kleidiai : add sme fp16 compute path for q4_0 gemm on aarch64 (#20043)

* Add Q4_1 OpenCL Kernels

* opencl: refactor transpose

* opencl: format

* opencl: refactor q4_1 unpack

* opencl: move `ggml_cl_mul_mat_q4_1_f32_adreno`

* opencl: refactor `ggml_cl_mul_mat_q4_1_f32_adreno` and kernels

* opencl: rename kernel files and kernes

* opencl: fix build for non adreno

* opencl: move code around and format

---------

Co-authored-by: Li He <lih@qti.qualcomm.com>

ggml/src/ggml-cpu/CMakeLists.txt

@@ -566,9 +566,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.16.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.22.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "0a9e9008adb6031f9e8cf70dff4a3321")
+        set(KLEIDIAI_ARCHIVE_MD5  "54049037570ab0ee0a0d126b2ba5ece1")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -608,6 +608,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
         set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -648,7 +649,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         if (NOT SME_ENABLED MATCHES -1)
             list(APPEND GGML_KLEIDIAI_SOURCES
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
@@ -656,10 +656,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
                 ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
-            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2+sme2+fp16")
         endif()
 
         if (NOT SVE_ENABLED MATCHES -1)

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
 // SPDX-License-Identifier: MIT
 //
 
@@ -9,7 +9,6 @@
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
-#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
 #include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h"
@@ -20,6 +19,7 @@
 #include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
+#include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
 
 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
 #include "kai_lhs_quant_pack_qsi8d32p_f32.h"
@@ -31,6 +31,7 @@
 #include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
 #include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
 #include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h"
+#include "kai_lhs_pack_f16pmrx2_f32_neon.h"
 
 #include "kai_common.h"
 
@@ -309,24 +310,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
     {
         /* SME GEMM */
         /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
         },
 
         /* .gemm_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f16pmrx2_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_pack_f16pmrx2_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_pack_f16pmrx2_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_void_fn10<kai_run_lhs_pack_f16pmrx2_f32_neon>,
         },
         /* SME GEMV */
         /* .kern_info = */ {

ggml/src/ggml-opencl/CMakeLists.txt

@@ -108,6 +108,8 @@ set(GGML_OPENCL_KERNELS
     mul_mm_q8_0_f32_l4_lm
     mul_mm_q6_k_f32_l4_lm
     mul_mm_q8_0_f32_8x4
+    gemv_noshuffle_q4_1_f32
+    gemm_noshuffle_q4_1_f32
     gemv_noshuffle_general_q8_0_f32
     mul
     norm

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

@@ -531,6 +531,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_restore_block_q4_0_noshuffle;
+    cl_kernel kernel_convert_block_q4_1_noshuffle;
+    cl_kernel kernel_restore_block_q4_1_noshuffle;
     cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q4_1_f32;
@@ -683,7 +685,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_8_buf;
     cl_kernel kernel_transpose_16_buf;
+    cl_kernel kernel_transpose_32_buf;
     cl_kernel kernel_transpose_16_4x1;
 
     // Gemm and Gemv related programs, kernels, etc
@@ -699,6 +703,8 @@ struct ggml_backend_opencl_context {
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
+    cl_kernel kernel_gemv_noshuffle_q4_1_f32;
+    cl_kernel kernel_gemm_noshuffle_q4_1_f32;
     cl_kernel kernel_mul_mm_q8_0_f32_8x4;
     cl_kernel CL_mul_mat_vec_q8_0_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
@@ -893,6 +899,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_1_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_1_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_1_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_1_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_1  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_1", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_1  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_1", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
@@ -2258,7 +2266,9 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_8_buf  = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_8_buf", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16_buf = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_buf", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_32_buf = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_buf", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16_4x1 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
@@ -2378,6 +2388,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // gemm_noshuffle_q4_1_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q4_1_f32.cl.h"
+       };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q4_1_f32.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_gemm_noshuffle_q4_1_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q4_1_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_q4_1_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q4_1_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q4_1_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q4_1_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q4_1_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mm_q8_0_f32_8x4
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2413,7 +2462,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         cl_program prog = build_program_from_source(
             backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv_general.c_str(), CL_gemv_compile_opts);
 
