llama-cpp-bin 9326.0.0__tar.gz → 9352.0.0__tar.gz

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9352.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9326.0.0
+Version: 9352.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9352.0.0}/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -2874,6 +2874,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_NORM:            return HTP_OP_NORM;
         case GGML_OP_L2_NORM:         return HTP_OP_L2_NORM;
         case GGML_OP_RMS_NORM:        return HTP_OP_RMS_NORM;
+        case GGML_OP_CONCAT:          return HTP_OP_CONCAT;
         case GGML_OP_SCALE:           return HTP_OP_SCALE;
         case GGML_OP_SQR:             return HTP_OP_SQR;
         case GGML_OP_SQRT:            return HTP_OP_SQRT;
@@ -3286,6 +3287,25 @@ static bool ggml_hexagon_supported_repeat(const struct ggml_hexagon_session * se
     return true;
 }
 
+static bool ggml_hexagon_supported_concat(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    int dim = ((const int32_t *) op->op_params)[0];
+    if (dim < 0 || dim >= GGML_MAX_DIMS) {
+        return false;
+    }
+
+    for (int i = 0; i < GGML_MAX_SRC; ++i) {
+        const struct ggml_tensor * src = op->src[i];
+        if (!src) {
+            continue;
+        }
+        if (src->type != GGML_TYPE_F32 && src->type != GGML_TYPE_I32 && src->type != GGML_TYPE_F16) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_fill(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const struct ggml_tensor * dst = op;
 
@@ -3434,6 +3454,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_cumsum(sess, op);
             break;
 
+        case GGML_OP_CONCAT:
+            supp = ggml_hexagon_supported_concat(sess, op);
+            break;
+
         case GGML_OP_FILL:
             supp = ggml_hexagon_supported_fill(sess, op);
             break;

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9352.0.0}/llama.cpp/ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -35,6 +35,7 @@ add_library(${HTP_LIB} SHARED
     ssm-conv.c
     cumsum-ops.c
     fill-ops.c
+    concat-ops.c
     diag-ops.c
     solve-tri-ops.c
     gated-delta-net-ops.c

llama_cpp_bin-9352.0.0/llama.cpp/ggml/src/ggml-hexagon/htp/concat-ops.c

@@ -0,0 +1,275 @@
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "hexagon_types.h"
+#include "hexagon_protos.h"
+#include "hvx_hexagon_protos.h"
+#include "hex-dma.h"
+#include "vtcm-utils.h"
+#include "hvx-utils.h"
+#include "hex-fastdiv.h"
+#include <string.h>
+
+struct htp_concat_context {
+    struct htp_ops_context * octx;
+    uint32_t dim;
+    uint32_t nrows_per_thread;
+    struct fastdiv_values div_ne0;
+    struct fastdiv_values div_ne1;
+    struct fastdiv_values div_ne2;
+};
+
+static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const uint32_t src0_ne0 = src0->ne[0];
+    const uint32_t src1_ne0 = src1->ne[0];
+    const uint32_t ne1      = dst->ne[1];
+
+    const uint32_t start_i = ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    if (start_i >= end_i) return;
+
+    dma_queue * q = octx->ctx->dma[ith];
+
+    uint8_t * spad0_base = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread;
+    uint8_t * spad1_base = octx->src1_spad.data + ith * octx->src1_spad.size_per_thread;
+
+    const uint32_t block_i = 32;
+    const uint32_t spad1_stride = block_i * sizeof(float);
+
+    int32_t offsets[32] __attribute__((aligned(128)));
+    for(int k=0; k<32; k++) {
+        offsets[k] = k * spad1_stride;
+    }
+    HVX_Vector vv = *(HVX_Vector*)offsets;
+    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 32);
+    const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(float), VLEN);
+    uint32_t mu = src1_ne0_padded * spad1_stride;
+
+    for (uint32_t i = start_i; i < end_i; i += block_i) {
+        uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
+
+        uint32_t src1_width_bytes = current_block_i * sizeof(float);
+        uint8_t * src1_ptr = (uint8_t *)src1->data + i * src1->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad1_base, src1_ptr), spad1_stride, src1->nb[0], src1_width_bytes, src1_ne0);
+
+        uint32_t src0_row_bytes = src0_ne0 * sizeof(float);
+        uint8_t * src0_ptr = (uint8_t *)src0->data + i * src0->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad0_base, src0_ptr), spad0_row_bytes, src0->nb[1], src0_row_bytes, current_block_i);
+
+        dma_queue_pop(q); // src1
+
+        HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
+
+        for (uint32_t j = 0; j < src1_ne0_padded; j += 32) {
+            #pragma unroll(4)
+            for (uint32_t ii = 0; ii < current_block_i; ii++) {
