add_f16_f16(a, b)
+#define HVX_OP_SUB_F16(a, b) hvx_vec_sub_f16_f16(a, b)
+#define HVX_OP_MUL_F16(a, b) hvx_vec_mul_f16_f16(a, b)
+
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO, ELEM_TYPE) \
 static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD_F32, float)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB_F32, float)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL_F32, float)
+
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f16, HVX_OP_ADD_F16, _Float16)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f16, HVX_OP_SUB_F16, _Float16)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f16, HVX_OP_MUL_F16, _Float16)
 
 // Dispatcher logic
 #define HVX_BINARY_DISPATCHER(OP_NAME) \
@@ -115,6 +128,10 @@ HVX_BINARY_DISPATCHER(hvx_add_f32)
 HVX_BINARY_DISPATCHER(hvx_sub_f32)
 HVX_BINARY_DISPATCHER(hvx_mul_f32)
 
+HVX_BINARY_DISPATCHER(hvx_add_f16)
+HVX_BINARY_DISPATCHER(hvx_sub_f16)
+HVX_BINARY_DISPATCHER(hvx_mul_f16)
+
 // Mul-Mul Optimized
 static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
     assert((unsigned long) dst % 128 == 0);
@@ -136,26 +153,25 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
 
     _Pragma("unroll(4)")
     for (; i < nvec; i++) {
-        HVX_Vector v1 = HVX_OP_MUL(vsrc0[i], vsrc1[i]);
+        HVX_Vector v1 = HVX_OP_MUL_F32(vsrc0[i], vsrc1[i]);
         vdst[i] = HVX_OP_MUL(v1, vsrc2[i]);
     }
 
     if (nloe) {
-        HVX_Vector v1 = HVX_OP_MUL(vsrc0[i], vsrc1[i]);
-        HVX_Vector v2 = HVX_OP_MUL(v1, vsrc2[i]);
+        HVX_Vector v1 = HVX_OP_MUL_F32(vsrc0[i], vsrc1[i]);
+        HVX_Vector v2 = HVX_OP_MUL_F32(v1, vsrc2[i]);
         hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v2);
     }
 }
 
 // Scalar Operations
 
-#define hvx_scalar_loop_body(dst_type, src_type, vec_store, scalar_op_macro)   \
+#define hvx_scalar_loop_body(dst_type, src_type, elem_size, vec_store, scalar_op_macro)   \
     do {                                                                       \
         dst_type * restrict vdst = (dst_type *) dst;                           \
         src_type * restrict vsrc = (src_type *) src;                           \
                                                                                \
-        const uint32_t elem_size = sizeof(float);                              \
-        const uint32_t epv  = 128 / elem_size;                                 \
+        const uint32_t epv  = 128 / (elem_size);                               \
         const uint32_t nvec = n / epv;                                         \
         const uint32_t nloe = n % epv;                                         \
                                                                                \
@@ -169,138 +185,88 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
         if (nloe) {                                                            \
             HVX_Vector v = vsrc[i];                                            \
             v = scalar_op_macro(v);                                            \
-            vec_store((void *) &vdst[i], nloe * elem_size, v);                 \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
         }                                                                      \
     } while(0)
 
-#define HVX_OP_ADD_SCALAR(v) \
+#define HVX_OP_ADD_SCALAR_F32(v) \
     ({ \
         const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, v); \
-        HVX_Vector out = HVX_OP_ADD(v, val_vec); \
+        HVX_Vector out = HVX_OP_ADD_F32(v, val_vec); \
         Q6_V_vmux_QVV(pred_inf, inf, out); \
     })
 
-#define HVX_OP_MUL_SCALAR(v) HVX_OP_MUL(v, val_vec)
-#define HVX_OP_SUB_SCALAR(v) HVX_OP_SUB(v, val_vec)
+#define HVX_OP_MUL_SCALAR_F32(v) HVX_OP_MUL_F32(v, val_vec)
+#define HVX_OP_SUB_SCALAR_F32(v) HVX_OP_SUB_F32(v, val_vec)
 
-// Add Scalar Variants
+#define HVX_OP_ADD_SCALAR_F16(v) \
+    ({ \
+        const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VhVh(inf, v); \
+        HVX_Vector out = HVX_OP_ADD_F16(v, val_vec); \
+        Q6_V_vmux_QVV(pred_inf, inf, out); \
+    })
 
-static inline void hvx_add_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD_SCALAR);
+#define HVX_OP_MUL_SCALAR_F16(v) HVX_OP_MUL_F16(v, val_vec)
+#define HVX_OP_SUB_SCALAR_F16(v) HVX_OP_SUB_F16(v, val_vec)
+
+// Scalar Variants
+
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(OP_NAME, OP_MACRO, SPLAT_MACRO, ELEM_TYPE) \
+static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src % 128 == 0); \
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) dst % 128 == 0); \
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) src % 128 == 0); \
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
+} \
+
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_add_scalar_f32, HVX_OP_ADD_SCALAR_F32, hvx_vec_splat_f32, float)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f32, HVX_OP_SUB_SCALAR_F32, hvx_vec_splat_f32, float)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_mul_scalar_f32, HVX_OP_MUL_SCALAR_F32, hvx_vec_splat_f32, float)
+
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_add_scalar_f16, HVX_OP_ADD_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f16, HVX_OP_SUB_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_mul_scalar_f16, HVX_OP_MUL_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+
+// Dispatcher logic
+#define HVX_BINARY_SCALAR_DISPATCHER(OP_NAME, ELEM_TYPE) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_aa(dst, src, val, num_elems); \
+    } else if (hex_is_aligned((void *) dst, 128)) { \
+        OP_NAME##_au(dst, src, val, num_elems); \
+    } else if (hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_ua(dst, src, val, num_elems); \
+    } else { \
+        OP_NAME##_uu(dst, src, val, num_elems); \
+    } \
 }
 
-static inline void hvx_add_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD_SCALAR);
-}
+HVX_BINARY_SCALAR_DISPATCHER(hvx_add_scalar_f32, float)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_sub_scalar_f32, float)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_mul_scalar_f32, float)
 
-static inline void hvx_add_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD_SCALAR);
-}
-
-static inline void hvx_add_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    static const float kInf = INFINITY;
-    const HVX_Vector inf = hvx_vec_splat_f32(kInf);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD_SCALAR);
-}
-
-// Sub Scalar Variants
-
-static inline void hvx_sub_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB_SCALAR);
-}
-
-// Mul Scalar Variants
-
-static inline void hvx_mul_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_add_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_add_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_add_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_add_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_add_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
-
-static inline void hvx_mul_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_mul_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_mul_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_mul_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_mul_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
-
-static inline void hvx_sub_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_sub_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_sub_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_sub_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_sub_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
+HVX_BINARY_SCALAR_DISPATCHER(hvx_add_scalar_f16, _Float16)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_sub_scalar_f16, _Float16)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_mul_scalar_f16, _Float16)
 
 // MIN Scalar variants
 
@@ -310,24 +276,24 @@ static inline void hvx_min_scalar_f32_aa(uint8_t * restrict dst, const uint8_t *
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
@@ -357,27 +323,27 @@ static inline void hvx_clamp_scalar_f32_aa(uint8_t * restrict dst, const uint8_t
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, const int num_elems) {
@@ -396,7 +362,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 // Square
 //
 
-#define hvx_sqr_loop_body(dst_type, src_type, vec_store)           \
+#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)           \
     do {                                                                   \
         dst_type * restrict vdst  = (dst_type *) dst;                      \
         src_type * restrict vsrc = (src_type *) src;                       \
@@ -410,10 +376,10 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
                                                                            \
         _Pragma("unroll(4)")                                               \
         for (; i < nvec; i++) {                                            \
-            vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]);                        \
+            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                        \
         }                                                                  \
         if (nloe) {                                                        \
-            HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]);                   \
+            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                   \
             vec_store((void *) &vdst[i], nloe * elem_size, v);             \
         }                                                                  \
     } while(0)
@@ -421,21 +387,21 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+    hvx_sqr_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
 }
 
 static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
-    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+    hvx_sqr_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
 }
 
 static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) src % 128 == 0);
-    hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+    hvx_sqr_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
 }
 
 static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+    hvx_sqr_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }
 
 static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
@@ -454,17 +420,24 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
     }
 }
 
