mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-02 20:51:23 +00:00
[rocm-libraries] ROCm/rocm-libraries#4302 (commit e62bd8a)
[CK_TILE] add tf32 support MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Proposed changes TF32 is added in CK on gfx942 and gfx950. This PR is to initiate tf32 in CK_TILE on gfx942 and gfx950. ## Checklist Please put an into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask. - [ ] I have added tests relevant to the introduced functionality, and the unit tests are passing locally - [ ] I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more than 30 seconds to run. - [ ] I have added inline documentation which enables the maintainers with understanding the motivation - [ ] I have removed the stale documentation which is no longer relevant after this pull request - [ ] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request - [x] I have run on all changed files - [ ] Any dependent changes have been merged ## Discussion
This commit is contained in:
yinglu
committed by
assistant-librarian[bot]
assistant-librarian[bot]
parent
652d3456ca
commit
d460ab35b6
@@ -6,8 +6,8 @@
|
||||
struct BasicInvoker
|
||||
{
|
||||
template <typename GemmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename DsDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
@@ -19,14 +19,30 @@ struct BasicInvoker
|
||||
typename CDEElementWise>
|
||||
static float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
|
||||
{
|
||||
// ADataTypeCompute: compute type (tf32_t for TF32 mode, used for warp gemm selection)
|
||||
// ADataTypeBuf: buffer/storage type (fp32 when tf32)
|
||||
using ADataTypeCompute = ADataType_;
|
||||
using BDataTypeCompute = BDataType_;
|
||||
using ADataTypeBuf = ck_tile::if_select_t<ADataType_, ck_tile::tf32_t, float, ADataType_>;
|
||||
using BDataTypeBuf = ck_tile::if_select_t<BDataType_, ck_tile::tf32_t, float, BDataType_>;
|
||||
|
||||
if constexpr(std::is_same_v<ADataTypeCompute, ck_tile::tf32_t>)
|
||||
{
|
||||
static_assert(std::is_same_v<ADataTypeCompute, BDataTypeCompute>,
|
||||
"ADataTypeCompute and BDataTypeCompute must be the same");
|
||||
}
|
||||
|
||||
if constexpr(Persistent)
|
||||
{
|
||||
std::cout << "WARNING: Ignoring persistent kernel option for basic gemm." << std::endl;
|
||||
}
|
||||
|
||||
constexpr bool is_fp32_input = std::is_same_v<ADataTypeBuf, float>;
|
||||
constexpr bool is_tf32_compute = std::is_same_v<ADataTypeCompute, ck_tile::tf32_t>;
|
||||
|
||||
// This part comes from the Codegen
|
||||
constexpr ck_tile::index_t M_Tile = 256;
|
||||
constexpr ck_tile::index_t N_Tile = 256;
|
||||
constexpr ck_tile::index_t M_Tile = is_fp32_input ? 128 : 256;
|
||||
constexpr ck_tile::index_t N_Tile = is_fp32_input ? 128 : 256;
|
||||
constexpr ck_tile::index_t K_Tile = 64;
|
||||
|
||||
#if CK_TILE_USE_WMMA
|
||||
@@ -38,12 +54,14 @@ struct BasicInvoker
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#else
|
||||
constexpr ck_tile::index_t M_Warp = 2;
|
||||
constexpr ck_tile::index_t N_Warp = 2;
|
||||
// gfx950: fp32 uses 16x16x16 tile (native MFMA)
|
||||
// tf32 uses 32x32x16 tile (3x bf16 32x32x16 MFMA emulation)
|
||||
constexpr ck_tile::index_t M_Warp = (is_fp32_input && !is_tf32_compute) ? 4 : 2;
|
||||
constexpr ck_tile::index_t N_Warp = (is_fp32_input && !is_tf32_compute) ? 4 : 2;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t M_Warp_Tile = (is_fp32_input && !is_tf32_compute) ? 16 : 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = (is_fp32_input && !is_tf32_compute) ? 16 : 32;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#endif
|
||||
|
||||
@@ -61,17 +79,21 @@ struct BasicInvoker
|
||||
BLayout,
|
||||
CLayout>;
|
||||
|
||||
using CodegenPipelineProblem = ck_tile::GemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CodegenGemmShape,
|
||||
CodegenGemmTraits>;
|
||||
using CodegenPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataTypeBuf,
|
||||
BDataTypeBuf,
|
||||
AccDataType,
|
||||
CodegenGemmShape,
|
||||
CodegenGemmTraits,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
ADataTypeCompute>;
|
||||
|
||||
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
|
||||
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
ck_tile::CShuffleEpilogueProblem<ADataTypeCompute,
|
||||
BDataTypeCompute,
|
||||
ck_tile::tuple<>,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
@@ -112,7 +134,7 @@ struct BasicInvoker
|
||||
}
|
||||
|
||||
// Declare rotating_mem_ptr here so it stays in scope until it is needed
|
||||
std::unique_ptr<ck_tile::RotatingMemWrapper<ADataType, BDataType>> rotating_mem_ptr;
|
||||
std::unique_ptr<ck_tile::RotatingMemWrapper<ADataTypeBuf, BDataTypeBuf>> rotating_mem_ptr;
|
||||
std::function<void()> preprocess;
|
||||
|
||||
auto clear_gemm_output = [&]() {
|
||||
@@ -125,16 +147,21 @@ struct BasicInvoker
|
||||
{
|
||||
std::cout << "Flushing cache..." << std::endl;
|
||||
|
||||
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
|
||||
ck_tile::HostTensor<ADataTypeBuf> a_m(ck_tile::host_tensor_descriptor(
|
||||
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
|
||||
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
|
||||
ck_tile::HostTensor<BDataTypeBuf> b_n(ck_tile::host_tensor_descriptor(
|
||||
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
|
||||
|
||||
auto size_a_buffer = a_m.get_element_space_size_in_bytes();
|
||||
auto size_b_buffer = b_n.get_element_space_size_in_bytes();
|
||||
|
||||
rotating_mem_ptr = std::make_unique<ck_tile::RotatingMemWrapper<ADataType, BDataType>>(
|
||||
kargs.as_ptr[0], kargs.bs_ptr[0], s.rotating_count_, size_a_buffer, size_b_buffer);
|
||||
rotating_mem_ptr =
|
||||
std::make_unique<ck_tile::RotatingMemWrapper<ADataTypeBuf, BDataTypeBuf>>(
|
||||
kargs.as_ptr[0],
|
||||
kargs.bs_ptr[0],
|
||||
s.rotating_count_,
|
||||
size_a_buffer,
|
||||
size_b_buffer);
|
||||
rotating_mem_ptr->Print();
|
||||
|
||||
preprocess = [&]() {
|
||||
|
||||
Reference in New Issue
Block a user