-        CL_CHECK((backend_ctx->CL_mul_mat_vec_q8_0_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q8_0_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q8_0_f32", &err), err));
         CL_CHECK(clReleaseProgram(prog));
         GGML_LOG_CONT(".");
     }
@@ -2923,6 +2972,82 @@ static void ggml_cl2_free(ggml_backend_t backend) {
     }
 }
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+static void transpose_2d(
+    ggml_backend_opencl_context * backend_ctx,
+    cl_kernel kernel,
+    cl_mem src, cl_mem dst, size_t size,
+    cl_int stride, cl_int rows,
+    bool blocking = true
+) {
+    static ggml_cl_buffer buf;
+
+    cl_event evt;
+    cl_int err;
+
+    buf.allocate(backend_ctx->context, size);
+
+    cl_mem trans;
+    cl_buffer_region region;
+
+    region.origin = 0;
+    region.size = size;
+    CL_CHECK((trans = clCreateSubBuffer(
+        buf.buffer, CL_MEM_READ_WRITE,
+        CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &src));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &trans));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_int), &stride));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &rows));
+
+    size_t local_size[3] = {64, 1, 1};
+    size_t global_size[3] = {(size_t)stride, (size_t)rows, 1};;
+    CL_CHECK(clEnqueueNDRangeKernel(backend_ctx->queue, kernel, 3, NULL,
+        global_size, local_size, 0, NULL, NULL));
+
+    if (blocking) {
+        CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, trans, dst, 0, 0, size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseEvent(evt));
+    } else {
+        CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, trans, dst, 0, 0, size, 0, NULL, NULL));
+    }
+
+    CL_CHECK(clReleaseMemObject(trans));
+}
+
+static void transpose_2d_as_8b(
+    ggml_backend_opencl_context * backend_ctx,
+    cl_mem src, cl_mem dst, size_t size,
+    cl_int stride, cl_int rows,
+    bool blocking = true
+) {
+    transpose_2d(backend_ctx, backend_ctx->kernel_transpose_8_buf,
+        src, dst, size, stride, rows, blocking);
+}
+
+static void transpose_2d_as_16b(
+    ggml_backend_opencl_context * backend_ctx,
+    cl_mem src, cl_mem dst, size_t size,
+    cl_int stride, cl_int rows,
+    bool blocking = true
+) {
+    transpose_2d(backend_ctx, backend_ctx->kernel_transpose_16_buf,
+        src, dst, size, stride, rows, blocking);
+}
+
+static void transpose_2d_as_32b(
+    ggml_backend_opencl_context * backend_ctx,
+    cl_mem src, cl_mem dst, size_t size,
+    cl_int stride, cl_int rows,
+    bool blocking = true
+) {
+    transpose_2d(backend_ctx, backend_ctx->kernel_transpose_32_buf,
+        src, dst, size, stride, rows, blocking);
+}
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
 //------------------------------------------------------------------------------
 // Tensor extra management
 //------------------------------------------------------------------------------
@@ -4271,7 +4396,15 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
         CL_CHECK(err);
 