-#undef HVX_OP_ADD
-#undef HVX_OP_SUB
-#undef HVX_OP_MUL
+#undef HVX_OP_ADD_F32
+#undef HVX_OP_SUB_F32
+#undef HVX_OP_MUL_F32
+#undef HVX_OP_ADD_F16
+#undef HVX_OP_SUB_F16
+#undef HVX_OP_MUL_F16
 #undef hvx_arith_loop_body
-#undef HVX_OP_ADD_SCALAR
-#undef HVX_OP_SUB_SCALAR
-#undef HVX_OP_MUL_SCALAR
+#undef HVX_OP_ADD_SCALAR_F32
+#undef HVX_OP_SUB_SCALAR_F32
+#undef HVX_OP_MUL_SCALAR_F32
+#undef HVX_OP_ADD_SCALAR_F16
+#undef HVX_OP_SUB_SCALAR_F16
+#undef HVX_OP_MUL_SCALAR_F16
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
 #undef DEFINE_HVX_BINARY_OP_VARIANTS
 #undef HVX_BINARY_DISPATCHER
+#undef UNUSED
 
 #endif // HVX_ARITH_H

ggml/src/ggml-hexagon/htp/hvx-base.h

@@ -38,7 +38,7 @@ static inline HVX_Vector hvx_vec_splat_f32(float v) {
     return Q6_V_vsplat_R(u.i);
 }
 
-static inline HVX_Vector hvx_vec_splat_f16(float v) {
+static inline HVX_Vector hvx_vec_splat_f16(_Float16 v) {
     union { __fp16 f; uint16_t i; } u = { .f = v };
     return Q6_Vh_vsplat_R(u.i);
 }
@@ -170,4 +170,71 @@ static inline HVX_Vector hvx_vec_i16_from_hf_rnd_sat(HVX_Vector vin) {
     return Q6_Vh_vround_VwVw_sat(vsf_1, vsf_0);
 }
 
+#if __HVX_ARCH__ < 79
+
+static inline HVX_VectorPair hvx_vec_mpyacc_f32_f16(HVX_VectorPair acc, HVX_Vector x, HVX_Vector y)
+{
+    HVX_VectorPair m = Q6_Wqf32_vmpy_VhfVhf(x, y);
+    HVX_Vector a0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_lo_W(m), Q6_V_lo_W(acc)));
+    HVX_Vector a1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_V_hi_W(m), Q6_V_hi_W(acc)));
+    return Q6_W_vcombine_VV(a1, a0);
+}
+
+#else
+
+static inline HVX_VectorPair hvx_vec_mpyacc_f32_f16(HVX_VectorPair acc, HVX_Vector x, HVX_Vector y)
+{
+    return Q6_Wsf_vmpyacc_WsfVhfVhf(acc, x, y);
+}
+
+#endif
+
+#if __HVX_ARCH__ < 79
+
+static inline HVX_Vector hvx_vec_add_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    const HVX_Vector negone = Q6_Vh_vsplat_R(0xBC00); // -1.0 in IEEE FP16
+    const HVX_Vector one    = Q6_Vh_vsplat_R(0x3C00); //  1.0 in IEEE FP16
+    HVX_VectorPair a_p = Q6_Wqf32_vmpy_VhfVhf(a, one);
+    HVX_VectorPair b_p = Q6_Wqf32_vmpy_VhfVhf(b, negone);
+    HVX_Vector a0 = Q6_Vqf32_vsub_Vqf32Vqf32(Q6_V_lo_W(a_p), Q6_V_lo_W(b_p));
+    HVX_Vector a1 = Q6_Vqf32_vsub_Vqf32Vqf32(Q6_V_hi_W(a_p), Q6_V_hi_W(b_p));
+    return Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(a1, a0));
+}
+
+static inline HVX_Vector hvx_vec_sub_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    const HVX_Vector negone = Q6_Vh_vsplat_R(0xBC00); // -1.0 in IEEE FP16
+    const HVX_Vector one    = Q6_Vh_vsplat_R(0x3C00); //  1.0 in IEEE FP16
+    HVX_VectorPair a_p = Q6_Wqf32_vmpy_VhfVhf(a, one);
+    HVX_VectorPair b_p = Q6_Wqf32_vmpy_VhfVhf(b, negone);
+    HVX_Vector a0 = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(a_p), Q6_V_lo_W(b_p));
+    HVX_Vector a1 = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(a_p), Q6_V_hi_W(b_p));
+    return Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(a1, a0));
+}
+
+static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_equals_Wqf32(Q6_Wqf32_vmpy_VhfVhf(a, b));
+}
+
+#else
+
+static inline HVX_Vector hvx_vec_add_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vadd_VhfVhf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_sub_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vsub_VhfVhf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vmpy_VhfVhf(a, b);
+}
+
+#endif // __HVX_ARCH__ < 79
+
 #endif /* HVX_BASE_H */

ggml/src/ggml-hexagon/htp/hvx-copy.h

@@ -42,11 +42,11 @@ static inline void hvx_splat_f32_u(uint8_t * restrict dst, float v, uint32_t n)
     hvx_splat_u(dst,  hvx_vec_splat_f32(v), n, sizeof(float));
 }
 
-static inline void hvx_splat_f16_a(uint8_t * restrict dst, float v, uint32_t n) {
+static inline void hvx_splat_f16_a(uint8_t * restrict dst, _Float16 v, uint32_t n) {
     hvx_splat_u(dst,  hvx_vec_splat_f16(v), n, sizeof(__fp16));
 }
 
-static inline void hvx_splat_f16_u(uint8_t * restrict dst, float v, uint32_t n) {
+static inline void hvx_splat_f16_u(uint8_t * restrict dst, _Float16 v, uint32_t n) {
     hvx_splat_u(dst,  hvx_vec_splat_f16(v), n, sizeof(__fp16));
 }
 

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -15,11 +15,144 @@
 #include "hvx-arith.h"
 
 #if __HVX_ARCH__ < 79
-#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
 #else
-#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
 #endif
 