+                size_t rt = (size_t)(spad1_base + j * spad1_stride + ii * sizeof(float));
+                Q6_vgather_ARMVw(&vtcm_tmp[ii], rt, mu, vv);
+                uint8_t * dst_ptr = spad0_base + ii * spad0_row_bytes + (src0_ne0 + j) * sizeof(float);
+                hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
+            }
+        }
+
+        dma_queue_pop(q); // src0
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i * dst->nb[1];
+        dma_queue_push(q, dma_make_ptr(dst_ptr, spad0_base), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(float), current_block_i);
+
+        dma_queue_pop(q);
+    }
+}
+
+static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const uint32_t src0_ne0 = src0->ne[0];
+    const uint32_t src1_ne0 = src1->ne[0];
+    const uint32_t ne1      = dst->ne[1];
+
+    const uint32_t start_i = ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    if (start_i >= end_i) return;
+
+    dma_queue * q = octx->ctx->dma[ith];
+
+    uint8_t * spad0_base = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread;
+    uint8_t * spad1_base = octx->src1_spad.data + ith * octx->src1_spad.size_per_thread;
+
+    const uint32_t block_i = 64;
+    const uint32_t spad1_stride = block_i * sizeof(__fp16);
+
+    int16_t offsets[64] __attribute__((aligned(128)));
+    for(int k=0; k<64; k++) {
+        offsets[k] = k * spad1_stride;
+    }
+    HVX_Vector vv = *(HVX_Vector*)offsets;
+    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 64);
+    const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(__fp16), VLEN);
+    uint32_t mu = src1_ne0_padded * spad1_stride;
+
+    for (uint32_t i = start_i; i < end_i; i += block_i) {
+        uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
+
+        uint32_t src1_width_bytes = current_block_i * sizeof(__fp16);
+        uint8_t * src1_ptr = (uint8_t *)src1->data + i * src1->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad1_base, src1_ptr), spad1_stride, src1->nb[0], src1_width_bytes, src1_ne0);
+
+        uint32_t src0_row_bytes = src0_ne0 * sizeof(__fp16);
+        uint8_t * src0_ptr = (uint8_t *)src0->data + i * src0->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad0_base, src0_ptr), spad0_row_bytes, src0->nb[1], src0_row_bytes, current_block_i);
+
+        dma_queue_pop(q); // src1
+
+        HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
+
+        for (uint32_t j = 0; j < src1_ne0_padded; j += 64) {
+            #pragma unroll(4)
+            for (uint32_t ii = 0; ii < current_block_i; ii++) {
+                size_t rt = (size_t)(spad1_base + j * spad1_stride + ii * sizeof(__fp16));
+                Q6_vgather_ARMVh(&vtcm_tmp[ii], rt, mu, vv);
+                uint8_t * dst_ptr = spad0_base + ii * spad0_row_bytes + (src0_ne0 + j) * sizeof(__fp16);
+                hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
+            }
+        }
+
+        dma_queue_pop(q); // src0
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i * dst->nb[1];
+        dma_queue_push(q, dma_make_ptr(dst_ptr, spad0_base), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(__fp16), current_block_i);
+
+        dma_queue_pop(q);
+    }
+}
+
+static void concat_generic(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const int dim = cctx->dim;
+    const uint32_t type_size = (dst->type == HTP_TYPE_F32 || dst->type == HTP_TYPE_I32) ? 4 : 2;
+
+    const uint32_t ne[4] = {dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]};
+    const uint32_t total_elements = ne[0] * ne[1] * ne[2] * ne[3];
+    const uint32_t chunk_size = (total_elements + nth - 1) / nth;
+
+    const uint32_t start_idx = MIN(ith * chunk_size, total_elements);
+    const uint32_t end_idx   = MIN(start_idx + chunk_size, total_elements);
+
+    // Naive scalar element-wise copy
+    for (uint32_t idx = start_idx; idx < end_idx; idx++) {
+        uint32_t idx_div_ne0 = fastdiv(idx, &cctx->div_ne0);
+        uint32_t i0 = idx - idx_div_ne0 * ne[0];
+
+        uint32_t idx_div_ne01 = fastdiv(idx_div_ne0, &cctx->div_ne1);
+        uint32_t i1 = idx_div_ne0 - idx_div_ne01 * ne[1];
+
+        uint32_t idx_div_ne012 = fastdiv(idx_div_ne01, &cctx->div_ne2);
+        uint32_t i2 = idx_div_ne01 - idx_div_ne012 * ne[2];
+        uint32_t i3 = idx_div_ne012;
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2] + i1 * dst->nb[1] + i0 * dst->nb[0];
+
+        uint32_t idx_dim = 0;
+        if (dim == 0) idx_dim = i0;
+        else if (dim == 1) idx_dim = i1;
+        else if (dim == 2) idx_dim = i2;
+        else if (dim == 3) idx_dim = i3;
+
+        const struct htp_tensor * src = (idx_dim < src0->ne[dim]) ? src0 : src1;
+