+    #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         cl_kernel kernel = backend_ctx->kernel_convert_block_q4_1;
+
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_q4_1_noshuffle;
+        }
+    #else
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_1;
+    #endif // GGML_OPENCL_USE_ADRENO_KERNELS
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
@@ -4287,6 +4420,22 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
         tensor->extra = extra;
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose q as ushort
+            transpose_2d_as_16b(backend_ctx, extra->q, extra->q, size_q, K/4, M);
+            // Transpose d as ushort
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/32, M);
+            // Transpose m as ushort
+            transpose_2d_as_16b(backend_ctx, extra->m, extra->m, size_m, K/32, M);
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
         return;
     }
     if (tensor->type == GGML_TYPE_MXFP4) {
@@ -4795,6 +4944,53 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
     if (tensor->type == GGML_TYPE_Q4_1) {
         ggml_tensor_extra_cl_q4_1 * extra = (ggml_tensor_extra_cl_q4_1 *)tensor->extra;
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            static ggml_cl_buffer buf_trans_q;
+            static ggml_cl_buffer buf_trans_m;
+            static ggml_cl_buffer buf_trans_d;
+            static ggml_cl_buffer buf_unpacked;
+
+            cl_int M = tensor->ne[1];
+            cl_int K = tensor->ne[0];
+
+            GGML_ASSERT(K % ggml_blck_size(tensor->type) == 0);
+
+            size_t size_q = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*ggml_blck_size(tensor->type)/2;
+            size_t size_d = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            size_t size_m = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            GGML_ASSERT(size_d + size_q + size_m == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+            buf_trans_q.allocate(backend_ctx->context, size_q);
+            buf_trans_m.allocate(backend_ctx->context, size_m);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));
+
+            // transpose q, d, m back
+            transpose_2d_as_16b(backend_ctx, extra->q, buf_trans_q.buffer, size_q, M, K/4);
+            transpose_2d_as_16b(backend_ctx, extra->d, buf_trans_d.buffer, size_d, M, K/32);
+            transpose_2d_as_16b(backend_ctx, extra->m, buf_trans_m.buffer, size_m, M, K/32);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q4_1_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_q.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &buf_trans_m.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &buf_unpacked.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_F0));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
+            return;
+        }
+#endif
+
         cl_int err;
         cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
             ggml_nbytes(tensor), NULL, &err);
@@ -4886,8 +5082,8 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
 
             int ne00 = tensor->ne[0];
             int ne01 = tensor->ne[1];
-            GGML_ASSERT(tensor->ne[2] == 1);  // ???
-            GGML_ASSERT(tensor->ne[3] == 1);  // ???
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
 
             CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
             CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
@@ -8371,6 +8567,180 @@ static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_ten
     CL_CHECK(clReleaseMemObject(D_sub_buffer));
 }
 
+static void ggml_cl_mul_mat_q4_1_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    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;
+    ggml_tensor_extra_cl_q4_1 * extra0_q4_1 = (ggml_tensor_extra_cl_q4_1 *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
+    cl_buffer_region    region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (ne1 == 1) {
+        cl_mem q_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for q
+        img_fmt = { CL_R, CL_UNSIGNED_INT32};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer = extra0_q4_1->q;
+        CL_CHECK((q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q4_1_f32;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &q_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q4_1->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q4_1->m));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne01));
+
+        size_t local_work_size[3] = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(q_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // transpose activations
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q4_1_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_q4_1->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q4_1->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q4_1->m));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_int),   &ne1));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     GGML_ASSERT(src0);
@@ -8736,6 +9106,16 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     int padding;
     // <--------------------------------------------> //
 
+    // NOTE: Kernels using image1d_buffer_t (e.g., src0_q) would normally require
+    // a limit check, but q4_0 / q4_1 tensors are very unlikely to exceed that
+    // limit, so the check is omitted.
+
+    // q4_1 x fp32
+    if (src0t == GGML_TYPE_Q4_1 && src1t == GGML_TYPE_F32) {
+            ggml_cl_mul_mat_q4_1_f32_adreno(backend, src0, src1, dst);
+            return;
+    }
+
     // q8_0 x fp32
     if (src0t == GGML_TYPE_Q8_0 && src1t == GGML_TYPE_F32 &&
         enable_adreno_trans_weight(backend_ctx, src0)) {