+// Compute div by scaler in f32. Requires first by expanding fp32 to fp16 and converting the result back to fp32.
+static inline HVX_Vector hvx_div_mul_f16_const_using_f32(HVX_Vector vec1_hf, HVX_Vector vec2_sf_const, HVX_Vector vec_hf_one_1_0) {
+#if __HVX_ARCH__ < 79
+    HVX_VectorPair src_to_f32 = Q6_Wqf32_vmpy_VhfVhf(vec1_hf, vec_hf_one_1_0);
+    HVX_Vector src_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(src_to_f32));
+    HVX_Vector src_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(src_to_f32));
+#else
+    HVX_VectorPair src_to_f32 = Q6_Wsf_vmpy_VhfVhf(vec1_hf, vec_hf_one_1_0);
+    HVX_Vector src_to_f32_0 = Q6_V_lo_W(src_to_f32);
+    HVX_Vector src_to_f32_1 = Q6_V_hi_W(src_to_f32);
+#endif
+
+    HVX_Vector div_f32_0 = HVX_OP_MUL_F32(src_to_f32_0, vec2_sf_const);
+    HVX_Vector div_f32_1 = HVX_OP_MUL_F32(src_to_f32_1, vec2_sf_const);
+
+#if __HVX_ARCH__ < 79
+    HVX_Vector res = hvx_vec_f32_to_f16(div_f32_0, div_f32_1);
+#else
+    HVX_Vector res = Q6_Vhf_vcvt_VsfVsf(div_f32_0, div_f32_1);
+#endif
+    return res;
+}
+
+#define hvx_div_scaler_f16_loop_body(dst_type, src_type, vec_store)                     \
+    do {                                                                                \
+        dst_type * restrict vdst = (dst_type *) dst;                                    \
+        src_type * restrict vsrc = (src_type *) src;                                    \
+        HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                     \
+                                                                                        \
+        const uint32_t nvec = n / VLEN_FP16;                                            \
+        const uint32_t nloe = n % VLEN_FP16;                                            \
+                                                                                        \
+        uint32_t i = 0;                                                                 \
+                                                                                        \
+        _Pragma("unroll(4)")                                                            \
+        for (; i < nvec; i++) {                                                         \
+            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
+            vdst[i] = res;                                                              \
+        }                                                                               \
+        if (nloe) {                                                                     \
+            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                      \
+        }                                                                               \
+    } while(0)
+
+static inline void hvx_div_scalar_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+static inline void hvx_div_scalar_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) dst % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+static inline void hvx_div_scalar_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) src % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+static inline void hvx_div_scalar_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+// Compute div by using hvx_vec_inverse_f32_guard. Requires first by exapnding fp32 to fp16 and convert the result back to fp32.
+static inline HVX_Vector hvx_vec_div_f16_using_f32(HVX_Vector vec1, HVX_Vector vec2, HVX_Vector f32_nan_inf_mask, HVX_Vector vec_hf_one_1_0) {
+#if __HVX_ARCH__ < 79
+    // Convert first input to fp32
+    HVX_VectorPair vec1_to_f32   = Q6_Wqf32_vmpy_VhfVhf(vec1, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec1_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vec1_to_f32));
+    HVX_Vector     vec1_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(vec1_to_f32));
+
+    // Convert second input to fp32
+    HVX_VectorPair vec2_to_f32   = Q6_Wqf32_vmpy_VhfVhf(vec2, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec2_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vec2_to_f32));
+    HVX_Vector     vec2_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(vec2_to_f32));
+#else
+    // Convert first input to fp32
+    HVX_VectorPair vec1_to_f32   = Q6_Wsf_vmpy_VhfVhf(vec1, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec1_to_f32_0 = Q6_V_lo_W(vec1_to_f32);
+    HVX_Vector     vec1_to_f32_1 = Q6_V_hi_W(vec1_to_f32);
+
+    // Convert second input to fp32
+    HVX_VectorPair vec2_to_f32   = Q6_Wsf_vmpy_VhfVhf(vec2, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec2_to_f32_0 = Q6_V_lo_W(vec2_to_f32);
+    HVX_Vector     vec2_to_f32_1 = Q6_V_hi_W(vec2_to_f32);
+#endif
+
+    // Inverse second input in fp32
+    HVX_Vector     vec2_inv_f32_0 = hvx_vec_inverse_f32_guard(vec2_to_f32_0, f32_nan_inf_mask);
+    HVX_Vector     vec2_inv_f32_1 = hvx_vec_inverse_f32_guard(vec2_to_f32_1, f32_nan_inf_mask);
+
+    // Multiply first input by inverse of second, in fp32
+    HVX_Vector     div_f32_0 = HVX_OP_MUL_F32(vec1_to_f32_0, vec2_inv_f32_0);
+    HVX_Vector     div_f32_1 = HVX_OP_MUL_F32(vec1_to_f32_1, vec2_inv_f32_1);
+
+    // Convert back to fp16
+#if __HVX_ARCH__ < 79
+    HVX_Vector     recip = hvx_vec_f32_to_f16(div_f32_0, div_f32_1);
+#else
+    HVX_Vector     recip = Q6_Vhf_vcvt_VsfVsf(div_f32_0, div_f32_1);
+#endif
+
+    return recip;
+}
+
+#define hvx_div_f16_loop_body(dst_type, src0_type, src1_type, vec_store)                  \
+    do {                                                                                  \
+        dst_type * restrict vdst = (dst_type *) dst;                                      \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                                  \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                                  \
+                                                                                          \
+        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                        \
+        const HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                 \
+                                                                                          \
+        const uint32_t nvec = n / VLEN_FP16;                                              \
+        const uint32_t nloe = n % VLEN_FP16;                                              \
+                                                                                          \
+        uint32_t i = 0;                                                                   \
+                                                                                          \
+        _Pragma("unroll(4)")                                                              \
+        for (; i < nvec; i++) {                                                           \
+            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            vdst[i] = res;                                                                \
+        }                                                                                 \
+        if (nloe) {                                                                       \
+            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                        \
+        }                                                                                 \
+    } while(0)
+
 #define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store)             \
     do {                                                                             \
         dst_type * restrict vdst = (dst_type *) dst;                                 \
@@ -36,81 +169,83 @@
         _Pragma("unroll(4)")                                                         \
         for (; i < nvec; i++) {                                                      \
             HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
-            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            HVX_Vector res = HVX_OP_MUL_F32(vsrc0[i], inv_src1);                     \
             vdst[i] = res;                                                           \
         }                                                                            \
         if (nloe) {                                                                  \
             HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
-            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            HVX_Vector res = HVX_OP_MUL_F32(vsrc0[i], inv_src1);                     \
             vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res);                   \
         }                                                                            \
     } while(0)
 
-// 3-letter suffix variants
-static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src0 % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_DIV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u); \
+} \
+
+// Dispatcher logic
+#define HVX_DIV_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128)) { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
+        } \
+    } else { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
+        } \
+    } \
 }
 
-static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src0 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
-}
+DEFINE_HVX_DIV_OP_VARIANTS(hvx_div_f32, hvx_div_f32_loop_body)
+DEFINE_HVX_DIV_OP_VARIANTS(hvx_div_f16, hvx_div_f16_loop_body)
 
-static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
-}
+HVX_DIV_DISPATCHER(hvx_div_f32)
+HVX_DIV_DISPATCHER(hvx_div_f16)
 
-static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
-}
-
-static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src0 % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src0 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128)) {
-        if (hex_is_aligned((void *) src0, 128)) {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_aau(dst, src0, src1, num_elems);
-        } else {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_auu(dst, src0, src1, num_elems);
-        }
-    } else {
-        if (hex_is_aligned((void *) src0, 128)) {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_uau(dst, src0, src1, num_elems);
-        } else {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_uuu(dst, src0, src1, num_elems);
-        }
-    }
-}
-
-#undef HVX_OP_MUL
+#undef HVX_OP_MUL_F32
 
 #endif // HVX_DIV_H

ggml/src/ggml-hexagon/htp/hvx-inverse.h

@@ -67,7 +67,7 @@ static inline HVX_Vector hvx_vec_inverse_f16(HVX_Vector vals) {
 
     HVX_Vector vcl0 = Q6_Vuh_vcl0_Vuh(rm);  //count leading zeros
 
-    // Get mantissa for 16-bit represenation
+    // Get mantissa for 16-bit representation
     HVX_Vector mant_recip = Q6_V_vand_VV(Q6_Vh_vasr_VhR(Q6_Vh_vasl_VhVh(rm, vcl0), 5), Q6_Vh_vsplat_R(0x03FF));
 
     //Compute Reciprocal Exponent
@@ -137,40 +137,74 @@ static inline HVX_Vector hvx_vec_inverse_f32_guard(HVX_Vector v_sf, HVX_Vector n
         }                                                                    \
     } while(0)
 
-static inline void hvx_inverse_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+static inline HVX_Vector hvx_vec_inverse_f16_guard(HVX_Vector v_sf, HVX_Vector nan_inf_mask) {
+    HVX_Vector out = hvx_vec_inverse_f16(v_sf);
+
+    HVX_Vector     masked_out = Q6_V_vand_VV(out, nan_inf_mask);
+    const HVX_VectorPred pred = Q6_Q_vcmp_eq_VhVh(nan_inf_mask, masked_out);
+
+    return Q6_V_vmux_QVV(pred, Q6_V_vzero(), out);
 }
 
