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

metal : add count_equal op (#18314)

Browse Source 
* add count equal for metal

* remove trailing whitespace

* updated doc ops table

* changed shmem to i32

* added multi tg and templating

* removed BLAS support from Metal docs

* Apply suggestions from code review

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

* add memset to set dst to 0

* metal : cleanup

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
gatbontonpc
2025-12-31 10:39:48 +02:00
committed by GitHub
co-authored by Georgi Gerganov
parent ecc343de63
commit 9a6369bb60
9 changed files with 585 additions and 324 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+66 -1
View File
@@ -448,7 +448,11 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_opt_step_sgd(ctx, idx);
} break;
default:
case GGML_OP_COUNT_EQUAL:
{
n_fuse = ggml_metal_op_count_equal(ctx, idx);
} break;
default:
{
GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(node->op));
GGML_ABORT("fatal error");
@@ -4090,3 +4094,64 @@ int ggml_metal_op_opt_step_sgd(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_count_equal(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(uint64_t, nb1, op->src[1], nb);
{
ggml_metal_kargs_memset args = { /*.val =*/ 0 };
auto pipeline = ggml_metal_library_get_pipeline_memset(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), 1);
ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, 1, 1, 1);
}
ggml_metal_op_concurrency_reset(ctx);
{
ggml_metal_kargs_count_equal args = {
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.nb10 =*/ nb10,
/*.nb11 =*/ nb11,
/*.nb12 =*/ nb12,
/*.nb13 =*/ nb13,
};
auto pipeline = ggml_metal_library_get_pipeline_count_equal(lib, op);
const size_t smem = pipeline.smem;
const int nth = 32*pipeline.nsg;
GGML_ASSERT(nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
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_set_threadgroup_memory_size(enc, smem, 0);
ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
}
return 1;
}