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

@@ -199,6 +199,58 @@ kernel void kernel_restore_block_q4_1(
     }
 }
 
+kernel void kernel_convert_block_q4_1_noshuffle(
+    global struct block_q4_1 * src0,
+    global uchar * dst_q,
+    global half  * dst_d,
+    global half  * dst_m
+) {
+    global struct block_q4_1 * b = (global struct block_q4_1 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_1/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+    global half  * m = (global half *) dst_m + get_global_id(0);
+
+    *d = b->d;
+    *m = b->m;
+    for (int i = 0; i < QK4_1/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0      ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
+        q[i + QK4_1/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
+
+#ifdef ADRENO_GPU
+        if (get_global_id(0) == 65536*4096) {
+            printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
+        }
+#endif
+    }
+}
+
+kernel void kernel_restore_block_q4_1_noshuffle(
+    global uchar * src_q,
+    global half  * src_d,
+    global half  * src_m,
+    global struct block_q4_1 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_1 * b = (global struct block_q4_1 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_1/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+    global half  * m = (global half *) src_m + get_global_id(0);
+
+    b->d = *d;
+    b->m = *m;
+    for (int i = 0; i < QK4_1/4; ++i) {
+        uchar x0 = q[i + 0      ] ;
+        uchar x1 = q[i + QK4_1/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
 //------------------------------------------------------------------------------
 // block_mxfp4
 //------------------------------------------------------------------------------

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl

@@ -0,0 +1,132 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_gemm_noshuffle_q4_1_f32(
+    global const ushort * src0_q,
+    global const half  * src0_d,
+    global const half  * src0_m,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    global const ushort* weight_ptr = src0_q + gx_2;
+    global const half*   scale_ptr  = src0_d + gx_2;
+    global const half*   min_ptr    = src0_m + gx_2;
+
+    for(int i = 0; i < k; i += 4) {
+        B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1);
+
+        ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m));
+
+        half4 scale = vload4(0, scale_ptr + (i/32)*(m));
+        half4 minv  = vload4(0,   min_ptr + (i/32)*(m));
+
+        // j=0
+        dequantized_weights.s0 = (bits4.s0 & (0x000F)) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = (bits4.s1 & (0x000F)) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = (bits4.s2 & (0x000F)) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = (bits4.s3 & (0x000F)) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1);
+        dequantized_weights.s0 = ((bits4.s0 & (0x00F0)) >> 4) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = ((bits4.s1 & (0x00F0)) >> 4) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0x00F0)) >> 4) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0x00F0)) >> 4) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1);
+        dequantized_weights.s0 = ((bits4.s0 & (0x0F00)) >> 8) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = ((bits4.s1 & (0x0F00)) >> 8) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0x0F00)) >> 8) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0x0F00)) >> 8) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1);
+        dequantized_weights.s0 = ((bits4.s0 & (0xF000)) >> 12) * scale.s0 + minv.s0;
+        dequantized_weights.s1 = ((bits4.s1 & (0xF000)) >> 12) * scale.s1 + minv.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0xF000)) >> 12) * scale.s2 + minv.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0xF000)) >> 12) * scale.s3 + minv.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl

@@ -121,7 +121,7 @@
 #ifdef ADRENO_GPU
 REQD_SUBGROUP_SIZE_64
 #endif
-__kernel void kernel_gemv_noshuffle(
+__kernel void kernel_gemv_noshuffle_q8_0_f32(
         __read_only  image1d_buffer_t src0_q,  // quantized A
         global half  * src0_d,  // A scales
         __read_only  image1d_buffer_t src1,    // B

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl

@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK4_0 32
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, minv, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, minv, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 + minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 + minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 + minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 + minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 + minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 + minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_1_f32(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+
+    private uint4     regA;
+    private half2     regS;
+    private half2     regM;
+    private float8    regB;
+
+    private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += NSUBGROUPS) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        regM = src0_m[gid + k * LINE_STRIDE_A]; // each fiber loads min of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

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

@@ -44,6 +44,19 @@ kernel void kernel_transpose_16_4x1(
     write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
 }
 
+// Transpose treating each element as 8-bit using buffer
+kernel void kernel_transpose_8_buf(
+    global const uchar * input,
+    global uchar * output,
+    const int ldi,
+    const int ldo
+) {
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+
+    output[x*ldo + y] = input[y*ldi + x];
+}
+
 // Transpose treating each element as 16-bit using buffer
 kernel void kernel_transpose_16_buf(
     global const ushort * input,
@@ -57,6 +70,19 @@ kernel void kernel_transpose_16_buf(
     output[x*ldo + y] = input[y*ldi + x];
 }
 
