cli: add --output option (#25484)

* hexagon: tile wide rows in pointwise unary ops to avoid VTCM overflow

* unary: reject permuted tensors for now (not used by models)

* hex-unary: replace divs with fastdiv

* hex-unary: add vtcm layout and host computed kernel params

* hex-unary: move fastdiv init into kernel params

* hex-unary: add specialized thread functions to improve generated code

* hex-unary: tracing instrumentation for unary ops

* hex-unary: factor out hvx kernels, streamline and remove more duplication

* ggml-hexagon: fix std::min collision with Windows min macro

* hex-cmake: make lto build happy

---------

Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>

common/arg.cpp

@@ -2849,7 +2849,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.out_file = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE,
-                    LLAMA_EXAMPLE_RESULTS, LLAMA_EXAMPLE_EXPORT_GRAPH_OPS}));
+                    LLAMA_EXAMPLE_RESULTS, LLAMA_EXAMPLE_EXPORT_GRAPH_OPS, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),

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

@@ -44,6 +44,7 @@
 #include "htp-ops.h"
 #include "htp/matmul-ops.h"
 #include "htp/flash-attn-ops.h"
+#include "htp/unary-ops.h"
 #include "htp_iface.h"
 #include "htp-drv.h"
 
@@ -170,8 +171,8 @@ static inline bool ggml_hexagon_is_hmx_weight_type(enum ggml_type type) {
     return type == GGML_TYPE_F16 || type == GGML_TYPE_F32 || ggml_hexagon_is_repack_type(type);
 }
 
-struct htp_mm_kernel_params;
 struct ggml_hexagon_session;
+
 static void ggml_hexagon_precompute_matmul_params(
     const struct ggml_hexagon_session * sess,
     const struct ggml_tensor * src0,
@@ -180,6 +181,15 @@ static void ggml_hexagon_precompute_matmul_params(
     struct htp_mm_kernel_params * kparams
 );
 
+static void ggml_hexagon_precompute_unary_params(
+    const struct ggml_hexagon_session * sess,
+    uint32_t op,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * src1,
+    const struct ggml_tensor * dst,
+    struct htp_unary_kernel_params * kparams
+);
+
 static void ggml_hexagon_precompute_fused_qkv_params(
     const struct ggml_hexagon_session * sess,
     const struct ggml_tensor * src0,
@@ -2591,6 +2601,74 @@ finalize:
     kparams->div_ne11     = init_fastdiv_values(ne11);
 }
 
