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

metal : add CONV_TRANSPOSE_2D (#16542)

Browse Source 
* initial: headers and metal-device.cpp updates

* adding conv_transpose_2d

* fix type

* fix type: int32->int64

* Update ggml/src/ggml-metal/ggml-metal.metal

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml-metal/ggml-metal.metal

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml-metal/ggml-metal.metal

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* add checks for src[0] and src[1]; add type checks

* Update ggml-metal.metal

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* add more tests, add optimization to threading

* add dynamic memory allocation in metal

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Ilia Ilmer
2025-10-17 09:33:58 +03:00
committed by GitHub
co-authored by Georgi Gerganov
parent 79967ec596
commit 9ad4f1931e
8 changed files with 198 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+60
View File
@@ -368,6 +368,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
n_fuse = ggml_metal_op_conv_transpose_2d(ctx, idx);
} break;
case GGML_OP_UPSCALE:
{
n_fuse = ggml_metal_op_upscale(ctx, idx);
@@ -3118,6 +3122,62 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_conv_transpose_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(uint32_t, nb, op, nb);
const int32_t s0 = ((const int32_t *)(op->op_params))[0];
const int32_t IC = op->src[1]->ne[2];
const int32_t IH = op->src[1]->ne[1];
const int32_t IW = op->src[1]->ne[0];
const int32_t KH = op->src[0]->ne[1];
const int32_t KW = op->src[0]->ne[0];
const int32_t OW = op->ne[0];
const int32_t OH = op->ne[1];
const int32_t OC = op->ne[2];
ggml_metal_kargs_conv_transpose_2d args = {
/*.IC =*/ IC,
/*.IH =*/ IH,
/*.IW =*/ IW,
/*.KH =*/ KH,
/*.KW =*/ KW,
/*.OC =*/ OC,
/*.s0 =*/ s0,
/*.nb0 =*/ nb0,
/*.nb1 =*/ nb1,
/*.nb2 =*/ nb2,
};
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_transpose_2d(lib, op);
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);
// Metal requires buffer size to be multiple of 16 bytes
const size_t smem = GGML_PAD(KW * KH * sizeof(float), 16);
ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
ggml_metal_encoder_dispatch_threadgroups(enc, OW, OH, OC, KW, KH, 1);
return 1;
}
int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);