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

ggml : extend ggml_pool_1d + metal (#16429)

Browse Source 
* chore: resolve conflicts

* feat: ggml metal impl

* fix: ggml_metal_kargs_pool_1d struct

* fix: require contiguous input

* chore: test pool_1d

* chore: limit pool1d test cases to p0=0 and s0=k0 to conform with asserts

* chore: add p0 and s0 to testing

* fix: allow padding for cpu and metal

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

* fix: correct single-threaded loop

* ggml : cleanup

* tests : add ne[1] != 1 tests

* fix: ne[1] handling in np

* cont : fixes

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Thore Koritzius
2026-01-16 16:59:56 +02:00
committed by GitHub
co-authored by Georgi Gerganov
parent 6ba6a3c76f
commit 388ce82241
10 changed files with 272 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+52
View File
@@ -432,6 +432,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_cpy(ctx, idx);
} break;
case GGML_OP_POOL_1D:
{
n_fuse = ggml_metal_op_pool_1d(ctx, idx);
} break;
case GGML_OP_POOL_2D:
{
n_fuse = ggml_metal_op_pool_2d(ctx, idx);
@@ -1622,6 +1626,54 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_pool_1d(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, ne, op, ne);
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
const int32_t * opts = op->op_params;
ggml_op_pool op_pool = (ggml_op_pool) opts[0];
const int32_t k0 = opts[1];
const int32_t s0 = opts[2];
const int32_t p0 = opts[3];
const int64_t IW = op->src[0]->ne[0];
const int64_t OW = op->ne[0];
const int64_t np = ggml_nelements(op);
ggml_metal_kargs_pool_1d args_pool_1d = {
/* .k0 = */ k0,
/* .s0 = */ s0,
/* .p0 = */ p0,
/* .IW = */ IW,
/* .OW = */ OW,
/* .np = */ np
};
auto pipeline = ggml_metal_library_get_pipeline_pool_1d(lib, op, op_pool);
const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), (int) np);
const int ntg = (np + nth - 1) / nth;
ggml_metal_encoder_set_pipeline(enc, pipeline);
ggml_metal_encoder_set_bytes (enc, &args_pool_1d, sizeof(args_pool_1d), 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), 2);
ggml_metal_encoder_dispatch_threadgroups(enc, ntg, 1, 1, nth, 1, 1);
return 1;
}
int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);