+        uint32_t s0 = i0;
+        uint32_t s1 = i1;
+        uint32_t s2 = i2;
+        uint32_t s3 = i3;
+
+        if (dim == 0 && src == src1) s0 -= src0->ne[0];
+        if (dim == 1 && src == src1) s1 -= src0->ne[1];
+        if (dim == 2 && src == src1) s2 -= src0->ne[2];
+        if (dim == 3 && src == src1) s3 -= src0->ne[3];
+
+        uint8_t * src_ptr = (uint8_t *)src->data + s3 * src->nb[3] + s2 * src->nb[2] + s1 * src->nb[1] + s0 * src->nb[0];
+
+        if (type_size == 4) {
+            *(float*)dst_ptr = *(float*)src_ptr;
+        } else {
+            *(__fp16*)dst_ptr = *(__fp16*)src_ptr;
+        }
+    }
+}
+
+int op_concat(struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    int dim = octx->op_params[0];
+
+    bool is_2d = dst->ne[2] == 1 && dst->ne[3] == 1;
+
+    const uint32_t type_size = (dst->type == HTP_TYPE_F32 || dst->type == HTP_TYPE_I32) ? 4 : 2;
+    bool is_src1_transposed  = (src1->nb[0] > src1->nb[1]);
+    bool is_src0_transposed  = (src0->nb[0] > src0->nb[1]);
+
+    uint32_t n_threads = octx->n_threads;
+    struct htp_concat_context cctx;
+    cctx.octx = octx;
+    cctx.dim = dim;
+    cctx.div_ne0 = init_fastdiv_values(dst->ne[0]);
+    cctx.div_ne1 = init_fastdiv_values(dst->ne[1]);
+    cctx.div_ne2 = init_fastdiv_values(dst->ne[2]);
+
+    void (*worker_func)(unsigned int, unsigned int, void *) = concat_generic;
+
+    if (dim == 0 && is_2d && is_src1_transposed && !is_src0_transposed) {
+        n_threads = MIN(dst->ne[1], n_threads);
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        uint32_t block_i = (type_size == 4) ? 32 : 64;
+
+        cctx.nrows_per_thread = hmx_ceil_div(dst->ne[1], n_threads);
+
+        // Allocate VTCM
+        uint32_t spad1_stride = block_i * type_size;
+
+        uint32_t src1_ne0_padded = hex_round_up(src1->ne[0], block_i);
+        uint32_t spad0_row_bytes = hex_round_up((src0->ne[0] + src1_ne0_padded) * type_size, VLEN);
+
+        octx->src0_spad.size_per_thread = block_i * spad0_row_bytes;
+        octx->src1_spad.size_per_thread = src1_ne0_padded * spad1_stride + block_i * VLEN;
+
+        octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
+        octx->src1_spad.size = n_threads * octx->src1_spad.size_per_thread;
+
+        if (octx->src0_spad.size + octx->src1_spad.size > octx->ctx->vtcm_size) {
+            return HTP_STATUS_VTCM_TOO_SMALL;
+        }
+
+        octx->src0_spad.data = octx->ctx->vtcm_base;
+        octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+
+        if (type_size == 4) {
+            worker_func = concat_2d_f32_transposed;
+        } else {
+            worker_func = concat_2d_f16_transposed;
+        }
+    }
+
+    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &cctx, n_threads);
+    return HTP_STATUS_OK;
+}

llama_cpp_bin-9352.0.0/llama.cpp/ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -0,0 +1,295 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+
+struct htp_copy_context {
+    struct htp_ops_context * octx;
+
+    uint32_t          src0_type_size;
+    uint32_t          src0_block_size;
+
+    uint32_t          dst_type_size;
+    uint32_t          dst_block_size;
+
+    uint32_t          src0_blocks_per_row;
+    uint32_t          dst_blocks_per_row;
+
+    uint32_t          src0_nrows_per_thread;
+};
+
+#define cpy_preamble                              \
+    const struct htp_tensor *src0 = octx->src[0]; \
+    const struct htp_tensor *dst  = octx->dst;    \
+                                                  \
+    const uint32_t ne00 = src0->ne[0];            \
+    const uint32_t ne01 = src0->ne[1];            \
+    const uint32_t ne02 = src0->ne[2];            \
+    const uint32_t ne03 = src0->ne[3];            \
+                                                  \
+    const uint32_t nb00 = src0->nb[0];            \
+    const uint32_t nb01 = src0->nb[1];            \
+    const uint32_t nb02 = src0->nb[2];            \
+    const uint32_t nb03 = src0->nb[3];            \
+                                                  \
+    const uint32_t  ne0 = dst->ne[0];             \
+    const uint32_t  ne1 = dst->ne[1];             \
+    const uint32_t  ne2 = dst->ne[2];             \
+    const uint32_t  ne3 = dst->ne[3];             \
+                                                  \
+    const uint32_t  nb0 = dst->nb[0];             \
+    const uint32_t  nb1 = dst->nb[1];             \
+    const uint32_t  nb2 = dst->nb[2];             \
+    const uint32_t  nb3 = dst->nb[3];             \
+                                                  \
+    const uint32_t   nr = ne01;
+