+// Transpose treating each element as 32-bit using buffer
+kernel void kernel_transpose_32_buf(
+    global const uint * input,
+    global uint * output,
+    const int ldi,
+    const int ldo
+) {
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+
+    output[x*ldo + y] = input[y*ldi + x];
+}
+
 // 32-bit transpose, loading/storing a 4x4 tile of elements
 kernel void kernel_transpose_32(
     __read_only image1d_buffer_t input,

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -31,6 +31,13 @@
 #define ROUNDUP_POW2(x, pow2) (((x) + ((pow2) - 1)) & ~((pow2) - 1))
 #define CEIL_DIV(M, N)        (((M) + (N) - 1) / (N))
 
+// Return a rectangular grid of workgroups with minimal over-provisioned workgroups.
+// Assumes that the total number of workgroups does not exceed max_per_dim^2.
+static inline void compute_2d_workgroups(uint32_t total_wg, uint32_t max_per_dim, uint32_t & wg_x, uint32_t & wg_y) {
+    wg_y = std::max(1u, CEIL_DIV(total_wg, max_per_dim));
+    wg_x = CEIL_DIV(total_wg, wg_y);
+}
+
 #ifdef GGML_WEBGPU_DEBUG
 #    define WEBGPU_LOG_DEBUG(msg)  std::cout << msg << std::endl
 #    define WEBGPU_DEBUG_BUF_ELEMS 512
@@ -69,8 +76,8 @@
 
 /* Constants */
 
-#define WEBGPU_NUM_PARAM_BUFS                16u
-#define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     8u
+#define WEBGPU_NUM_PARAM_BUFS                48u
+#define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     16u
 #define WEBGPU_WAIT_ANY_TIMEOUT_MS           0
 // Maximum number of in-flight submissions per-thread, to avoid exhausting the
 // parameter buffer pool
@@ -1146,8 +1153,9 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
     };
 
     // Calculate workgroup dimensions
-    uint32_t wg_x = 1;
-    uint32_t wg_y = 1;
+    uint32_t       wg_x = 1;
+    uint32_t       wg_y = 1;
+    const uint32_t max_wg_per_dim = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
 
     if (use_fast && is_vec) {
         auto decisions = static_cast<ggml_webgpu_mul_mat_vec_shader_decisions *>(pipeline.context.get());
@@ -1155,9 +1163,7 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
         uint32_t batches       = dst->ne[2] * dst->ne[3];
         uint32_t output_groups = CEIL_DIV(dst->ne[0], decisions->outputs_per_wg);
         uint32_t total_wg      = output_groups * batches;
-        // TODO: split large sizes into multiple batches to avoid way over-provisioning workgroups
-        wg_x = std::min(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
-        wg_y = CEIL_DIV(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
+        compute_2d_workgroups(total_wg, max_wg_per_dim, wg_x, wg_y);
     } else if (use_fast) {
         auto decisions = static_cast<ggml_webgpu_mul_mat_shader_decisions *>(pipeline.context.get());
 
@@ -1176,12 +1182,14 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
             wg_m              = CEIL_DIV(dst->ne[0], tile_m_s);
             wg_n              = CEIL_DIV(dst->ne[1], tile_n_s);
         }
-        wg_x = wg_m * wg_n * dst->ne[2] * dst->ne[3];
+        uint32_t total_wg = wg_m * wg_n * dst->ne[2] * dst->ne[3];
+        compute_2d_workgroups(total_wg, max_wg_per_dim, wg_x, wg_y);
+
     } else {  // legacy
         auto     decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
         uint32_t wg_size   = decisions->wg_size;
-        wg_x               = CEIL_DIV(dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3], wg_size);
-        wg_y               = 1;
+        uint32_t total_wg  = CEIL_DIV(dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3], wg_size);
+        compute_2d_workgroups(total_wg, max_wg_per_dim, wg_x, wg_y);
     }
 
     return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, wg_y);