+static void ggml_hexagon_precompute_unary_params(
+    const struct ggml_hexagon_session * sess,
+    uint32_t op,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * src1,
+    const struct ggml_tensor * dst,
+    struct htp_unary_kernel_params * kparams
+) {
+    memset(kparams, 0, sizeof(*kparams));
+
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = (std::min)((uint32_t)sess->n_threads, src0_nrows);
+
+    kparams->n_threads = n_threads;
+
+    const size_t src0_data_row_size = src0->ne[0] * sizeof(float);
+    const size_t dst_data_row_size  = dst->ne[0]  * sizeof(float);
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);
+
+    kparams->src0_row_size_aligned = src0_row_size_aligned;
+    kparams->dst_row_size_aligned  = dst_row_size_aligned;
+
+    size_t src1_data_row_size = 0;
+    size_t src1_row_size_aligned = 0;
+    bool broadcast_weight = false;
+
+    if (op == HTP_OP_RMS_NORM_MUL) {
+        GGML_ASSERT(src1 != nullptr);
+        src1_data_row_size = src1->ne[0] * sizeof(float);
+        src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
+        broadcast_weight = (src1->ne[1] * src1->ne[2] * src1->ne[3] == 1);
+    }
+
+    kparams->src1_row_size_aligned = src1_row_size_aligned;
+    kparams->broadcast_weight      = broadcast_weight;
+
+    struct htp_unary_vtcm_layout L;
+    uint32_t col_tile = 0;
+    uint32_t vtcm_row_per_thread = 0;
+
+    htp_unary_vtcm_layout_build(&L, op, src0->ne[0], dst->ne[0],
+                                op == HTP_OP_RMS_NORM_MUL ? src1->ne[0] : 0,
+                                broadcast_weight, n_threads, sess->vtcm_size,
+                                &col_tile, &vtcm_row_per_thread);
+
+    kparams->col_tile = col_tile;
+    kparams->vtcm_row_per_thread = vtcm_row_per_thread;
+    kparams->vtcm_size = L.total_bytes;
+
+    kparams->vtcm_src0_size_per_thread = L.src0_bytes;
+    kparams->vtcm_src1_size_per_thread = L.src1_bytes;
+    kparams->vtcm_dst_size_per_thread  = L.dst_bytes;
+
+    kparams->vtcm_src0_size = L.src0_bytes * n_threads;
+    kparams->vtcm_src1_size = L.src1_bytes * n_threads;
+    kparams->vtcm_dst_size  = L.dst_bytes * n_threads;
+
+    kparams->block = col_tile ? 0 : ((L.src0_bytes / 2) / src0_row_size_aligned);
+
+    const uint32_t tiles_per_row = col_tile > 0 ? (src0->ne[0] + col_tile - 1) / col_tile : 1;
+    kparams->div_ne01  = init_fastdiv_values(src0->ne[1]);
+    kparams->div_ne02  = init_fastdiv_values(src0->ne[2]);
+    kparams->div_ne012 = init_fastdiv_values(src0->ne[1] * src0->ne[2]);
+    kparams->div_tpr   = init_fastdiv_values(tiles_per_row);
+}
+
 static void ggml_hexagon_precompute_fused_qkv_params(
     const struct ggml_hexagon_session * sess,
     const struct ggml_tensor * src0, // Wk
@@ -2866,6 +2944,9 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
         return false;
     }
 
+    if (ggml_is_permuted(src0) || ggml_is_permuted(dst)) {
+        return false;
+    }
     if (!ggml_are_same_shape(src0, dst)) {
         return false;
     }
@@ -2912,6 +2993,9 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     if (dst->type != GGML_TYPE_F32) {
         return false;
     }
+    if (ggml_is_permuted(src0)) {
+        return false;
+    }
     if (!ggml_are_same_shape(src0, dst)) {
         return false;
     }
@@ -3451,6 +3535,15 @@ static bool try_fuse_node(const ggml_hexagon_session * sess, const ggml_cgraph *
         if (next_node->op == GGML_OP_MUL && op_is_compute(next_node) && ggml_can_fuse(graph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
             htp_opnode node(n, {}, HTP_OP_RMS_NORM_MUL);
             node.add_fused(next_node);
+
+            auto inputs = node.get_inputs();
+            const struct ggml_tensor * src0 = inputs[0];
+            const struct ggml_tensor * src1 = inputs.size() > 1 ? inputs[1] : nullptr;
+            ggml_hexagon_precompute_unary_params(sess,
+                node.opcode, src0, src1, node.dst(),
+                (struct htp_unary_kernel_params *)node.kernel_params
+            );
+
             nodes.push_back(std::move(node));
             i++; // skip the fused MUL node
             return true;
@@ -3555,6 +3648,14 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                     node.node,
                     (struct htp_fa_kernel_params *)node.kernel_params
                 );
+            } else if (htp_op_is_unary(node.opcode)) {
+                auto inputs = node.get_inputs();
+                const struct ggml_tensor * src0 = inputs[0];
+                const struct ggml_tensor * src1 = inputs.size() > 1 ? inputs[1] : nullptr;
+                ggml_hexagon_precompute_unary_params(sess,
+                    node.opcode, src0, src1, node.dst(),
+                    (struct htp_unary_kernel_params *)node.kernel_params
+                );
             }
             computed_nodes.push_back(std::move(node));
         }

ggml/src/ggml-hexagon/htp-opnode.h

@@ -12,6 +12,7 @@
 #include "htp-ops.h"
 #include "htp/matmul-ops.h"
 #include "htp/flash-attn-ops.h"
+#include "htp/unary-ops.h"
 
