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

Unify the naming of the math functions used by the host and kernel (#262)

Browse Source 
* Use the unified naming for math functions on host and HIP kernel

* Corresponding change/simplification in reduction host/profiler/examples due to unified math functions renaming

* Renaming GetReductionZeroVal() to GetIdentityValue()

* Tiny renaming in profile_reduce_impl.hpp

* More renaming in profile_reduce_impl.hpp

* Replace zeroVal by identiyVal

* Remove ck_ prefix in the naming of ck::math provided functions
This commit is contained in:
Qianfeng
2022-06-02 10:49:53 +08:00
committed by GitHub
parent b6eaf3eb7e
commit 86185bd7ce
22 changed files with 198 additions and 417 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
View File

@@ -171,7 +171,7 @@ struct GridwiseReduction_mk_to_m_multiblock
AccDataType beta,
OutDataType* const __restrict__ p_out_value_global)
{
const auto zeroVal = ReduceOperation::GetReductionZeroVal();
const auto identityVal = ReduceOperation::GetIdentityValue();
// LDS
__shared__ AccDataType p_reduce_work_buffer[BlockSize];
@@ -179,7 +179,7 @@ struct GridwiseReduction_mk_to_m_multiblock
const auto in_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
in_grid_desc_m_k.GetElementSpaceSize(),
type_convert<InDataType>(zeroVal));
type_convert<InDataType>(identityVal));
auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
@@ -191,7 +191,7 @@ struct GridwiseReduction_mk_to_m_multiblock
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = zeroVal; });
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
@@ -358,12 +358,12 @@ struct GridwiseReduction_mk_to_m_multiblock
__shared__ AccDataType p_reduce_work_val_buffer[BlockSize];
__shared__ IndexDataType p_reduce_work_idx_buffer[BlockSize];
const auto zeroVal = ReduceOperation::GetReductionZeroVal();
const auto identityVal = ReduceOperation::GetIdentityValue();
const auto in_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
in_grid_desc_m_k.GetElementSpaceSize(),
type_convert<InDataType>(zeroVal));
type_convert<InDataType>(identityVal));
const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
@@ -418,7 +418,7 @@ struct GridwiseReduction_mk_to_m_multiblock
thread_k_cluster_id * KThreadSliceSize));
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
accu_value_buf(I) = zeroVal;
accu_value_buf(I) = identityVal;
accu_index_buf(I) = 0;
});
@@ -459,7 +459,7 @@ struct GridwiseReduction_mk_to_m_multiblock
in_thread_idx_buf);
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
AccDataType tmpValue = zeroVal;
AccDataType tmpValue = identityVal;
IndexDataType tmpIndex = 0;
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
@@ -512,7 +512,7 @@ struct GridwiseReduction_mk_to_m_multiblock
in_thread_val_buf(Number<offset>{}));
});
AccDataType tmpValue = zeroVal;
AccDataType tmpValue = identityVal;
IndexDataType tmpIndex = 0;
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {

12
include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
View File

@@ -135,12 +135,12 @@ struct GridwiseReduction_mk_to_m_threadwise
ReduceOperation,
PropagateNan>;
const auto zeroVal = ReduceOperation::GetReductionZeroVal();
const auto identityVal = ReduceOperation::GetIdentityValue();
const auto in_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
in_grid_desc_m_k.GetElementSpaceSize(),
type_convert<InDataType>(zeroVal));
type_convert<InDataType>(identityVal));
auto dst_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
@@ -149,7 +149,7 @@ struct GridwiseReduction_mk_to_m_threadwise
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = zeroVal; });
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
@@ -276,12 +276,12 @@ struct GridwiseReduction_mk_to_m_threadwise
(void)acc_elementwise_op;
const auto zeroVal = ReduceOperation::GetReductionZeroVal();
const auto identityVal = ReduceOperation::GetIdentityValue();
const auto in_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
in_grid_desc_m_k.GetElementSpaceSize(),
type_convert<InDataType>(zeroVal));
type_convert<InDataType>(identityVal));
const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
@@ -303,7 +303,7 @@ struct GridwiseReduction_mk_to_m_threadwise
StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, MThreadSliceSize, true> accu_index_buf;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
accu_value_buf(I) = zeroVal;
accu_value_buf(I) = identityVal;
accu_index_buf(I) = 0;
});

4
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View File

@@ -816,10 +816,10 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
false>;
// Global write Gemm shuffle + reduction
const auto d_zeroVal = DReduceOperation::GetReductionZeroVal();
const auto d_identityVal = DReduceOperation::GetIdentityValue();
static_for<0, mreduce_per_thread, 1>{}(
[&](auto I) { d_thread_buf(I) = d_zeroVal; });
[&](auto I) { d_thread_buf(I) = d_identityVal; });
// reduce in VGPR
static_for<0, mreduce_per_thread, 1>{}([&](auto im) {