Skip to content

GitLab

Commit 7733dd88 authored by Chao Liu's avatar Chao Liu
Browse files

use readfirstlane to force result into SGPR to reduce VGPR usage

parent 3b3cfae5
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 447 additions and 464 deletions
composable_kernel/include/tensor_description/dynamic_multi_index_transform.hpp
View file @ 7733dd88
......@@ -423,17 +423,6 @@ struct DynamicMerge
LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
"wrong! inconsistent # of dimension");
#if 0
// I only want to do this check, if idx_diff_up is know at compile-time
if(idx_diff_up[Number<0>{}] == 0)
{
static_for<0, NDimLow, 1>{}([&idx_diff_low](auto i){
idx_diff_low(i) = 0;
});
return;
}
#endif
// CalculateLowerIndex(idx_diff_low_const) has multiple integer divisions.
// However,
// 1) If idx_diff_up is known at compile-time, then idx_diff_low_const
......@@ -449,7 +438,19 @@ struct DynamicMerge
// computed at
// run-time each time this function is called, and can be very expensive.
LowerIndex idx_diff_low_const;
#if !CK_HACK_DYNAMIC_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
CalculateLowerIndex(idx_diff_low_const, idx_diff_up);
#else
index_t tmp = idx_diff_up[Number<0>{}];
static_for<0, NDimLow - 1, 1>{}([&](auto i) {
idx_diff_low_const(i) = tmp / low_lengths_scan_[i];
tmp -= idx_diff_low_const[i] * low_lengths_scan_[i];
});
// Hack: this force result into SGPR. Need to make sure the result is thread invariant
idx_diff_low_const(Number<NDimLow - 1>{}) = __builtin_amdgcn_readfirstlane(tmp);
#endif
// do carry check on each low dimension in reversed order
// do not need to check the first dimension
......
composable_kernel/include/tensor_operation/blockwise_dynamic_tensor_slice_transfer.hpp
View file @ 7733dd88
......@@ -121,9 +121,9 @@ struct BlockwiseDynamicTensorSliceTransfer_v1r1
ThreadwiseTransfer threadwise_transfer_;
};
// this version is very likely to have scratch memory issue, due to:
// this version tend to have scratch memory issue, due to:
// 1. ThreadwiseDynamicTensorSliceTransfer_v1r1 keeps reference to tensor descriptor
// 2. threadwise_dynamic_tensor_slice_transfer_v1r1 constructs new tensor coordinate
// 2. ThreadwiseDynamicTensorSliceTransfer_v1r1::Run() constructs new tensor coordinate
template <index_t BlockSize,
typename BlockSrcData,
typename BlockDstData,
......@@ -289,7 +289,7 @@ struct BlockwiseDynamicTensorSliceTransfer_v2r1
// this version does following things to avoid scratch memory issue
// 1. ThreadwiseDynamicTensorSliceTransfer_v1r2 does not keep reference to tensor descriptor
// 2. threadwise_dynamic_tensor_slice_transfer_v1r2 does not construct new tensor coordinate
// 2. ThreadwiseDynamicTensorSliceTransfer_v1r2::Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename BlockSrcData,
typename BlockDstData,
......@@ -465,7 +465,7 @@ struct BlockwiseDynamicTensorSliceTransfer_v2r2
// this version does following things to avoid scratch memory issue
// 1. BlockwiseDynamicTensorSliceTransfer_v2r3 doesn't allocate thread buffer (array) as member
// 2. ThreadwiseDynamicTensorSliceTransfer_v1r2 does not keep reference to tensor descriptor
// 3. threadwise_dynamic_tensor_slice_transfer_v1r2 does not construct new tensor coordinate
// 3. ThreadwiseDynamicTensorSliceTransfer_v1r2::Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename BlockSrcData,
typename BlockDstData,
......@@ -485,7 +485,9 @@ template <index_t BlockSize,
AddressSpace DstAddressSpace,
InMemoryDataOperation DstInMemOp,
index_t SrcDataStride,
index_t DstDataStride>
index_t DstDataStride,
index_t ThreadTransferMoveBackSrcCoord = true,
index_t ThreadTransferMoveBackDstCoord = true>
struct BlockwiseDynamicTensorSliceTransfer_v2r3
{
static constexpr index_t nDim =
......@@ -607,20 +609,25 @@ struct BlockwiseDynamicTensorSliceTransfer_v2r3
AddressSpace::Vgpr,
InMemoryDataOperation::Set,