 struct htp_opnode {
     ggml_tensor * node = nullptr;
@@ -362,6 +363,9 @@ struct htp_opformat {
                 path = "hvx";
             }
             snprintf(str, max_size, "%s vtcm %d", path, (int) kparams->vtcm_size);
+        } else if (htp_op_is_unary(node.opcode)) {
+            const auto * kparams = (const struct htp_unary_kernel_params *) node.kernel_params;
+            snprintf(str, max_size, "%s vtcm %d", kparams->col_tile ? "wide-row" : "row-block", (int) kparams->vtcm_size);
         } else {
             snprintf(str, max_size, "----");
         }

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -39,8 +39,8 @@ add_library(${HTP_LIB} SHARED
     diag-ops.c
     solve-tri-ops.c
     pad-ops.c
-    flash-attn-ops.c
     matmul-ops.c
+    flash-attn-ops.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -120,7 +120,6 @@ int op_concat(struct htp_ops_context * octx);
 int op_diag(struct htp_ops_context * octx);
 int op_solve_tri(struct htp_ops_context * octx);
 int op_gated_delta_net(struct htp_ops_context * octx);
-int op_tri(struct htp_ops_context * octx);
 int op_pad(struct htp_ops_context * octx);
 
 #endif /* HTP_CTX_H */

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

@@ -0,0 +1,257 @@
+#ifndef HVX_NORM_H
+#define HVX_NORM_H
+
+#include <stdint.h>
+#include "hvx-base.h"
+#include "hvx-reduce.h"
+#include "hvx-inverse.h"
+#include "hvx-sqrt.h"
+#include "hvx-repl.h"
+
+static inline void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
+                                         uint8_t * restrict dst,
+                                         const int num_elems,
+                                         float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP32;    // number of full vectors
+    const int nloe = num_elems % VLEN_FP32;    // leftover elements
+
+    // Compute sum of squares for full vectors
+    HVX_Vector sum_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    // Reduce HVX sum
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    // Scale full vectors
+    HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        v_dst[i] = Q6_Vsf_equals_Vqf32(v2);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        HVX_Vector result = Q6_Vsf_equals_Vqf32(v2);
+
+        // Store with masking to avoid overwriting memory beyond the tensor
+        hvx_vec_store_a(&v_dst[nvec], nloe * 4, result);
+    }
+}
+
+static inline void hvx_fast_rms_norm_mul_f32(const uint8_t * restrict src,
+                                             const uint8_t * restrict weight,
+                                             uint8_t * restrict dst,
+                                             const int num_elems,
+                                             float     epsilon) {
+    const HVX_Vector * restrict v_src    = (const HVX_Vector *) src;
+    const HVX_Vector * restrict v_weight = (const HVX_Vector *) weight;
+    HVX_Vector * restrict v_dst          = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP32;    // number of full vectors
+    const int nloe = num_elems % VLEN_FP32;    // leftover elements
+
+    // Compute sum of squares for full vectors
+    HVX_Vector sum_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    // Reduce HVX sum
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    // Scale and multiply
+    HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        HVX_Vector v3 = Q6_Vsf_equals_Vqf32(v2);
+        HVX_Vector result = Q6_Vqf32_vmpy_VsfVsf(v3, v_weight[i]);
+        v_dst[i] = Q6_Vsf_equals_Vqf32(result);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        HVX_Vector v3 = Q6_Vsf_equals_Vqf32(v2);
+        HVX_Vector result = Q6_Vqf32_vmpy_VsfVsf(v3, v_weight[nvec]);
+        HVX_Vector res_v = Q6_Vsf_equals_Vqf32(result);
+
+        // Store with masking to avoid overwriting memory beyond the tensor
+        hvx_vec_store_a(&v_dst[nvec], nloe * 4, res_v);
+    }
+}
+
+static inline void hvx_fast_norm_f32(const uint8_t * restrict src,
+                                     uint8_t * restrict dst,
+                                     const int num_elems,
+                                     float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP32;    // number of full vectors
