Skip to content

GitLab

"...composable_kernel_rocm.git" did not exist on "8b83b087cd9cf7e0ba04a4ded685086c5df7dbec"
Unverified Commit 4c850c90 authored by Rostyslav Geyyer's avatar Rostyslav Geyyer Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-1815

parents ce30621d 1973903f
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 562 additions and 301 deletions
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp
View file @ 4c850c90
......@@ -13,7 +13,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multiply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
......
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_xdl_bf16_i8.cpp
View file @ 4c850c90
......@@ -13,7 +13,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multiply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
......
example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp
View file @ 4c850c90
......@@ -63,7 +63,7 @@ using DeviceGemmInstance =
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>;
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<4,4,4>>;
// clang-format on
struct ProblemSize final
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp
View file @ 4c850c90
......@@ -144,12 +144,12 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
......@@ -446,12 +446,12 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
static constexpr index_t PrefetchStages = 1;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp
View file @ 4c850c90
......@@ -153,12 +153,12 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
......@@ -646,12 +646,12 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp
View file @ 4c850c90
......@@ -146,12 +146,12 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp
View file @ 4c850c90
......@@ -147,12 +147,12 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
static constexpr index_t GlobalBufferNum = 2;
static constexpr index_t HotloopUnroll = 2;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % HotloopUnroll == 1)
{
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
View file @ 4c850c90
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -501,29 +501,24 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
// for sanity check of vector memory access
for(index_t i = 0; i < NumATensor; ++i)
{
as_mz_consecutive_[i] = a_ms_ks_strides[i][NumDimM - 1] == 1;
as_kz_consecutive_[i] = a_ms_ks_strides[i][NumDimM + NumDimK - 1] == 1;
as_max_read_elems_[i] =
tie(as_continous_dim_[i], as_max_read_elems_[i]) =
CalculateMaxRead<NumDimM, NumDimK>(a_ms_ks_lengths[i], a_ms_ks_strides[i]);
}
for(index_t i = 0; i < NumBTensor; ++i)
{
bs_nz_consecutive_[i] = b_ns_ks_strides[i][NumDimN - 1] == 1;
bs_kz_consecutive_[i] = b_ns_ks_strides[i][NumDimN + NumDimK - 1] == 1;
bs_max_read_elems_[i] =
tie(bs_continous_dim_[i], bs_max_read_elems_[i]) =
CalculateMaxRead<NumDimN, NumDimK>(b_ns_ks_lengths[i], b_ns_ks_strides[i]);
}
for(index_t i = 0; i < NumDTensor; ++i)
{
ds_nz_consecutive_[i] = d_ms_ns_strides[i][NumDimM + NumDimN - 1] == 1;
ds_max_read_elems_[i] =
tie(ds_continous_dim_[i], ds_max_read_elems_[i]) =
CalculateMaxRead<NumDimM, NumDimN>(d_ms_ns_lengths[i], d_ms_ns_strides[i]);
}
e_nz_consecutive_ = e_ms_ns_stride[NumDimM + NumDimN - 1] == 1;
e_max_write_elems_ = CalculateMaxRead<NumDimM, NumDimN>(e_ms_ns_length, e_ms_ns_stride);
tie(e_continous_dim_, e_max_write_elems_) =
CalculateMaxRead<NumDimM, NumDimN>(e_ms_ns_length, e_ms_ns_stride);
}
// pointers
......@@ -553,14 +548,11 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
// Describe whether the last part of a given dimension of A/B/D/E is consecutive
// in the memory or not.
std::array<bool, NumATensor> as_mz_consecutive_;
std::array<bool, NumATensor> as_kz_consecutive_;
std::array<bool, NumBTensor> bs_nz_consecutive_;
std::array<bool, NumBTensor> bs_kz_consecutive_;
std::array<bool, NumDTensor> ds_nz_consecutive_;
bool e_nz_consecutive_;
// Describe whether the last part of a given dimension of A/B/D/E is continues dim.