-static inline void hvx_inverse_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+#define hvx_inverse_f16_loop_body(dst_type, src_type, vec_store)             \
+    do {                                                                     \
+        dst_type * restrict vdst = (dst_type *) dst;                         \
+        src_type * restrict vsrc = (src_type *) src;                         \
+                                                                             \
+        const HVX_Vector nan_inf_mask = Q6_Vh_vsplat_R(0x7c00);              \
+                                                                             \
+        const uint32_t nvec = n / VLEN_FP16;                                 \
+        const uint32_t nloe = n % VLEN_FP16;                                 \
+                                                                             \
+        uint32_t i = 0;                                                      \
+                                                                             \
+        _Pragma("unroll(4)")                                                 \
+        for (; i < nvec; i++) {                                              \
+             vdst[i] = hvx_vec_inverse_f16_guard(vsrc[i], nan_inf_mask);     \
+        }                                                                    \
+        if (nloe) {                                                          \
+            HVX_Vector v = hvx_vec_inverse_f16_guard(vsrc[i], nan_inf_mask); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);             \
+        }                                                                    \
+    } while(0)
+
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_INV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) src % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, hvx_vec_store_u); \
+} \
+
+// Dispatcher logic
+#define HVX_INV_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_aa(dst, src, num_elems); \
+    } else if (hex_is_aligned((void *) dst, 128)) { \
+        OP_NAME##_au(dst, src, num_elems); \
+    } else if (hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_ua(dst, src, num_elems); \
+    } else { \
+        OP_NAME##_uu(dst, src, num_elems); \
+    } \
 }
 
-static inline void hvx_inverse_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) src % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
-}
+DEFINE_HVX_INV_OP_VARIANTS(hvx_inverse_f32, hvx_inverse_f32_loop_body)
+DEFINE_HVX_INV_OP_VARIANTS(hvx_inverse_f16, hvx_inverse_f16_loop_body)
 
-static inline void hvx_inverse_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    hvx_inverse_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_inverse_f32(uint8_t * restrict dst, uint8_t * restrict src, const int num_elems) {
-    if ((unsigned long) dst % 128 == 0) {
-        if ((unsigned long) src % 128 == 0) {
-            hvx_inverse_f32_aa(dst, src, num_elems);
-        } else {
-            hvx_inverse_f32_au(dst, src, num_elems);
-        }
-    } else {
-        if ((unsigned long) src % 128 == 0) {
-            hvx_inverse_f32_ua(dst, src, num_elems);
-        } else {
-            hvx_inverse_f32_uu(dst, src, num_elems);
-        }
-    }
-}
+HVX_INV_DISPATCHER(hvx_inverse_f32)
+HVX_INV_DISPATCHER(hvx_inverse_f16)
 
 #endif // HVX_INVERSE_H

ggml/src/ggml-hexagon/htp/hvx-reduce.h

@@ -46,6 +46,21 @@ static inline HVX_Vector hvx_vec_reduce_sum_qf32(HVX_Vector in) {
 
 #if __HVX_ARCH__ > 75
 
+static inline HVX_Vector hvx_vec_reduce_sum_f32x4(HVX_Vector_x4 in) {
+    HVX_VectorPair sum_p01 = Q6_W_vshuff_VVR(in.v[1], in.v[0], 4);
+    HVX_VectorPair sum_p23 = Q6_W_vshuff_VVR(in.v[3], in.v[2], 4);
+    HVX_Vector  sum_sf01  = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sum_p01), Q6_V_hi_W(sum_p01));
+    HVX_Vector  sum_sf23  = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sum_p23), Q6_V_hi_W(sum_p23));
+
+    HVX_VectorPair sum_p0123 = Q6_W_vshuff_VVR(sum_sf23, sum_sf01, 8);
+    HVX_Vector  sum_sf       = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sum_p0123), Q6_V_hi_W(sum_p0123));
+
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 2));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 4));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 8));
+    return sum_sf;
+}
+
 static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
     HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
     HVX_Vector  sum_sf  = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));
@@ -72,6 +87,21 @@ static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n)
 
 #else
 
+static inline HVX_Vector hvx_vec_reduce_sum_f32x4(HVX_Vector_x4 in) {
+    HVX_VectorPair sum_p01  = Q6_W_vshuff_VVR(in.v[1], in.v[0], 4);
+    HVX_VectorPair sum_p23  = Q6_W_vshuff_VVR(in.v[3], in.v[2], 4);
+    HVX_Vector     sum_qf01 = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sum_p01), Q6_V_hi_W(sum_p01));
+    HVX_Vector     sum_qf23 = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sum_p23), Q6_V_hi_W(sum_p23));
+
+    HVX_VectorPair sum_p0123 = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(sum_qf23), Q6_Vsf_equals_Vqf32(sum_qf01), 8);
+    HVX_Vector     sum_qf    = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sum_p0123), Q6_V_hi_W(sum_p0123));
+
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 2));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 4));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 8));
+    return Q6_Vsf_equals_Vqf32(sum_qf);
+}
+
 static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
     HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
     HVX_Vector  sum_qf  = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -1234,27 +1234,24 @@ static void vec_dot_f16_f16_aa_1x1(const int n, float * restrict s, const void *
     uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16; // leftover elements
 
-    HVX_Vector rsum = Q6_V_vsplat_R(0);
+    HVX_VectorPair rsum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
 
     uint32_t i = 0;
 
     #pragma unroll(4)
     for (i = 0; i < nvec; i++) {
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x[i], y[i]);
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        rsum_p = hvx_vec_mpyacc_f32_f16(rsum_p, x[i], y[i]);
     }
 
     if (nloe) {
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
         HVX_Vector x_hf = Q6_V_vand_QV(bmask, x[i]);
         HVX_Vector y_hf = Q6_V_vand_QV(bmask, y[i]);
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        rsum_p = hvx_vec_mpyacc_f32_f16(rsum_p, x_hf, y_hf);
     }
 
-    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
-    hvx_vec_store_u(&s[0], 4, rsum);
+    HVX_Vector rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum_p), Q6_V_hi_W(rsum_p)));
+    hvx_vec_store_u(s, 4, hvx_vec_reduce_sum_f32(rsum));
 }
 
 static void vec_dot_f16_f16_aa_2x1(const int n, float * restrict s0,
@@ -1267,35 +1264,30 @@ static void vec_dot_f16_f16_aa_2x1(const int n, float * restrict s0,
     uint32_t nvec = n / VLEN_FP16;
     uint32_t nloe = n % VLEN_FP16;
 
-    HVX_Vector rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector rsum1 = Q6_V_vsplat_R(0);
+    HVX_VectorPair rsum0_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair rsum1_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
 
     uint32_t i = 0;
 
     #pragma unroll(2)
     for (i = 0; i < nvec; i++) {
         HVX_Vector y_hf = y[i];
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0[i], y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1[i], y_hf);
-
-        rsum0 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum0, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)));
-        rsum1 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum1, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)));
+        rsum0_p = hvx_vec_mpyacc_f32_f16(rsum0_p, x0[i], y_hf);
+        rsum1_p = hvx_vec_mpyacc_f32_f16(rsum1_p, x1[i], y_hf);
     }
 
     if (nloe) {
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
+        HVX_Vector y_hf  = Q6_V_vand_QV(bmask, y[i]);
         HVX_Vector x0_hf = Q6_V_vand_QV(bmask, x0[i]);
         HVX_Vector x1_hf = Q6_V_vand_QV(bmask, x1[i]);
-        HVX_Vector y_hf  = Q6_V_vand_QV(bmask, y[i]);
-
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum0, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)));
-        rsum1 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum1, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)));
+        rsum0_p = hvx_vec_mpyacc_f32_f16(rsum0_p, x0_hf, y_hf);
+        rsum1_p = hvx_vec_mpyacc_f32_f16(rsum1_p, x1_hf, y_hf);
     }
 