+    const int nloe = num_elems % VLEN_FP32;    // leftover elements
+
+    // Compute sum of squares and sum of values for full vectors
+    HVX_Vector sum_sq_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector sum_x_v  = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, v2);
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(v1, Q6_V_vzero()));
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, v2);
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(v1, Q6_V_vzero()));
+    }
+
+    // Reduce HVX sums
+    sum_sq_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_sq_v));
+    sum_x_v  = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_x_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_sq_v      = Q6_Vqf32_vmpy_VsfVsf(sum_sq_v, denom_v);
+    HVX_Vector mean_x_v       = Q6_Vqf32_vmpy_VsfVsf(sum_x_v,  denom_v);
+    HVX_Vector mean_x_sq_v    = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(mean_x_v), Q6_Vsf_equals_Vqf32(mean_x_v));
+    HVX_Vector var_v          = Q6_Vqf32_vsub_Vqf32Vqf32(mean_sq_v, mean_x_sq_v);
+    HVX_Vector var_epsilon_v  = Q6_Vqf32_vadd_Vqf32Vsf(var_v, epsilon_v);
+
+    // scale = rsqrt(variance + epsilon),  mean_x broadcast for subtraction
+    HVX_Vector scale_v  = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(var_epsilon_v));
+    HVX_Vector mean_x_b = hvx_vec_repl_f32(Q6_Vsf_equals_Vqf32(mean_x_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v1, mean_x_b);
+        HVX_Vector v3 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v2), scale_v);
+        v_dst[i] = Q6_Vsf_equals_Vqf32(v3);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v1, mean_x_b);
+        HVX_Vector v3 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v2), scale_v);
+        HVX_Vector result = Q6_Vsf_equals_Vqf32(v3);
+
+        // Store with masking to avoid overwriting memory beyond the tensor
+        hvx_vec_store_a(&v_dst[nvec], nloe * 4, result);
+    }
+}
+
+static inline void hvx_fast_l2_norm_f32(const uint8_t * restrict src,
+                                        uint8_t * restrict dst,
+                                        const int num_elems,
+                                        float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    HVX_Vector sum_v = hvx_vec_splat_f32(0.0f);
+
+    const int nvec = num_elems / VLEN_FP32;
+    const int nloe = num_elems % VLEN_FP32;
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector sq = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, sq);
+    }
+
+    // Include tail elements in the sum-of-squares using a predicate mask
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector sq = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, sq);
+    }
+
+    // Compute scale = 1/fmax(sqrt(sum), epsilon) entirely in HVX registers.
+    // hvx_vec_rsqrt_f32 + hvx_vec_inverse_f32 avoids scalar extraction.
+    HVX_Vector sum_sf    = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+    HVX_Vector rsqrt_v   = hvx_vec_rsqrt_f32(sum_sf);              // 1/sqrt(sum)
+    HVX_Vector sqrt_v    = hvx_vec_inverse_f32(rsqrt_v);            // sqrt(sum)
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+    HVX_Vector denom_v   = Q6_Vsf_vmax_VsfVsf(sqrt_v, epsilon_v);  // fmax(sqrt(sum), epsilon)
+    HVX_Vector scale_v   = hvx_vec_inverse_f32(denom_v);            // 1/fmax(sqrt(sum), epsilon)
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_Vector v1 = v_src[i];
+        v_dst[i]      = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(v1, scale_v));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector result = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(v1, scale_v));
+        hvx_vec_store_a(&v_dst[nvec], nloe * 4, result);
+    }
+}
+
+#endif // HVX_NORM_H

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

@@ -19,5 +19,6 @@
 #include "hvx-base.h"
 #include "hvx-pow.h"
 #include "hvx-log.h"
+#include "hvx-norm.h"
 
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -667,7 +667,7 @@ static int execute_op(struct htp_ops_context * octx) {
             return op_gated_delta_net(octx);
 
         case HTP_OP_TRI:
-            return op_tri(octx);
+            return op_unary(octx);
 
         case HTP_OP_INVALID:
             break;

