Gather mm new kernel and small refactoring (#2040)

This commit is contained in:
Angelos Katharopoulos
2025-04-14 16:37:36 -07:00
committed by GitHub
parent e9e268336b
commit 99eefd2ec0
23 changed files with 1260 additions and 378 deletions

View File

@@ -584,6 +584,44 @@ MTL::ComputePipelineState* get_steel_gemm_masked_kernel(
return d.get_kernel(kernel_name, lib);
}
MTL::ComputePipelineState* get_steel_gemm_gather_kernel(
metal::Device& d,
const std::string& kernel_name,
const std::string& hash_name,
const metal::MTLFCList& func_consts,
const array& out,
bool transpose_a,
bool transpose_b,
int bm,
int bn,
int bk,
int wm,
int wn,
bool rhs) {
const auto& lib_name = kernel_name;
auto lib = d.get_library(lib_name, [&]() {
std::string kernel_source;
concatenate(
kernel_source,
metal::utils(),
metal::gemm(),
metal::steel_gemm_gather(),
get_template_definition(
lib_name,
rhs ? "gather_mm_rhs" : "gather_mm",
get_type_string(out.dtype()),
bm,
bn,
bk,
wm,
wn,
transpose_a,
transpose_b));
return kernel_source;
});
return d.get_kernel(kernel_name, lib, hash_name, func_consts);
}
MTL::ComputePipelineState* get_gemv_masked_kernel(
metal::Device& d,
const std::string& kernel_name,