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl

@@ -679,19 +679,24 @@ struct MulMatParams {
 @group(0) @binding(3) var<uniform> params: MulMatParams;
 
 @compute @workgroup_size(256)
-fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+fn main(@builtin(local_invocation_id) local_id: vec3<u32>,
+        @builtin(workgroup_id) wg_id: vec3<u32>,
+        @builtin(num_workgroups) num_wg: vec3<u32>) {
+    let wg_linear = wg_id.y * num_wg.x + wg_id.x;
+    let global_idx = wg_linear * 256u + local_id.x;
+
     let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
-    if (global_id.x >= total) {
+    if (global_idx >= total) {
         return;
     }
 
     let dst2_stride = params.m * params.n;
     let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
 
-    let dst3_idx = global_id.x / dst3_stride;
+    let dst3_idx = global_idx / dst3_stride;
     let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
     let src13_idx = dst3_idx; // src1 is not broadcast
-    let dst3_rem = global_id.x % dst3_stride;
+    let dst3_rem = global_idx % dst3_stride;
 
     let dst2_idx = dst3_rem / dst2_stride;
     let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl

@@ -54,7 +54,8 @@ var<workgroup> shmem: array<f16, TILE_SRC0_SHMEM + TILE_SRC1_SHMEM>;
 
 @compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
 fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
-        @builtin(local_invocation_id) local_id: vec3<u32>) {
+        @builtin(local_invocation_id) local_id: vec3<u32>,
+        @builtin(num_workgroups) num_wg: vec3<u32>) {
 
     let thread_id = local_id.x;
     let local_m = get_local_m(thread_id);
@@ -64,9 +65,16 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
     let wg_m_count = (params.m + WORKGROUP_SIZE_M * TILE_M - 1u) / (WORKGROUP_SIZE_M * TILE_M);
     let wg_per_matrix = wg_m_count * wg_n_count;
 
-    let batch_idx = wg_id.x / wg_per_matrix;
+    let wg_linear = wg_id.y * num_wg.x + wg_id.x;
 
-    let wg_in_batch = wg_id.x % wg_per_matrix;
+    let batch_idx = wg_linear / wg_per_matrix;
+
+    let total_batches = params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    if (batch_idx >= total_batches) {
+        return;
+    }
+
+    let wg_in_batch = wg_linear % wg_per_matrix;
     let wg_m = wg_in_batch % wg_m_count;
     let wg_n = wg_in_batch / wg_m_count;
 

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl

@@ -69,7 +69,8 @@ var<workgroup> shmem: array<f16, SHMEM_SIZE>;
 @compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
 fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
         @builtin(local_invocation_id) local_id: vec3<u32>,
-        @builtin(subgroup_id) subgroup_id: u32) {
+        @builtin(subgroup_id) subgroup_id: u32,
+        @builtin(num_workgroups) num_wg: vec3<u32>) {
 
     let thread_id = local_id.x;
     let subgroup_m = subgroup_id % SUBGROUP_M;
@@ -79,9 +80,16 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
     let wg_n_count = (params.n + WG_N_SG_TILE_SIZE - 1) / WG_N_SG_TILE_SIZE;
     let wg_per_matrix = wg_m_count * wg_n_count;
 
-    let batch_idx = wg_id.x / wg_per_matrix;
+    let wg_linear = wg_id.y * num_wg.x + wg_id.x;
 
-    let wg_in_batch = wg_id.x % wg_per_matrix;
+    let batch_idx = wg_linear / wg_per_matrix;
+
+    let total_batches = params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    if (batch_idx >= total_batches) {
+        return;
+    }
+
+    let wg_in_batch = wg_linear % wg_per_matrix;
     let wg_m = wg_in_batch % wg_m_count;
     let wg_n = wg_in_batch / wg_m_count;