+#define DEFINE_CPY_SAMESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                       \
+static void cpy_thread_##NAME##_sameshape(unsigned int nth, unsigned int ith, void * data) {                   \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
+    struct htp_ops_context * octx = ct->octx;                                                                  \
+    cpy_preamble;                                                                                              \
+    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
+    const uint32_t ir0 = dr * ith;                                                                             \
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
+    if (ir0 >= nr) return;                                                                                     \
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {                                                                \
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
+            _Pragma("unroll(4)")                                                                               \
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;                     \
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;                    \
+                hex_l2fetch(src0_ptr, ne00 * ELEM_SIZE, nb01, 2);                                              \
+                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
+            }                                                                                                  \
+        }                                                                                                      \
+    }                                                                                                          \
+}
+
+DEFINE_CPY_SAMESHAPE(f32, float, 4)
+DEFINE_CPY_SAMESHAPE(f16, __fp16, 2)
+
+#define DEFINE_CPY_RESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                         \
+static void cpy_thread_##NAME##_reshape(unsigned int nth, unsigned int ith, void * data) {                     \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
+    struct htp_ops_context * octx = ct->octx;                                                                  \
+    cpy_preamble;                                                                                              \
+    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
+    const uint32_t ir0 = dr * ith;                                                                             \
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
+    if (ir0 >= nr) return;                                                                                     \
+    const bool src0_contig = (nb00 == ELEM_SIZE)   &&                                                          \
+                             (nb01 == ne00 * nb00) &&                                                          \
+                             (nb02 == ne01 * nb01) &&                                                          \
+                             (nb03 == ne02 * nb02);                                                            \
+    const bool dst_contig  = (nb0  == ELEM_SIZE)   &&                                                          \
+                             (nb1  == ne0  * nb0)  &&                                                          \
+                             (nb2  == ne1  * nb1)  &&                                                          \
+                             (nb3  == ne2  * nb2);                                                             \
+    if (src0_contig && dst_contig) {                                                                           \
+        for (int64_t i03 = 0; i03 < ne03; i03++) {                                                             \
+            for (int64_t i02 = 0; i02 < ne02; i02++) {                                                         \
+                uint8_t * src_ptr = (uint8_t *) src0->data + i03*nb03 + i02*nb02 + ir0*nb01;                   \
+                uint32_t  flat    = ((i03*ne02 + i02)*ne01 + ir0) * ne00;                                      \
+                uint8_t * dst_ptr = (uint8_t *) dst->data  + flat * ELEM_SIZE;                                 \
+                hvx_copy_uu(dst_ptr, src_ptr, (ir1 - ir0) * ne00, ELEM_SIZE);                                  \
+            }                                                                                                  \
+        }                                                                                                      \