-    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(Q6_Vsf_equals_Vqf32(rsum0), Q6_Vsf_equals_Vqf32(rsum1));
+    HVX_Vector rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum0_p), Q6_V_hi_W(rsum0_p)));
+    HVX_Vector rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum1_p), Q6_V_hi_W(rsum1_p)));
+    HVX_Vector rsum  = hvx_vec_reduce_sum_f32x2(rsum0, rsum1);
     hvx_vec_store_u(s0, 8, rsum);
 }
 
@@ -1311,10 +1303,10 @@ static void vec_dot_f16_f16_aa_2x2(const int n, float * restrict s0, float * res
     uint32_t nloe = n % VLEN_FP16;
 
     // Row sums (sf) - 4 accumulators for 2×2 tile
-    HVX_Vector r0_c0_sum = Q6_V_vsplat_R(0);
-    HVX_Vector r0_c1_sum = Q6_V_vsplat_R(0);
-    HVX_Vector r1_c0_sum = Q6_V_vsplat_R(0);
-    HVX_Vector r1_c1_sum = Q6_V_vsplat_R(0);
+    HVX_VectorPair r0_c0_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair r0_c1_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair r1_c0_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
+    HVX_VectorPair r1_c1_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
 
     uint32_t i = 0;
 
@@ -1326,20 +1318,10 @@ static void vec_dot_f16_f16_aa_2x2(const int n, float * restrict s0, float * res
         HVX_Vector c1_hf = y1[i];
 
         // Compute 4 dot products: r0×c0, r0×c1, r1×c0, r1×c1
-        HVX_VectorPair r0_c0_qf_p = Q6_Wqf32_vmpy_VhfVhf(r0_hf, c0_hf);
-        HVX_VectorPair r0_c1_qf_p = Q6_Wqf32_vmpy_VhfVhf(r0_hf, c1_hf);
-        HVX_VectorPair r1_c0_qf_p = Q6_Wqf32_vmpy_VhfVhf(r1_hf, c0_hf);
-        HVX_VectorPair r1_c1_qf_p = Q6_Wqf32_vmpy_VhfVhf(r1_hf, c1_hf);
-
-        HVX_Vector r0_c0_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r0_c0_qf_p), Q6_V_hi_W(r0_c0_qf_p));
-        HVX_Vector r0_c1_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r0_c1_qf_p), Q6_V_hi_W(r0_c1_qf_p));
-        HVX_Vector r1_c0_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r1_c0_qf_p), Q6_V_hi_W(r1_c0_qf_p));
-        HVX_Vector r1_c1_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r1_c1_qf_p), Q6_V_hi_W(r1_c1_qf_p));
-
-        r0_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c0_qf, r0_c0_sum));
-        r0_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c1_qf, r0_c1_sum));
-        r1_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c0_qf, r1_c0_sum));
-        r1_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c1_qf, r1_c1_sum));
+        r0_c0_sum_p = hvx_vec_mpyacc_f32_f16(r0_c0_sum_p, r0_hf, c0_hf);
+        r0_c1_sum_p = hvx_vec_mpyacc_f32_f16(r0_c1_sum_p, r0_hf, c1_hf);
+        r1_c0_sum_p = hvx_vec_mpyacc_f32_f16(r1_c0_sum_p, r1_hf, c0_hf);
+        r1_c1_sum_p = hvx_vec_mpyacc_f32_f16(r1_c1_sum_p, r1_hf, c1_hf);
     }
 
     if (nloe) {
@@ -1350,23 +1332,17 @@ static void vec_dot_f16_f16_aa_2x2(const int n, float * restrict s0, float * res
         HVX_Vector c0_hf = Q6_V_vand_QV(bmask, y0[i]);
         HVX_Vector c1_hf = Q6_V_vand_QV(bmask, y1[i]);
 
-        HVX_VectorPair r0_c0_qf_p = Q6_Wqf32_vmpy_VhfVhf(r0_hf, c0_hf);
-        HVX_VectorPair r0_c1_qf_p = Q6_Wqf32_vmpy_VhfVhf(r0_hf, c1_hf);
-        HVX_VectorPair r1_c0_qf_p = Q6_Wqf32_vmpy_VhfVhf(r1_hf, c0_hf);
-        HVX_VectorPair r1_c1_qf_p = Q6_Wqf32_vmpy_VhfVhf(r1_hf, c1_hf);
-
-        HVX_Vector r0_c0_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r0_c0_qf_p), Q6_V_hi_W(r0_c0_qf_p));
-        HVX_Vector r0_c1_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r0_c1_qf_p), Q6_V_hi_W(r0_c1_qf_p));
-        HVX_Vector r1_c0_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r1_c0_qf_p), Q6_V_hi_W(r1_c0_qf_p));
-        HVX_Vector r1_c1_qf = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(r1_c1_qf_p), Q6_V_hi_W(r1_c1_qf_p));
-
-        r0_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c0_qf, r0_c0_sum));
-        r0_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c1_qf, r0_c1_sum));
-        r1_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c0_qf, r1_c0_sum));
-        r1_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c1_qf, r1_c1_sum));
-
+        r0_c0_sum_p = hvx_vec_mpyacc_f32_f16(r0_c0_sum_p, r0_hf, c0_hf);
+        r0_c1_sum_p = hvx_vec_mpyacc_f32_f16(r0_c1_sum_p, r0_hf, c1_hf);
+        r1_c0_sum_p = hvx_vec_mpyacc_f32_f16(r1_c0_sum_p, r1_hf, c0_hf);
+        r1_c1_sum_p = hvx_vec_mpyacc_f32_f16(r1_c1_sum_p, r1_hf, c1_hf);
     }
 
+    HVX_Vector r0_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(r0_c0_sum_p), Q6_V_hi_W(r0_c0_sum_p)));
+    HVX_Vector r0_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(r0_c1_sum_p), Q6_V_hi_W(r0_c1_sum_p)));
+    HVX_Vector r1_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(r1_c0_sum_p), Q6_V_hi_W(r1_c0_sum_p)));
+    HVX_Vector r1_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(r1_c1_sum_p), Q6_V_hi_W(r1_c1_sum_p)));
+
     // Reduce and store results
     HVX_Vector r0_r1_c0_sum = hvx_vec_reduce_sum_f32x2(r0_c0_sum, r1_c0_sum);
     HVX_Vector r0_r1_c1_sum = hvx_vec_reduce_sum_f32x2(r0_c1_sum, r1_c1_sum);

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -18,7 +18,7 @@
 #include "htp-msg.h"
 #include "htp-ops.h"
 
-// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we cant include ggml.h
+// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we can't include ggml.h
 #define HTP_ROPE_TYPE_NORMAL 0
 #define HTP_ROPE_TYPE_NEOX   2
 
@@ -400,7 +400,9 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads = octx->n_threads;
+    const uint32_t ne0 = dst->ne[0];
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t dst_row_size  = dst->nb[1];
@@ -465,17 +467,14 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     rctx.dst_row_size_aligned  = dst_row_size_aligned;
     rctx.theta_cache_offset    = theta_cache_size_aligned;
 
-    uint32_t ne0 = dst->ne[0];
-    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
     rctx.src0_nrows = src0_nrows;
+    rctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
          rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs = MIN(n_threads, src0_nrows);
-        rctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, rope_job_f32, &rctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, rope_job_f32, &rctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hexagon/htp/set-rows-ops.c