std::array<index_t, NumATensor> as_continous_dim_;
std::array<index_t, NumATensor> bs_continous_dim_;
std::array<index_t, NumBTensor> ds_continous_dim_;
index_t e_continous_dim_;
std::array<index_t, NumATensor> as_max_read_elems_;
std::array<index_t, NumBTensor> bs_max_read_elems_;
......@@ -659,9 +651,9 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
const bool valid_a_vector_size =
arg.as_max_read_elems_[i] % ABlockTransferSrcScalarPerVector == 0;
const bool valid_a_access_dim_m =
ABlockTransferSrcVectorDim == 1 && arg.as_mz_consecutive_[i];
ABlockTransferSrcVectorDim == 1 && arg.as_continous_dim_[i] == 0;
const bool valid_a_access_dim_k =
ABlockTransferSrcVectorDim == 2 && arg.as_kz_consecutive_[i];
ABlockTransferSrcVectorDim == 2 && arg.as_continous_dim_[i] == 1;
const bool valid_a_access_dim = valid_a_access_dim_m || valid_a_access_dim_k;
if(!((valid_a_vector_size && valid_a_access_dim) ||
ABlockTransferSrcScalarPerVector == 1))
......@@ -679,9 +671,9 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
const bool valid_b_vector_size =
arg.bs_max_read_elems_[i] % BBlockTransferSrcScalarPerVector == 0;
const bool valid_b_access_dim_n =
BBlockTransferSrcVectorDim == 1 && arg.bs_nz_consecutive_[i];
BBlockTransferSrcVectorDim == 1 && arg.bs_continous_dim_[i] == 0;
const bool valid_b_access_dim_k =
BBlockTransferSrcVectorDim == 2 && arg.bs_kz_consecutive_[i];
BBlockTransferSrcVectorDim == 2 && arg.bs_continous_dim_[i] == 1;
const bool valid_b_access_dim = valid_b_access_dim_n || valid_b_access_dim_k;
if(!((valid_b_vector_size && valid_b_access_dim) ||
BBlockTransferSrcScalarPerVector == 1))
......@@ -699,7 +691,7 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
const bool valid_d_vector_size =
arg.ds_max_read_elems_[i] % CDEBlockTransferScalarPerVector_NPerBlock == 0;
// Vector read of Ds is always on N dimension.
const bool valid_d_access_dim = arg.ds_nz_consecutive_[i];
const bool valid_d_access_dim = arg.ds_continous_dim_[i] == 1;
if(!((valid_d_vector_size && valid_d_access_dim) ||
CDEBlockTransferScalarPerVector_NPerBlock == 1))
{
......@@ -714,7 +706,7 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
const bool valid_e_vector_size =
arg.e_max_write_elems_ % CDEBlockTransferScalarPerVector_NPerBlock == 0;
// Vector write of E is always on N dimension.
const bool valid_e_access_dim = arg.e_nz_consecutive_;
const bool valid_e_access_dim = arg.e_continous_dim_ == 1;
if(!((valid_e_vector_size && valid_e_access_dim) ||
CDEBlockTransferScalarPerVector_NPerBlock == 1))
{
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 4c850c90
......@@ -442,25 +442,19 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
}
// for sanity check of vector memory access
a_mz_consecutive_ = a_ms_ks_strides[NumDimM - 1] == 1;
a_kz_consecutive_ = a_ms_ks_strides[NumDimM + NumDimK - 1] == 1;
a_max_read_elems_ =
tie(a_continous_dim_, a_max_read_elems_) =
CalculateMaxRead<NumDimM, NumDimK>(a_ms_ks_lengths, a_ms_ks_strides);
b_nz_consecutive_ = b_ns_ks_strides[NumDimN - 1] == 1;
b_kz_consecutive_ = b_ns_ks_strides[NumDimN + NumDimK - 1] == 1;
b_max_read_elems_ =
tie(b_continous_dim_, b_max_read_elems_) =
CalculateMaxRead<NumDimN, NumDimK>(b_ns_ks_lengths, b_ns_ks_strides);
for(index_t i = 0; i < NumDTensor; ++i)
{
ds_nz_consecutive_[i] = ds_ms_ns_strides[i][NumDimM + NumDimN - 1] == 1;
ds_max_read_elems_[i] =
tie(ds_continous_dim_[i], ds_max_read_elems_[i]) =
CalculateMaxRead<NumDimM, NumDimN>(ds_ms_ns_lengths[i], ds_ms_ns_strides[i]);
}
e_nz_consecutive_ = e_ms_ns_strides[NumDimM + NumDimN - 1] == 1;
e_max_write_elems_ =
tie(e_continous_dim_, e_max_write_elems_) =
CalculateMaxRead<NumDimM, NumDimN>(e_ms_ns_lengths, e_ms_ns_strides);
}
......@@ -501,14 +495,11 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
// Describe whether the last part of a given dimension of A/B/D/E is consecutive
// in the memory or not.
