Skip to content

GitLab

Commit e00a943e authored by myamlak's avatar myamlak
Browse files

Merge remote-tracking branch 'origin/develop' into myamlak/cgemm

parents ffe12e2e 9f71ff48
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 436 additions and 272 deletions
include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp
View file @ e00a943e
...@@ -385,8 +385,11 @@ struct DeviceGemmXdlSplitK ...@@ -385,8 +385,11 @@ struct DeviceGemmXdlSplitK
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", " std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl; << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
} }
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
ShowInfo(arg);
const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0); const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_, if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
...@@ -408,50 +411,30 @@ struct DeviceGemmXdlSplitK ...@@ -408,50 +411,30 @@ struct DeviceGemmXdlSplitK
float ave_time = 0; float ave_time = 0;
const auto Run = [&](const auto& kernel) { const auto Run = [&](const auto& kernel) {
if(nrepeat > 0) // FIXME: this should be moved outside of DeviceOp
{ hipGetErrorString(
ShowInfo(arg); hipMemset(arg.p_c_grid_,
ave_time = launch_and_time_kernel(kernel, 0,
nrepeat, arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_.GetElementSpaceSize() *
dim3(grid_size), sizeof(CDataType)));
dim3(BlockSize),
0, ave_time = launch_and_time_kernel(stream_config,
arg.p_a_grid_, kernel,
arg.p_b_grid_, dim3(grid_size),
arg.p_c_grid_, dim3(BlockSize),
arg.a_grid_desc_kbatch_k0_m_k1_, 0,
arg.b_grid_desc_kbatch_k0_n_k1_, arg.p_a_grid_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_, arg.p_b_grid_,
arg.a_element_op_, arg.p_c_grid_,
arg.b_element_op_, arg.a_grid_desc_kbatch_k0_m_k1_,
arg.c_element_op_, arg.b_grid_desc_kbatch_k0_n_k1_,
arg.block_2_ctile_map_); arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
} arg.a_element_op_,
arg.b_element_op_,
if(kbatch > 1 || nrepeat <= 0) arg.c_element_op_,
{ arg.block_2_ctile_map_);
hipGetErrorString(
hipMemset(arg.p_c_grid_,
0,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_.GetElementSpaceSize() *
sizeof(CDataType)));
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
}; };
if(has_main_k0_block_loop) if(has_main_k0_block_loop)
{ {
if(kbatch == 1) if(kbatch == 1)
...@@ -531,9 +514,10 @@ struct DeviceGemmXdlSplitK ...@@ -531,9 +514,10 @@ struct DeviceGemmXdlSplitK
} }
// polymorphic // polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
} }
}; };
......
include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ e00a943e
...@@ -391,8 +391,11 @@ struct DeviceGemmXdlSplitKCShuffle ...@@ -391,8 +391,11 @@ struct DeviceGemmXdlSplitKCShuffle
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", " std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl; << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
} }
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
ShowInfo(arg);
const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0); const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_, if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
...@@ -414,51 +417,29 @@ struct DeviceGemmXdlSplitKCShuffle ...@@ -414,51 +417,29 @@ struct DeviceGemmXdlSplitKCShuffle
float ave_time = 0; float ave_time = 0;
const auto Run = [&](const auto& kernel) { const auto Run = [&](const auto& kernel) {
if(nrepeat > 0) hipGetErrorString(hipMemset(
{ arg.p_c_grid_,
ShowInfo(arg); 0,
ave_time = arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
launch_and_time_kernel(kernel, sizeof(CDataType)));
nrepeat,
dim3(grid_size), launch_and_time_kernel(stream_config,
dim3(BlockSize), kernel,
0, dim3(grid_size),
arg.p_a_grid_, dim3(BlockSize),
arg.p_b_grid_, 0,
arg.p_c_grid_, arg.p_a_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_, arg.p_b_grid_,
arg.b_grid_desc_kbatch_k0_n_k1_, arg.p_c_grid_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_, arg.a_grid_desc_kbatch_k0_m_k1_,
arg.a_element_op_, arg.b_grid_desc_kbatch_k0_n_k1_,