@@ -128,6 +128,8 @@ static void set_rows_thread_f16_f32(unsigned int nth, unsigned int ith, void *da
 int op_set_rows(struct htp_ops_context * octx) {
     set_rows_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     if (octx->src0.type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
@@ -149,15 +151,14 @@ int op_set_rows(struct htp_ops_context * octx) {
     srctx.div_ne12 = init_fastdiv_values(ne12);
     srctx.div_ne11 = init_fastdiv_values(ne11);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    srctx.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    srctx.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
     switch(octx->dst.type) {
     case HTP_TYPE_F32:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_threads);
         break;
     case HTP_TYPE_F16:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_threads);
         break;
     default:
         return HTP_STATUS_NO_SUPPORT;

ggml/src/ggml-hexagon/htp/softmax-ops.c

@@ -353,7 +353,8 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t src1_row_size = src0_row_size;
@@ -393,12 +394,9 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
     octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
     octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
 
-    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
-        smctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_jobs);
+        smctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
+        worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hexagon/htp/sum-rows-ops.c

@@ -102,11 +102,9 @@ int op_sum_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    const int      n_threads  = octx->n_threads;
     const uint32_t src0_nrows = ne01 * ne02 * ne03;
-
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-    uint32_t rows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    const uint32_t n_threads = MIN(octx->n_threads, src0_nrows);
+    const uint32_t rows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     bool opt_path = false;
     if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
@@ -124,7 +122,7 @@ int op_sum_rows(struct htp_ops_context * octx) {
         .opt_path        = opt_path,
     };
 
-    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_thread_f32, &smctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_thread_f32, &smctx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -301,8 +301,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const int      n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t dst_row_size  = dst->nb[1];
@@ -338,11 +338,9 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
          octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
         struct htp_unary_context uctx = {
             .octx                  = octx,
-            .src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs,
+            .src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads,
             .src0_nrows            = src0_nrows,
 
             .data_src0             = (const uint8_t *)src0->data,
@@ -361,7 +359,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             .nc                    = src0->ne[0],
         };
 
-        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hexagon/htp/worker-pool.c

@@ -56,7 +56,7 @@ static void worker_pool_main(void * context) {
         unsigned int n = atomic_load(&pool->n_jobs);
         unsigned int i = atomic_fetch_add(&pool->next_job, 1);
         if (i >= n) {
-            // Spurios wakeup
+            // Spurious wakeup
             continue;
         }
 

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1281,7 +1281,7 @@ struct ggml_metal_buffer {
     bool use_residency_sets;
 
     // optional MTLResidencySet
-    // note: cannot use explicity "id<MTLResidencySet>" here because it is not available on certain OSes
+    // note: cannot use explicitly "id<MTLResidencySet>" here because it is not available on certain OSes
     id rset;
 
     // pointers to global device

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -631,7 +631,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
 
     if (!inplace) {
-        // run a separete kernel to cpy src->dst
+        // run a separate kernel to cpy src->dst
         // not sure how to avoid this
         // TODO: make a simpler cpy_bytes kernel
 
@@ -1644,7 +1644,7 @@ int ggml_metal_op_set(ggml_metal_op_t ctx, int idx) {
     const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
 
     if (!inplace) {
-        // run a separete kernel to cpy src->dst
+        // run a separate kernel to cpy src->dst
         // not sure how to avoid this
         // TODO: make a simpler cpy_bytes kernel
 
@@ -2005,7 +2005,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         const int16_t r0ptg  = nypsg*nsg;         // num src0 rows per threadgroup
               int16_t r1ptg  = 4;                 // num src1 rows per threadgroup
 
-        // note: not sure how optimal are those across all different hardware. there might be someting cleverer
+        // note: not sure how optimal are those across all different hardware. there might be something cleverer
         switch (ne11) {
             case 2:
                 r1ptg = 2; break;

ggml/src/ggml-metal/ggml-metal.cpp

@@ -14,7 +14,7 @@
 #define GGML_METAL_MAX_DEVICES 16
 
 // number of Metal devices
-// note: can be overriden with GGML_METAL_DEVICES env to simulate virtual devices
+// note: can be overridden with GGML_METAL_DEVICES env to simulate virtual devices
 static int g_devices = 1;
 
 ////////////////////////////////////////////////////////////////////////////////

ggml/src/ggml-metal/ggml-metal.metal

@@ -4218,7 +4218,7 @@ kernel void kernel_im2col(
 template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
 template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 
-// TODO: obolete -- remove
+// TODO: obsolete -- remove
 //typedef void (im2col_ext_t)(
 //        constant ggml_metal_kargs_im2col & args,
 //        device const float * x,

ggml/src/ggml-opencl/CMakeLists.txt

@@ -63,6 +63,7 @@ set(GGML_OPENCL_KERNELS
     cpy
     cvt
     diag_mask_inf
+    diag
     div
     gelu
     gemv_noshuffle_general
@@ -112,6 +113,7 @@ set(GGML_OPENCL_KERNELS
     gemm_noshuffle_q4_1_f32
     gemv_noshuffle_general_q8_0_f32
     mul
+    neg
     norm
     relu
     rms_norm
@@ -134,6 +136,7 @@ set(GGML_OPENCL_KERNELS
     tsembd
     upscale
     tanh
+    exp
     expm1
     softplus
     pad

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -313,7 +313,7 @@ struct ProfilingInfo {
     cl_ulong cmd_duration_ns;
     // The time for the kernel to complete - COMPLETE - END
     cl_ulong cmd_complete_duration_ns;
-    // Total time to finish the kernel - COMPELTE - QUEUED
+    // Total time to finish the kernel - COMPLETE - QUEUED
     cl_ulong cmd_total_duration_ns;
     // Global and local work sizes.
     size_t global_size[3];
@@ -416,7 +416,6 @@ struct ggml_backend_opencl_context {
     cl_program program_add;
     cl_program program_add_id;
     cl_program program_clamp;
-    cl_program program_cpy;
     cl_program program_cvt;
     cl_program program_diag_mask_inf;
     cl_program program_gelu;
@@ -500,6 +499,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_rms_norm, kernel_rms_norm_mul;
     cl_kernel kernel_group_norm, kernel_group_norm_mul_add;
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
+    cl_kernel kernel_diag_f32;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
@@ -514,7 +514,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
-    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
+    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32, kernel_cpy_i32_i32;
     cl_kernel kernel_mul_mat_f32_f32;
     cl_kernel kernel_mul_mat_f16_f16;
     cl_kernel kernel_mul_mat_f16_f32_1row;
@@ -550,6 +550,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_pad;
     cl_kernel kernel_tanh_f32, kernel_tanh_f32_4, kernel_tanh_f32_nc;
     cl_kernel kernel_tanh_f16, kernel_tanh_f16_4, kernel_tanh_f16_nc;
+    cl_kernel kernel_neg_f32, kernel_neg_f32_4, kernel_neg_f32_nc;
+    cl_kernel kernel_neg_f16, kernel_neg_f16_4, kernel_neg_f16_nc;
+    cl_kernel kernel_exp_f32, kernel_exp_f32_4, kernel_exp_f32_nc;
+    cl_kernel kernel_exp_f16, kernel_exp_f16_4, kernel_exp_f16_nc;
     cl_kernel kernel_expm1_f32, kernel_expm1_f32_4, kernel_expm1_f32_nc;
     cl_kernel kernel_expm1_f16, kernel_expm1_f16_4, kernel_expm1_f16_nc;
     cl_kernel kernel_softplus_f32, kernel_softplus_f32_4, kernel_softplus_f32_nc;
@@ -873,13 +877,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("cpy.cl");
 #endif
-        backend_ctx->program_cpy =
+        cl_program prog =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_cpy_f16_f16 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f16_f16", &err), err));
-        CL_CHECK((backend_ctx->kernel_cpy_f16_f32 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f16_f32", &err), err));
-        CL_CHECK((backend_ctx->kernel_cpy_f32_f16 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f32_f16", &err), err));
-        CL_CHECK((backend_ctx->kernel_cpy_f32_f32 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f16_f16 = clCreateKernel(prog, "kernel_cpy_f16_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f16_f32 = clCreateKernel(prog, "kernel_cpy_f16_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f32_f16 = clCreateKernel(prog, "kernel_cpy_f32_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f32_f32 = clCreateKernel(prog, "kernel_cpy_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_i32_i32 = clCreateKernel(prog, "kernel_cpy_i32_i32", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -932,6 +937,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // diag
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "diag.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("diag.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_diag_f32 = clCreateKernel(prog, "kernel_diag_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // gelu
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1979,6 +2001,48 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // neg
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "neg.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("neg.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_neg_f32    = clCreateKernel(prog, "kernel_neg_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_neg_f32_4  = clCreateKernel(prog, "kernel_neg_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_neg_f32_nc = clCreateKernel(prog, "kernel_neg_f32_nc", &err), err));
+        CL_CHECK((backend_ctx->kernel_neg_f16    = clCreateKernel(prog, "kernel_neg_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_neg_f16_4  = clCreateKernel(prog, "kernel_neg_f16_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_neg_f16_nc = clCreateKernel(prog, "kernel_neg_f16_nc", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // exp
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "exp.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("exp.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_exp_f32    = clCreateKernel(prog, "kernel_exp_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_exp_f32_4  = clCreateKernel(prog, "kernel_exp_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_exp_f32_nc = clCreateKernel(prog, "kernel_exp_f32_nc", &err), err));
+        CL_CHECK((backend_ctx->kernel_exp_f16    = clCreateKernel(prog, "kernel_exp_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_exp_f16_4  = clCreateKernel(prog, "kernel_exp_f16_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_exp_f16_nc = clCreateKernel(prog, "kernel_exp_f16_nc", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // expm1
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2555,7 +2619,7 @@ static std::vector<ggml_backend_device> ggml_opencl_probe_devices(ggml_backend_r
 