+        return;                                                                                                \
+    }                                                                                                          \
+    const bool reshape_flat_fast = (ne03 == 1 && ne2 == 1 && ne3 == 1) &&                                      \
+                                   (ne0 == ne00 * ne01) && (ne1 == ne02) &&                                    \
+                                   (nb00 == ELEM_SIZE) && (nb0 == ELEM_SIZE);                                  \
+    if (reshape_flat_fast) {                                                                                   \
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
+                uint8_t * src0_ptr = (uint8_t *) src0->data + i01 * nb01 + i02 * nb02;                         \
+                uint8_t * dst_ptr  = (uint8_t *) dst->data  + i01 * ne00 * ELEM_SIZE + i02 * nb1;              \
+                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
+            }                                                                                                  \
+        }                                                                                                      \
+        return;                                                                                                \
+    }                                                                                                          \
+    int64_t k10 = 0;                                                                                           \
+    int64_t i11 = 0;                                                                                           \
+    int64_t i12 = 0;                                                                                           \
+    int64_t i13 = 0;                                                                                           \
+    const int64_t nk00 = ct->src0_blocks_per_row;                                                              \
+    const int64_t nk0  = ct->dst_blocks_per_row;                                                               \
+    for (int64_t i03 = 0; i03 < ne03; i03++) {                                                                 \
+        for (int64_t i02 = 0; i02 < ne02; i02++) {                                                             \
+            k10 += nk00 * ir0;                                                                                 \
+            while (k10 >= nk0) {                                                                               \
+                k10 -= nk0;                                                                                    \
+                if (++i11 == ne1) {                                                                            \
+                    i11 = 0;                                                                                   \
+                    if (++i12 == ne2) {                                                                        \
+                        i12 = 0;                                                                               \
+                        if (++i13 == ne3) {                                                                    \
+                            i13 = 0;                                                                           \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+            for (int64_t i01 = ir0; i01 < ir1; i01++) {                                                        \
+                for (int64_t k00 = 0; k00 < nk00; k00++) {                                                     \
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); \
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);  \
+                    memcpy(dst_ptr, src0_ptr, ELEM_SIZE);                                                      \
+                    if (++k10 == nk0) {                                                                        \
+                        k10 = 0;                                                                               \
+                        if (++i11 == ne1) {                                                                    \
+                            i11 = 0;                                                                           \
+                            if (++i12 == ne2) {                                                                \
+                                i12 = 0;                                                                       \
+                                if (++i13 == ne3) {                                                            \