bool a_mz_consecutive_;
bool a_kz_consecutive_;
bool b_nz_consecutive_;
bool b_kz_consecutive_;
std::array<bool, NumDTensor> ds_nz_consecutive_;
bool e_nz_consecutive_;
// Describe whether the last part of a given dimension of A/B/D/E is continues dim.
index_t a_continous_dim_;
index_t b_continous_dim_;
std::array<index_t, NumDTensor> ds_continous_dim_;
index_t e_continous_dim_;
index_t a_max_read_elems_;
index_t b_max_read_elems_;
......@@ -624,8 +615,10 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
const bool valid_a_vector_size =
arg.a_max_read_elems_ % ABlockTransferSrcScalarPerVector == 0;
const bool valid_a_access_dim_m = ABlockTransferSrcVectorDim == 1 && arg.a_mz_consecutive_;
const bool valid_a_access_dim_k = ABlockTransferSrcVectorDim == 2 && arg.a_kz_consecutive_;
const bool valid_a_access_dim_m =
ABlockTransferSrcVectorDim == 1 && arg.a_continous_dim_ == 0;
const bool valid_a_access_dim_k =
ABlockTransferSrcVectorDim == 2 && arg.a_continous_dim_ == 1;
const bool valid_a_access_dim =
valid_a_access_dim_m || valid_a_access_dim_k || ABlockTransferSrcScalarPerVector == 1;
if(!(valid_a_vector_size && valid_a_access_dim))
......@@ -635,8 +628,10 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
const bool valid_b_vector_size =
arg.b_max_read_elems_ % BBlockTransferSrcScalarPerVector == 0;
const bool valid_b_access_dim_n = BBlockTransferSrcVectorDim == 1 && arg.b_nz_consecutive_;
const bool valid_b_access_dim_k = BBlockTransferSrcVectorDim == 2 && arg.b_kz_consecutive_;
const bool valid_b_access_dim_n =
BBlockTransferSrcVectorDim == 1 && arg.b_continous_dim_ == 0;
const bool valid_b_access_dim_k =
BBlockTransferSrcVectorDim == 2 && arg.b_continous_dim_ == 1;
const bool valid_b_access_dim =
valid_b_access_dim_n || valid_b_access_dim_k || BBlockTransferSrcScalarPerVector == 1;
if(!(valid_b_vector_size && valid_b_access_dim))
......@@ -650,7 +645,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
arg.ds_max_read_elems_[i] % CDEBlockTransferScalarPerVector_NPerBlock == 0;
// Vector read of Ds is always on N dimension.
const bool valid_d_access_dim =
arg.ds_nz_consecutive_[i] || CDEBlockTransferScalarPerVector_NPerBlock == 1;
arg.ds_continous_dim_[i] == 1 || CDEBlockTransferScalarPerVector_NPerBlock == 1;
if(!(valid_d_vector_size && valid_d_access_dim))
{
valid_ds_access = false;
......@@ -665,7 +660,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
arg.e_max_write_elems_ % CDEBlockTransferScalarPerVector_NPerBlock == 0;
// Vector write of E is always on N dimension.
const bool valid_e_access_dim =
arg.e_nz_consecutive_ || CDEBlockTransferScalarPerVector_NPerBlock == 1;
arg.e_continous_dim_ == 1 || CDEBlockTransferScalarPerVector_NPerBlock == 1;
if(!(valid_e_vector_size && valid_e_access_dim))
{
return false;
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp
View file @ 4c850c90
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -50,25 +50,53 @@ auto CalculateMaxRead(const std::vector<index_t>& lengths, const std::vector<ind
}
// Determine the beginning and end idx of the group representing the FCD.