     cl_platform_id platform_ids[NPLAT];
     if (clGetPlatformIDs(NPLAT, platform_ids, &n_platforms) != CL_SUCCESS) {
-        GGML_LOG_ERROR("ggml_opencl: plaform IDs not available.\n");
+        GGML_LOG_ERROR("ggml_opencl: platform IDs not available.\n");
         return found_devices;
     }
 
@@ -3339,7 +3403,7 @@ static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
     CL_CHECK(clReleaseEvent(evt));
 }
 
-// Syncronizes the 'backend_ctx's device with others so that commands
+// Synchronizes the 'backend_ctx's device with others so that commands
 // enqueued to it won't start until commands in the other devices have
 // completed.
 static void sync_with_other_backends(ggml_backend_opencl_context * backend_ctx) {
@@ -3544,9 +3608,21 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                         default:
                             return false;
                     }
+                case GGML_TYPE_I32:
+                    switch (op->type) {
+                        case GGML_TYPE_I32:
+                            return true;
+                        default:
+                            return false;
+                    }
                 default:
                     return false;
             }
+        case GGML_OP_SET: {
+            return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32) &&
+                    op->type == op->src[0]->type &&
+                    op->type == op->src[1]->type;
+        }
         case GGML_OP_SCALE:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:
@@ -3580,6 +3656,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_SIGMOID:
                     return ggml_is_contiguous(op->src[0]);
                 case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_NEG:
+                case GGML_UNARY_OP_EXP:
                    return op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16;
                 case GGML_UNARY_OP_EXPM1:
                    return op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16;
@@ -3665,6 +3743,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
             return true;
+        case GGML_OP_DIAG:
+            return true;
         case GGML_OP_DIAG_MASK_INF:
             return op->ne[3] == 1;
         case GGML_OP_ROPE: {
@@ -3985,7 +4065,7 @@ struct ggml_backend_opencl_buffer_context {
 
     // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
     // before any tensor is initialized (at the beginning of alloc_tensor_range).
-    // Hence, there is alway a buffer object in this vector. When each tensor is
+    // Hence, there is always a buffer object in this vector. When each tensor is
     // being initialized, this original buffer object will be released if both
     // flattening and small allocation are enabled, and additional buffer
     // objects will be created in init_tensor to represent flattened quantized
@@ -4120,7 +4200,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         //GGML_ASSERT(offset == 0);
 
         // We create subbuffers from the original tensor buffer for scales and
-        // quants - i.e., scales and quants are aliases into the buffer obejct
+        // quants - i.e., scales and quants are aliases into the buffer object
         // that backs the original tensor. This is a cleaner way to adapt to the
         // new memory management.
         // In the old code, we allocate new buffers for scales and quants
@@ -7569,6 +7649,170 @@ static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 }
 
+static void ggml_cl_neg(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int,      ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb,  dst,  nb);
+
+    cl_kernel kernel;
+
+    if (ggml_is_contiguous(src0)) {
+        // Handle contiguous input
+        int n = ggml_nelements(dst);
+        if (n % 4 == 0) {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_neg_f32_4;
+            } else {
+                kernel = backend_ctx->kernel_neg_f16_4;
+            }
+            n /= 4;
+        } else {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_neg_f32;
+            } else {
+                kernel = backend_ctx->kernel_neg_f16;
+            }
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_int),   &n));
+
+        size_t global_work_size[] = {(size_t)CEIL_DIV(n, 64)*64, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        // Handle non-contiguous input
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_neg_f32_nc;
+        } else {
+            kernel = backend_ctx->kernel_neg_f16_nc;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb3));
+
+        int nth = 64;
+
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_exp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int,      ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb,  dst,  nb);
+
+    cl_kernel kernel;
+
+    if (ggml_is_contiguous(src0)) {
+        // Handle contiguous input
+        int n = ggml_nelements(dst);
+        if (n % 4 == 0) {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_exp_f32_4;
+            } else {
+                kernel = backend_ctx->kernel_exp_f16_4;
+            }
+            n /= 4;
+        } else {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_exp_f32;
+            } else {
+                kernel = backend_ctx->kernel_exp_f16;
+            }
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_int),   &n));
+
+        size_t global_work_size[] = {(size_t)CEIL_DIV(n, 64)*64, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        // Handle non-contiguous input
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_exp_f32_nc;
+        } else {
+            kernel = backend_ctx->kernel_exp_f16_nc;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb3));
+
+        int nth = 64;
+
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
 static void ggml_cl_expm1(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -10782,28 +11026,13 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
     // GGML_OP_DUP and GGML_OP_CONT happen between src0 and dst.
     UNUSED(dst);
 
-    const int ne00 = src0 ? src0->ne[0] : 0;
-    const int ne01 = src0 ? src0->ne[1] : 0;
-    const int ne02 = src0 ? src0->ne[2] : 0;
-    const int ne03 = src0 ? src0->ne[3] : 0;
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int,      ne1, src1, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb1, src1, nb);
 
-    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
-
-    const int ne10 = src1 ? src1->ne[0] : 0;
-    const int ne11 = src1 ? src1->ne[1] : 0;
-    const int ne12 = src1 ? src1->ne[2] : 0;
-    const int ne13 = src1 ? src1->ne[3] : 0;
-
-    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
-    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
-    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
-    const cl_ulong nb13 = src1 ? src1->nb[3] : 0;
-
-    const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-    const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+    const enum ggml_type src0t = src0->type;
+    const enum ggml_type src1t = src1->type;
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -10840,6 +11069,15 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
                     GGML_ASSERT(false && "not implemented");
             }
             break;
+        case GGML_TYPE_I32:
+            switch (src1t) {
+                case GGML_TYPE_I32:
+                    kernel = backend_ctx->kernel_cpy_i32_i32;
+                    break;
+                default:
+                    GGML_ASSERT(false && "not implemented");
+            }
+            break;
         default:
             GGML_ASSERT(false && "not implemented");
     }
@@ -10878,6 +11116,89 @@ static void ggml_cl_dup(ggml_backend_t backend, const ggml_tensor * src0, const
     UNUSED(src1);
 }
 