+                                    i13 = 0;                                                                   \
+                                }                                                                              \
+                            }                                                                                  \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+            k10 += nk00 * (ne01 - ir1);                                                                        \
+            while (k10 >= nk0) {                                                                               \
+                k10 -= nk0;                                                                                    \
+                if (++i11 == ne1) {                                                                            \
+                    i11 = 0;                                                                                   \
+                    if (++i12 == ne2) {                                                                        \
+                        i12 = 0;                                                                               \
+                        if (++i13 == ne3) {                                                                    \
+                            i13 = 0;                                                                           \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+        }                                                                                                      \
+    }                                                                                                          \
+}
+
+DEFINE_CPY_RESHAPE(f32, float, 4)
+DEFINE_CPY_RESHAPE(f16, __fp16, 2)
+
+static void cpy_thread_f16_f32_sameshape(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;
+    struct htp_ops_context * octx = ct->octx;
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+    if (ir0 >= nr) return;
+
+    // copy by rows
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            #pragma unroll(2)
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                hex_l2fetch(src0_ptr, ne00 * sizeof(float), nb01, 2);
+                hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
+            }
+        }
+    }
+}
+
+static void cpy_thread_f32_f16_sameshape(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;
+    struct htp_ops_context * octx = ct->octx;
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+    if (ir0 >= nr) return;
+
+    // copy by rows
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            #pragma unroll(2)
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                hex_l2fetch(src0_ptr, ne00 * sizeof(__fp16), nb01, 2);
+                hvx_copy_f32_f16_uu(dst_ptr, src0_ptr, ne00);
+            }
+        }
+    }
+}
+
+int op_cpy(struct htp_ops_context * octx) {
+    cpy_preamble;
+
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
+    struct htp_copy_context ct;
+    ct.octx = octx;
+
+    switch (src0->type) {
+    case HTP_TYPE_F32: ct.src0_type_size = 4; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
+    case HTP_TYPE_F16: ct.src0_type_size = 2; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    switch (dst->type) {
+    case HTP_TYPE_F32: ct.dst_type_size = 4; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
+    case HTP_TYPE_F16: ct.dst_type_size = 2; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const bool sametype   = (src0->type == dst->type);
+    const bool transposed = (nb00 > nb01) || (nb0 > nb1);
+    const bool sameshape  = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
+
+    ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
+
+    worker_callback_t copy_fun;
+
+    if (sametype && sameshape) {
+        if (src0->type == HTP_TYPE_F32) {
+            copy_fun = cpy_thread_f32_sameshape;
+        } else {
+            copy_fun = cpy_thread_f16_sameshape;
+        }
+    } else if (sameshape) {
+        /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
+            copy_fun = cpy_thread_f16_f32_sameshape;
+        else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16)
+            copy_fun = cpy_thread_f32_f16_sameshape;
+        else
+            return HTP_STATUS_NO_SUPPORT;
+    } else if (sametype) {
+        if (src0->type == HTP_TYPE_F32) {
+            copy_fun = cpy_thread_f32_reshape;
+        } else {
+            copy_fun = cpy_thread_f16_reshape;
+        }
+    } else {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
+
+    return HTP_STATUS_OK;
+}