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

@@ -0,0 +1,162 @@
+#ifndef HTP_UNARY_OPS_H
+#define HTP_UNARY_OPS_H
+
+#include "hex-common.h"
+#include "htp-ops.h"
+
+// Op-specific struct for precomputed unary params
+struct htp_unary_kernel_params {
+    uint32_t  n_threads;
+    uint32_t  col_tile;
+    uint32_t  vtcm_row_per_thread;
+    uint32_t  block;
+    uint32_t  broadcast_weight;
+
+    uint32_t  vtcm_src0_size_per_thread;
+    uint32_t  vtcm_src1_size_per_thread;
+    uint32_t  vtcm_dst_size_per_thread;
+
+    uint32_t  vtcm_src0_size;
+    uint32_t  vtcm_src1_size;
+    uint32_t  vtcm_dst_size;
+
+    uint32_t  src0_row_size_aligned;
+    uint32_t  src1_row_size_aligned;
+    uint32_t  dst_row_size_aligned;
+
+    uint32_t  vtcm_size;
+
+    // Fastdiv helpers
+    struct fastdiv_values div_ne01;
+    struct fastdiv_values div_ne02;
+    struct fastdiv_values div_ne012;
+    struct fastdiv_values div_tpr;
+};
+
+#if defined(__cplusplus)
+static_assert(sizeof(struct htp_unary_kernel_params) <= 128, "htp_unary_kernel_params is too large for kernel_params blob");
+#else
+_Static_assert(sizeof(struct htp_unary_kernel_params) <= 128, "htp_unary_kernel_params is too large for kernel_params blob");
+#endif
+
+static inline bool htp_op_is_unary(uint32_t opcode) {
+    switch (opcode) {
+        case HTP_OP_NORM:
+        case HTP_OP_RMS_NORM:
+        case HTP_OP_RMS_NORM_MUL:
+        case HTP_OP_SCALE:
+        case HTP_OP_SQR:
+        case HTP_OP_SQRT:
+        case HTP_OP_UNARY_NEG:
+        case HTP_OP_UNARY_EXP:
+        case HTP_OP_UNARY_SIGMOID:
+        case HTP_OP_UNARY_SOFTPLUS:
+        case HTP_OP_UNARY_TANH:
+        case HTP_OP_L2_NORM:
+        case HTP_OP_TRI:
+            return true;
+        default:
+            return false;
+    }
+}
+
+struct htp_unary_vtcm_layout {
+    size_t total_bytes;
+    size_t off_src0;
+    size_t off_src1;
+    size_t off_dst;
+
+    size_t src0_bytes;
+    size_t src1_bytes;
+    size_t dst_bytes;
+};
+
+static inline void htp_unary_vtcm_layout_build(
+    struct htp_unary_vtcm_layout * L,
+    uint32_t op,
+    uint32_t ne00,
+    uint32_t ne10,
+    uint32_t ne11,
+    bool broadcast_weight,
+    uint32_t n_threads,
+    size_t vtcm_size,
+    uint32_t * out_col_tile,
+    uint32_t * out_vtcm_row_per_thread
+) {
+    const size_t src0_data_row_size = ne00 * sizeof(float);
+    const size_t dst_data_row_size  = ne10 * sizeof(float);
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);
+
+    size_t src1_row_size_aligned = 0;
+    if (op == HTP_OP_RMS_NORM_MUL) {
+        const size_t src1_data_row_size = ne11 * sizeof(float);
+        src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
+    }
+
+    size_t vtcm_size_per_row = 0;
+    size_t vtcm_row_per_thread = 0;
+
+    if (op == HTP_OP_RMS_NORM_MUL) {
+        if (broadcast_weight) {
+            size_t available_vtcm = vtcm_size;
+            size_t src1_vtcm_total = n_threads * src1_row_size_aligned;
+            if (available_vtcm > src1_vtcm_total) {
+                available_vtcm -= src1_vtcm_total;
+            } else {
+                available_vtcm = 0;
+            }
+            vtcm_size_per_row = 2 * (src0_row_size_aligned + dst_row_size_aligned);
+            vtcm_row_per_thread = available_vtcm / (n_threads * vtcm_size_per_row);
+        } else {
+            vtcm_size_per_row = 2 * (src0_row_size_aligned + dst_row_size_aligned + src1_row_size_aligned);
+            vtcm_row_per_thread = vtcm_size / (n_threads * vtcm_size_per_row);
+        }
+    } else {
+        vtcm_size_per_row   = 2 * (src0_row_size_aligned + dst_row_size_aligned);
+        vtcm_row_per_thread = vtcm_size / (n_threads * vtcm_size_per_row);
+    }
+
+    const bool is_reduction = (op == HTP_OP_NORM || op == HTP_OP_RMS_NORM ||
+                               op == HTP_OP_RMS_NORM_MUL || op == HTP_OP_L2_NORM);
+    uint32_t col_tile = 0;
+
+    if (vtcm_row_per_thread == 0 && !is_reduction) {
+        const size_t per_thread_budget = vtcm_size / n_threads;
+        const size_t col_tile_bytes = hex_align_down(per_thread_budget / 4, 128);
+        col_tile = (uint32_t) (col_tile_bytes / sizeof(float));
+
+        L->src0_bytes = col_tile_bytes * 2;
+        L->dst_bytes  = col_tile_bytes * 2;
+        L->src1_bytes = 0;
+    } else {
+        L->src0_bytes = src0_row_size_aligned * vtcm_row_per_thread * 2;
+        L->dst_bytes  = dst_row_size_aligned * vtcm_row_per_thread * 2;
+        if (op == HTP_OP_RMS_NORM_MUL) {
+            if (broadcast_weight) {
+                L->src1_bytes = src1_row_size_aligned;
+            } else {
+                L->src1_bytes = src1_row_size_aligned * vtcm_row_per_thread * 2;
+            }
+        } else {
+            L->src1_bytes = 0;
+        }
+    }
+
+    L->off_src0 = 0;
+    if (op == HTP_OP_RMS_NORM_MUL) {
+        L->off_src1 = L->off_src0 + L->src0_bytes * n_threads;
+        L->off_dst  = L->off_src1 + L->src1_bytes * n_threads;
+    } else {
+        L->off_src1 = 0;
+        L->off_dst  = L->off_src0 + L->src0_bytes * n_threads;
+    }
+
+    L->total_bytes = L->off_dst + L->dst_bytes * n_threads;
+
+    *out_col_tile = col_tile;
+    *out_vtcm_row_per_thread = vtcm_row_per_thread;
+}
+
+#endif /* HTP_UNARY_OPS_H */