+static void ggml_cl_set(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32) &&
+        src1->type == src0->type && dst->type == src0->type);
+
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int,      ne1, src1, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb1, src1, nb);
+    GGML_TENSOR_LOCALS(int,      ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb,  dst,  nb);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const cl_ulong pnb1    = ((const int32_t *)dst->op_params)[0];
+    const cl_ulong pnb2    = ((const int32_t *)dst->op_params)[1];
+    const cl_ulong pnb3    = ((const int32_t *)dst->op_params)[2];
+    const cl_ulong offs    = ((const int32_t *)dst->op_params)[3];
+    const bool     inplace = (bool)((const int32_t *)dst->op_params)[4];
+
+    cl_kernel kernel = nullptr;
+
+    // for inplace case, dst is a view of src0 and is updated on top of it
+    // so for non-inplace case, copy src0 to dst first
+    if (!inplace) {
+        ggml_cl_cpy(backend, src0, dst, nullptr);
+    }
+
+    // then copy src1 to dst with specified offset
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_cpy_f32_f32;
+    } else if (src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+        kernel = backend_ctx->kernel_cpy_i32_i32;
+    } else {
+        GGML_ASSERT(false && "not implemented");
+    }
+
+    offsetd += offs;
+    cl_ulong nb = ggml_element_size(dst);
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &pnb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &pnb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &pnb3));
+
+    int max_local_size = backend_ctx->get_kernel_workgroup_size(kernel);
+
+    const int nth = MIN(max_local_size, ne00);
+
+    size_t global_work_size[] = {(size_t)ne11*nth, (size_t)ne12, (size_t)ne13};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -10940,6 +11261,49 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr
     }
 }
 
+static void ggml_cl_diag(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int,      ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb,  dst,  nb);
+
+    cl_kernel kernel = backend_ctx->kernel_diag_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_int),   &ne0));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb3));
+
+    int nth = 64;
+
+    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -11651,6 +12015,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_cpy;
             break;
+        case GGML_OP_SET:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_set;
+            break;
         case GGML_OP_DUP:
         case GGML_OP_CONT:
             if (!any_on_device) {
@@ -11750,6 +12120,18 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_tanh;
                     break;
+                case GGML_UNARY_OP_NEG:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_neg;
+                    break;
+                case GGML_UNARY_OP_EXP:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_exp;
+                    break;
                 case GGML_UNARY_OP_EXPM1:
                     if (!any_on_device) {
                         return false;
@@ -11876,6 +12258,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_nop;
             break;
+        case GGML_OP_DIAG:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_diag;
+            break;
         case GGML_OP_DIAG_MASK_INF:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/cpy.cl

@@ -182,3 +182,48 @@ kernel void kernel_cpy_f32_f32(
         dst_data[i00] = src[0];
     }
 }
+
+kernel void kernel_cpy_i32_i32(
+        global int * src0,
+        ulong offset0,
+        global int * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global int*)((global char*)src0 + offset0);
+    dst = (global int*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global int * dst_data = (global int *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const int * src = (global int *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

ggml/src/ggml-opencl/kernels/diag.cl

@@ -0,0 +1,27 @@
+kernel void kernel_diag_f32(
+    global const char * src0,
+    ulong               offset0,
+    global       char * dst,
+    ulong               offsetd,
+    ulong               nb01,
+    ulong               nb02,
+    ulong               nb03,
+    int                 ne0,
+    ulong               nb0,
+    ulong               nb2,
+    ulong               nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    global const float * src0_ptr = (global const float *)(src0 +           i2*nb02 + i3*nb03);
+    global       float * dst_ptr  = (global       float *)(dst  + i1*nb01 + i2*nb2  + i3*nb3);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}

ggml/src/ggml-opencl/kernels/exp.cl

@@ -0,0 +1,125 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_exp_f32(
+        global const float * src0,
+        ulong                offset0,
+        global       float * dst,
+        ulong                offsetd,
+        int                  n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = exp(src0[get_global_id(0)]);
+}
+
+kernel void kernel_exp_f32_4(
+        global const float4 * src0,
+        ulong                 offset0,
+        global       float4 * dst,
+        ulong                 offsetd,
+        int                   n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = exp(src0[get_global_id(0)]);
+}
+
+kernel void kernel_exp_f16(
+        global const half * src0,
+        ulong               offset0,
+        global       half * dst,
+        ulong               offsetd,
+        int                 n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = exp(src0[get_global_id(0)]);
+}
+
+kernel void kernel_exp_f16_4(
+        global const half4 * src0,
+        ulong                offset0,
+        global       half4 * dst,
+        ulong                offsetd,
+        int                  n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = exp(src0[get_global_id(0)]);
+}
+
+kernel void kernel_exp_f32_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const float * x = (global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       float * y = (global       float *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = exp(*x);
+    }
+}
+
+kernel void kernel_exp_f16_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const half * x = (global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       half * y = (global       half *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = exp(*x);
+    }
+}

ggml/src/ggml-opencl/kernels/neg.cl

@@ -0,0 +1,125 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_neg_f32(
+        global const float * src0,
+        ulong                offset0,
+        global       float * dst,
+        ulong                offsetd,
+        int                  n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = -src0[get_global_id(0)];
+}
+
+kernel void kernel_neg_f32_4(
+        global const float4 * src0,
+        ulong                 offset0,
+        global       float4 * dst,
+        ulong                 offsetd,
+        int                   n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = -src0[get_global_id(0)];
+}
+
+kernel void kernel_neg_f16(
+        global const half * src0,
+        ulong               offset0,
+        global       half * dst,
+        ulong               offsetd,
+        int                 n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = -src0[get_global_id(0)];
+}
+
+kernel void kernel_neg_f16_4(
+        global const half4 * src0,
+        ulong                offset0,
+        global       half4 * dst,
+        ulong                offsetd,
+        int                  n
+) {
+    if (get_global_id(0) >= n) {
+        return;
+    }
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = -src0[get_global_id(0)];
+}
+
+kernel void kernel_neg_f32_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const float * x = (global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       float * y = (global       float *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = -*x;
+    }
+}
+
+kernel void kernel_neg_f16_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const half * x = (global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       half * y = (global       half *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = -*x;
+    }
+}

ggml/src/ggml-sycl/common.hpp

@@ -76,10 +76,10 @@ extern int g_ggml_sycl_prioritize_dmmv;
 
 
 #define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
-#define VER_4VEC 610 // todo for hardward optimize.
-#define VER_GEN9 700 // todo for hardward optimize.
-#define VER_GEN12 1000000 // todo for hardward optimize.
-#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardward optimize.
+#define VER_4VEC 610 // todo for hardware optimize.
+#define VER_GEN9 700 // todo for hardware optimize.
+#define VER_GEN12 1000000 // todo for hardware optimize.
+#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardware optimize.
 
 #define GGML_SYCL_MAX_NODES 8192 // TODO: adapt to hardwares
 

ggml/src/ggml-sycl/quants.hpp

@@ -29,7 +29,7 @@ namespace ggml_sycl_reordered {
 // [qs0, qs1, qs2, ..., qsN]  [d0, d1, d2, ..., dN]
 //
 // Notes: out-of-bounds qs will run into d values
-// Aligment relies on the allocated size of qs
+// Alignment relies on the allocated size of qs
 
 template <ggml_type type> struct block_q_t;
 

ggml/src/ggml-sycl/softmax.cpp

@@ -37,7 +37,7 @@ struct soft_max_params {
 };
 
 // When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
-// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
+// As we want to keep pragma unroll for all other cases we suppress the clang transformation warning here.
 #ifdef __clang__
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wpass-failed"