index_t begin_idx, end_idx;
if(strides[NumDim1 - 1] == 1)
index_t begin_idx, end_idx, continous_dim, consecutive_stride = 1;
if(strides[NumDim1 - 1] == 1 && strides[NumDim1 + NumDim2 - 1] == 1)
{
begin_idx = 0;
end_idx = NumDim1 - 1;
// MZ or KZ are ones
bool dims1_are_ones = true;
for(index_t dim_idx = 0; dim_idx < NumDim1; dim_idx++)
{
if(lengths[dim_idx] != 1)
{
dims1_are_ones = false;
}
}
if(dims1_are_ones)
{
begin_idx = NumDim1;
end_idx = NumDim1 + NumDim2 - 1;
continous_dim = 1;
}
else
{
begin_idx = 0;
end_idx = NumDim1 - 1;
continous_dim = 0;
}
}
else if(strides[NumDim1 - 1] == 1)
{
begin_idx = 0;
end_idx = NumDim1 - 1;
continous_dim = 0;
}
else if(strides[NumDim1 + NumDim2 - 1] == 1)
{
begin_idx = NumDim1;
end_idx = NumDim1 + NumDim2 - 1;
begin_idx = NumDim1;
end_idx = NumDim1 + NumDim2 - 1;
continous_dim = 1;
}
else
{
// The dimension consecutive in memory is not the last dimension of any group, so only
// one element can be read/written at once.
return 1;
consecutive_stride = 1;
continous_dim = 0;
return make_tuple(continous_dim, consecutive_stride);
}
index_t consecutive_stride = 1;
for(index_t dim_idx = end_idx; dim_idx >= begin_idx; --dim_idx)
{
if(strides[dim_idx] == consecutive_stride)
......@@ -81,7 +109,7 @@ auto CalculateMaxRead(const std::vector<index_t>& lengths, const std::vector<ind
}
}
const index_t max_subsequent_elems = consecutive_stride;
return max_subsequent_elems;
return make_tuple(continous_dim, max_subsequent_elems);
}
} // namespace device
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View file @ 4c850c90
......@@ -93,9 +93,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
View file @ 4c850c90
......@@ -54,9 +54,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
__shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp
View file @ 4c850c90
......@@ -66,9 +66,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
__shared__ FloatA p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatA)];
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View file @ 4c850c90
......@@ -59,9 +59,12 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
......@@ -113,9 +116,12 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 4c850c90
......@@ -97,9 +97,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ 4c850c90
......@@ -106,10 +106,12 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
const long_index_t e_batch_offset = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
......@@ -170,10 +172,13 @@ __global__ void
}
else
{
const long_index_t a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid + a_batch_offset + a_n_offset,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp
View file @ 4c850c90
......@@ -85,12 +85,17 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
const long_index_t a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
const long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
......@@ -142,12 +147,17 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
const long_index_t a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
const long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ 4c850c90
......@@ -161,11 +161,11 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t a_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t b_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t e_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 4c850c90
......@@ -908,6 +908,51 @@ struct OffsettedBlockToCTileMap
UnderlyingBlockToCTileMap block_to_ctile_map_;
index_t block_start_;
};
// second version with 2 offsets
template <typename UnderlyingBlockToCTileMap>
struct OffsettedBlockToCTileMap2
{
using underlying_type = UnderlyingBlockToCTileMap;
__host__ __device__ OffsettedBlockToCTileMap2(UnderlyingBlockToCTileMap block_to_ctile_map,
index_t group_offset,
index_t tile_offset)
: block_to_ctile_map_{block_to_ctile_map},
group_offset_{group_offset},
tile_offset_{tile_offset}
{
}
template <typename TopIdx>
__host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
{
return block_to_ctile_map_.CalculateBottomIndex(
make_multi_index(idx_top[Number<0>{}] + tile_offset_ - group_offset_));
}
template <typename CTileIdx, typename CTileDim>
__host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
const CTileDim& c_tile_dim) const
{
return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
}
template <typename CGridDesc_M_N>
__host__ constexpr bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
{
return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
}
__host__ __device__ constexpr index_t CalculateGridSize(index_t M, index_t N) const
{
return block_to_ctile_map_.CalculateGridSize(M, N);
}
__device__ void UpdateTileOffset(index_t offset) { tile_offset_ = offset; }
UnderlyingBlockToCTileMap block_to_ctile_map_;
index_t group_offset_;
index_t tile_offset_;
};
/**
* @brief Simple tile mapping which creates 3D grid of block of threads.
......
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