tools/cli/cli-context.cpp

@@ -162,6 +162,14 @@ bool cli_context::init() {
 
     fetch_server_props();
 
+    if (!params.out_file.empty()) {
+        output_file.emplace(params.out_file);
+        if (!output_file->is_open()) {
+            ui::show_error(string_format("failed to open output file '%s'", params.out_file.c_str()));
+            return false;
+        }
+    }
+
     return true;
 }
 
@@ -323,7 +331,14 @@ bool cli_context::stage_media_file(const std::string & fname, const std::string
     return true;
 }
 
-bool cli_context::generate_completion(std::string & assistant_content, cli_timings & timings) {
+void cli_context::write_output_file(const std::string & content) {
+    if (output_file) {
+        (*output_file) << content;
+        output_file->flush();
+    }
+}
+
+bool cli_context::generate_completion(generated_content & content_out, cli_timings & timings) {
     json body = {
         {"messages",          impl->messages},
         {"stream",            true},
@@ -364,13 +379,14 @@ bool cli_context::generate_completion(std::string & assistant_content, cli_timin
         if (delta.contains("reasoning_content") && delta.at("reasoning_content").is_string()) {
             const std::string text = delta.at("reasoning_content").get<std::string>();
             if (!text.empty()) {
+                content_out.reasoning += text;
                 a.push(ui::ASSISTANT_DISPLAY_MODE_REASONING, text);
             }
         }
         if (delta.contains("content") && delta.at("content").is_string()) {
             const std::string text = delta.at("content").get<std::string>();
             if (!text.empty()) {
-                assistant_content += text;
+                content_out.content += text;
                 a.push(ui::ASSISTANT_DISPLAY_MODE_CONTENT, text);
             }
         }
@@ -520,10 +536,12 @@ int cli_context::run() {
                 continue;
             }
             ui::show_message(string_format("Loaded media from '%s'", fname.c_str()));
+            write_output_file(string_format("User: Added media: %s\n", fname.c_str()));
             continue;
         } else if (string_starts_with(buffer, "/read ")) {
             std::string fname = string_strip(buffer.substr(6));
             add_text_file(fname);
+            write_output_file(string_format("User: Added text file: %s\n", fname.c_str()));
             continue;
         } else if (string_starts_with(buffer, "/glob ")) {
             std::error_code ec;
@@ -568,9 +586,11 @@ int cli_context::run() {
                     continue;
                 }
 