SrcDataStride,
1>;
using ThreadwiseWrite = ThreadwiseDynamicTensorSliceTransfer_v1r2<decltype(thread_buffer_desc_),
BlockDstDesc,
ThreadSliceLengths,
DstDimAccessOrder,
DstVectorWriteDim,
1,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp,
1,
DstDataStride>;
1,
ThreadTransferMoveBackSrcCoord,
true>;
using ThreadwiseWrite =
ThreadwiseDynamicTensorSliceTransfer_v1r2<decltype(thread_buffer_desc_),
BlockDstDesc,
ThreadSliceLengths,
DstDimAccessOrder,
DstVectorWriteDim,
1,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp,
1,
DstDataStride,
true,
ThreadTransferMoveBackDstCoord>;
ThreadwiseRead threadwise_read_;
ThreadwiseWrite threadwise_write_;
......
composable_kernel/include/tensor_operation/gridwise_dynamic_gemm.hpp
View file @ 7733dd88
......@@ -459,11 +459,24 @@ struct GridwiseDynamicGemm_km_kn_mn_v1r2
const index_t N = b_k_n_global_desc.GetLength(I1);
// divide block work by [M, N]
#if 0
const index_t m_block_work_num = M / MPerBlock;
const index_t n_block_work_num = N / NPerBlock;
#else
// Hack: this force result into SGPR
const index_t m_block_work_num = __builtin_amdgcn_readfirstlane(M / MPerBlock);
const index_t n_block_work_num = __builtin_amdgcn_readfirstlane(N / NPerBlock);
#endif
#if 0
const index_t m_block_work_id = get_block_1d_id() / n_block_work_num;
const index_t n_block_work_id = get_block_1d_id() - m_block_work_id * n_block_work_num;
#else
// Hack: this force result into SGPR
const index_t m_block_work_id =
__builtin_amdgcn_readfirstlane(get_block_1d_id() / n_block_work_num);
const index_t n_block_work_id = get_block_1d_id() - m_block_work_id * n_block_work_num;
#endif
const index_t m_block_data_on_global = m_block_work_id * MPerBlock;
const index_t n_block_data_on_global = n_block_work_id * NPerBlock;
......@@ -505,10 +518,13 @@ struct GridwiseDynamicGemm_km_kn_mn_v1r2
AddressSpace::Lds,
InMemoryDataOperation::Set,
1,
1>(a_k_m_global_desc,
make_multi_index(0, m_block_data_on_global),
a_k_m_block_desc,
make_multi_index(0, 0));
1,
true,
true>(
a_k_m_global_desc,
make_multi_index(0, m_block_data_on_global),
a_k_m_block_desc,
make_multi_index(0, 0));
// B matrix blockwise copy
auto b_block_copy =
......@@ -531,10 +547,17 @@ struct GridwiseDynamicGemm_km_kn_mn_v1r2
AddressSpace::Lds,
InMemoryDataOperation::Set,
1,
1>(b_k_n_global_desc,
make_multi_index(0, n_block_data_on_global),
b_k_n_block_desc,
make_multi_index(0, 0));
1,
#if 0
true.
#else
false,
#endif
true>(
b_k_n_global_desc,
make_multi_index(0, n_block_data_on_global),
b_k_n_block_desc,
make_multi_index(0, 0));
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......@@ -599,7 +622,14 @@ struct GridwiseDynamicGemm_km_kn_mn_v1r2
threadwise_matrix_set_zero(c_m0m1_n0n1_thread_mtx_desc, p_c_thread);
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock, 0);
#if 0
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock, 0);
#else
// HACK: fuse threadwise copy move-back coordinate with move src slice window
constexpr auto b_block_slice_copy_step =
b_block_copy.threadwise_read_.GetCoordinateStepBack() + make_multi_index(KPerBlock, 0);
#endif
// LDS double buffer: preload data into LDS
{
......
composable_kernel/include/utility/config.amd.hpp.in
View file @ 7733dd88
......@@ -74,6 +74,14 @@
#define CK_EXPERIMENTAL_IMPLICIT_GEMM_BACKWARD_DATA_V4R1_INPUT_SKIP_OUT_OF_BOUND_CHECK 0
#endif
// hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
// thread-invariant, otherwise it's a bug
// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
#ifndef CK_HACK_DYNAMIC_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
#define CK_HACK_DYNAMIC_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
#endif
// workaround: put all workaround here
// workaround for unnecessary VGPA <--> AGRP data movement when using mfma LLVM intrinsic
#ifndef CK_WORKAROUND_SWDEV_229564
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment