Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

metal: accelerated conv2d (#17175)

Browse Source 
* metal: accelerated conv2d

* cont : cleanup

---------

Co-authored-by: bghira <bghira@users.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
bagheera
2025-11-13 13:32:44 +02:00
committed by GitHub
co-authored by bghira Georgi Gerganov
parent 2776db6c81
commit 0cfb19166b
7 changed files with 258 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+83 -5
View File
@@ -10,6 +10,7 @@
#include <cassert>
#include <algorithm>
#include <limits>
static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
if (!t) {
@@ -364,6 +365,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_im2col(ctx, idx);
} break;
case GGML_OP_CONV_2D:
{
n_fuse = ggml_metal_op_conv_2d(ctx, idx);
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{
n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
@@ -1036,11 +1041,6 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
nth = std::min(nth, nk0);
if (nth*nrptg > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
nrptg = 1;
}
ggml_metal_kargs_set_rows args = {
/*.nk0 =*/ nk0,
/*.ne01 =*/ ne01,
@@ -3082,6 +3082,84 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);
ggml_metal_library_t lib = ctx->lib;
ggml_metal_encoder_t enc = ctx->enc;
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
GGML_ASSERT(ggml_is_contiguous(op->src[0]));
GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
GGML_ASSERT(op->type == GGML_TYPE_F32);
GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
const int32_t s0 = ((const int32_t *) op->op_params)[0];
const int32_t s1 = ((const int32_t *) op->op_params)[1];
const int32_t p0 = ((const int32_t *) op->op_params)[2];
const int32_t p1 = ((const int32_t *) op->op_params)[3];
const int32_t d0 = ((const int32_t *) op->op_params)[4];
const int32_t d1 = ((const int32_t *) op->op_params)[5];
ggml_metal_kargs_conv_2d args = {
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.nb10 =*/ nb10,
/*.nb11 =*/ nb11,
/*.nb12 =*/ nb12,
/*.nb13 =*/ nb13,
/*.nb0 =*/ nb0,
/*.nb1 =*/ nb1,
/*.nb2 =*/ nb2,
/*.nb3 =*/ nb3,
/*.IW =*/ ne10,
/*.IH =*/ ne11,
/*.KW =*/ ne00,
/*.KH =*/ ne01,
/*.IC =*/ ne02,
/*.OC =*/ ne03,
/*.OW =*/ ne0,
/*.OH =*/ ne1,
/*.N =*/ ne3,
/*.s0 =*/ s0,
/*.s1 =*/ s1,
/*.p0 =*/ p0,
/*.p1 =*/ p1,
/*.d0 =*/ d0,
/*.d1 =*/ d1,
};
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_2d(lib, op);
int nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
nth = std::min(nth, 256);
nth = std::max(nth, 1);
const uint64_t n_out = ggml_nelements(op);
uint64_t tg = (n_out + nth - 1)/nth;
tg = std::max<uint64_t>(tg, 1);
tg = std::min<uint64_t>(tg, (uint64_t) std::numeric_limits<int>::max());
ggml_metal_encoder_set_pipeline(enc, pipeline);
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 3);
ggml_metal_encoder_dispatch_threadgroups(enc, tg, 1, 1, nth, 1, 1);
return 1;
}
int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);