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

metal: add CONV_3D (#19927)

Browse Source 
* Apply suggestions from code review

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

* metal:add conv_3d backend

Rebased with master and resolved conflicts.

* Resolved issues related to changes in variable names

* kernel void kernel_upscale_bilinear_f32 was missing in my branch, added back, should pass all tests now

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Rashid Ul Islam
2026-03-23 09:45:34 +02:00
committed by GitHub
co-authored by Georgi Gerganov
parent 7a0b6a635e
commit 177c75852a
7 changed files with 232 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+75
View File
@@ -394,6 +394,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_conv_transpose_2d(ctx, idx);
} break;
case GGML_OP_CONV_3D:
{
n_fuse = ggml_metal_op_conv_3d(ctx, idx);
} break;
case GGML_OP_UPSCALE:
{
n_fuse = ggml_metal_op_upscale(ctx, idx);
@@ -3697,6 +3701,77 @@ int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_conv_3d(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;
// 1. Extract standard dimensions and byte strides
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
// 2. Extract hyperparams from op_params
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 s2 = ((const int32_t *)(op->op_params))[2];
const int32_t p0 = ((const int32_t *)(op->op_params))[3];
const int32_t p1 = ((const int32_t *)(op->op_params))[4];
const int32_t p2 = ((const int32_t *)(op->op_params))[5];
const int32_t d0 = ((const int32_t *)(op->op_params))[6];
const int32_t d1 = ((const int32_t *)(op->op_params))[7];
const int32_t d2 = ((const int32_t *)(op->op_params))[8];
const int32_t IC = ((const int32_t *)(op->op_params))[9];
const int32_t N = ((const int32_t *)(op->op_params))[10];
const int32_t OC = ((const int32_t *)(op->op_params))[11];
// 3. Build the parameter struct using the macro-generated variables
ggml_metal_kargs_conv_3d args = {
/*.IW =*/ (int32_t)op->src[1]->ne[0],
/*.IH =*/ (int32_t)op->src[1]->ne[1],
/*.ID =*/ (int32_t)op->src[1]->ne[2],
/*.OW =*/ (int32_t)op->ne[0],
/*.OH =*/ (int32_t)op->ne[1],
/*.OD =*/ (int32_t)op->ne[2],
/*.KW =*/ (int32_t)op->src[0]->ne[0],
/*.KH =*/ (int32_t)op->src[0]->ne[1],
/*.KD =*/ (int32_t)op->src[0]->ne[2],
s0, s1, s2,
p0, p1, p2,
d0, d1, d2,
IC, N, OC,
nb00, nb01, nb02, nb03, // Weight strides
nb10, nb11, nb12, nb13, // Input strides
nb0, nb1, nb2, nb3 // Output strides
};
// 4. Fetch the JIT pipeline
auto pipeline = ggml_metal_library_get_pipeline_conv_3d(lib, op);
// 5. Grid mapping
int nth0 = 32; // Standard SIMD width for Apple Silicon
int nth1 = 1;
int nth2 = 1;
int64_t spatial_volume = args.OW * args.OH * args.OD;
int ntg0 = (spatial_volume + nth0 - 1) / nth0;
int ntg1 = args.OC;
int ntg2 = args.N;
// 6. Bind and Dispatch via the ggml C wrapper
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, ntg0, ntg1, ntg2, nth0, nth1, nth2);
return 1;
}
int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);