-                if (!add_text_file((rel_path / rel).string())) {
+                const std::string full_path = (curdir / rel).string();
+                if (!add_text_file(full_path)) {
                     continue;
                 }
+                write_output_file(string_format("User: Added text file: %s\n", full_path.c_str()));
 
                 if (++count >= FILE_GLOB_MAX_RESULTS) {
                     ui::show_error(string_format("Maximum number of globbed files allowed (%zu) reached.", FILE_GLOB_MAX_RESULTS));
@@ -586,16 +606,34 @@ int cli_context::run() {
         // generate response
         if (add_user_msg) {
             push_user_message(cur_msg);
+            write_output_file(string_format("User:\n%s\n\n", cur_msg.c_str()));
             cur_msg.clear();
         }
+
         cli_timings timings;
-        std::string assistant_content;
-        generate_completion(assistant_content, timings);
+        generated_content content;
+        generate_completion(content, timings);
+
         impl->messages.push_back({
             {"role",    "assistant"},
-            {"content", assistant_content}
+            {"content", content.content}
         });
 
+        if (output_file) {
+            std::string out_content = "Assistant:\n";
+            if (!content.reasoning.empty()) {
+                out_content += "[Start thinking]\n\n";
+                out_content += content.reasoning;
+                out_content += "[End thinking]\n\n";
+            }
+            out_content += content.content;
+            if (!out_content.empty() && out_content.back() != '\n') {
+                out_content += "\n";
+            }
+            out_content += "\n";
+            write_output_file(out_content);
+        }
+
         if (params.show_timings) {
             ui::show_info(string_format(
                 "\n[ Prompt: %.1f t/s | Generation: %.1f t/s ]",
@@ -619,4 +657,8 @@ void cli_context::shutdown() {
         server->stop();
         server.reset();
     }
+    if (output_file) {
+        output_file->close();
+        output_file.reset();
+    }
 }

tools/cli/cli-context.h

@@ -9,6 +9,7 @@
 #include <memory>
 #include <optional>
 #include <string>
+#include <fstream>
 
 struct cli_timings {
     double prompt_per_second    = 0.0;
@@ -32,6 +33,8 @@ struct cli_context {
     bool has_audio  = false;
     bool has_video  = false;
 
+    std::optional<std::ofstream> output_file;
+
     cli_context(const common_params & params);
     ~cli_context();
 
@@ -49,7 +52,11 @@ struct cli_context {
     static std::atomic<bool> & interrupted();
 
 private:
-    bool generate_completion(std::string & assistant_content, cli_timings & timings);
+    struct generated_content {
+        std::string reasoning;
+        std::string content;
+    };
+    bool generate_completion(generated_content & content_out, cli_timings & timings);
     void fetch_server_props();
     void add_system_prompt();
     void push_user_message(const std::string & text);
@@ -62,5 +69,8 @@ private:
     // "image", "audio", "video"; returns false if the file cannot be read
     bool stage_media_file(const std::string & fname, const std::string & type);
 
+    // no-op if output file is not set
+    void write_output_file(const std::string & content);
+
     std::unique_ptr<cli_context_impl> impl;
 };