ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 05:01:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Compile for gfx908 and gfx90a (#130)

Browse Source 
* adding compilation for multiple targets

* fix build

* clean

* update Jekinsfile

* update readme

* update Jenkins

* use ck::half_t instead of ushort for bf16

* rename enum classes

* clean

* rename

* clean
This commit is contained in:
Chao Liu
2022-03-31 12:33:34 -05:00
committed by GitHub
parent ecf337bab5
commit cd167e492a
227 changed files with 1398 additions and 2944 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
View File

@@ -207,9 +207,9 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
CM0M1N0N1ThreadDesc{}.GetLength(I2) == N0,
"wrong");
auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatA>(
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_k_m0_m1_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatB>(
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_k_n0_n1_thread_desc_.GetElementSpaceSize());
constexpr auto threadwise_gemm =

4
include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r3.hpp
View File

@@ -220,9 +220,9 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I2) == BN0,
"wrong");
auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatA>(
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatB>(
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
constexpr auto threadwise_contraction =

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
View File

@@ -119,7 +119,7 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
constexpr auto a_block_mtx = ABlockDesc_E1_K1_E2{};
// thread A buffer for GEMM
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
StaticBuffer<AddressSpaceEnum::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
a_thread_buf;
constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,

6
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View File

@@ -42,7 +42,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
StaticBufferTupleOfVector<AddressSpaceEnum_t::Vgpr,
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
FloatAcc,
MRepeat * NRepeat,
xdlops_gemm.GetRegSizePerXdlops(),
@@ -250,9 +250,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, MRepeat, 1>{}([&](auto m0) {

2
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v4r1.hpp
View File

@@ -16,7 +16,7 @@ namespace ck {
template <index_t BlockSize,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,

2
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
View File

@@ -14,7 +14,7 @@ namespace ck {
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename ThreadSliceLengths,
typename ThreadClusterLengths,

2
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r1.hpp
View File

@@ -15,7 +15,7 @@ namespace ck {
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename ElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,

2
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r2.hpp
View File

@@ -15,7 +15,7 @@ namespace ck {
// 3. Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename ElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,

2
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r3.hpp
View File

@@ -15,7 +15,7 @@ namespace ck {
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename ElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,