arg.b_element_op_, arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.c_element_op_, arg.a_element_op_,
arg.block_2_ctile_map_); arg.b_element_op_,
} arg.c_element_op_,
arg.block_2_ctile_map_);
if(kbatch > 1 || nrepeat <= 0)
{
hipGetErrorString(hipMemset(
arg.p_c_grid_,
0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
}; };
if(has_main_k0_block_loop) if(has_main_k0_block_loop)
{ {
if(kbatch == 1) if(kbatch == 1)
...@@ -542,9 +523,10 @@ struct DeviceGemmXdlSplitKCShuffle ...@@ -542,9 +523,10 @@ struct DeviceGemmXdlSplitKCShuffle
} }
// polymorphic // polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
} }
}; };
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp
View file @ e00a943e
...@@ -449,7 +449,7 @@ struct DeviceGroupedGemmXdl ...@@ -449,7 +449,7 @@ struct DeviceGroupedGemmXdl
{ {
using Argument = DeviceGroupedGemmXdl::Argument; using Argument = DeviceGroupedGemmXdl::Argument;
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
StaticallyIndexedArray<GemmDescKernelArg, MaxGroupCount> gemm_desc_kernel_args; StaticallyIndexedArray<GemmDescKernelArg, MaxGroupCount> gemm_desc_kernel_args;
...@@ -510,8 +510,8 @@ struct DeviceGroupedGemmXdl ...@@ -510,8 +510,8 @@ struct DeviceGroupedGemmXdl
true, true,
MaxGroupCount>; MaxGroupCount>;
ave_time = launch_and_time_kernel(kernel, ave_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.grid_size_), dim3(arg.grid_size_),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -534,8 +534,8 @@ struct DeviceGroupedGemmXdl ...@@ -534,8 +534,8 @@ struct DeviceGroupedGemmXdl
false, false,
MaxGroupCount>; MaxGroupCount>;
ave_time = launch_and_time_kernel(kernel, ave_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.grid_size_), dim3(arg.grid_size_),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -550,9 +550,10 @@ struct DeviceGroupedGemmXdl ...@@ -550,9 +550,10 @@ struct DeviceGroupedGemmXdl
} }
// polymorphic // polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
} }
}; };
......
include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp
View file @ e00a943e
...@@ -204,7 +204,7 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd ...@@ -204,7 +204,7 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
using gridwise_reduce = GridwiseReduction_mk_to_m_threadwise<InDataType, using gridwise_reduce = GridwiseReduction_mk_to_m_threadwise<InDataType,
OutDataType, OutDataType,
...@@ -241,8 +241,8 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd ...@@ -241,8 +241,8 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
const index_t grid_size = (ReduceM / ReduceM_BlockTileSize); const index_t grid_size = (ReduceM / ReduceM_BlockTileSize);
return launch_and_time_kernel(kernel, return launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(grid_size), dim3(grid_size),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -257,9 +257,10 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd ...@@ -257,9 +257,10 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
arg.p_out_indices_dev_); arg.p_out_indices_dev_);
} }
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
} }
}; };
......
include/ck/tensor_operation/gpu/device/device_reduce_blockwise.hpp
View file @ e00a943e
...@@ -211,7 +211,7 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl ...@@ -211,7 +211,7 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto in_grid_desc_m_k = const auto in_grid_desc_m_k =
DeviceReduceBlockWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_); DeviceReduceBlockWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_);
...@@ -253,8 +253,8 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl ...@@ -253,8 +253,8 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl
InElementwiseOperation, InElementwiseOperation,
AccElementwiseOperation>; AccElementwiseOperation>;
avg_time = launch_and_time_kernel(kernel, avg_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.gridSize), dim3(arg.gridSize),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -272,9 +272,10 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl ...@@ -272,9 +272,10 @@ struct DeviceReduceBlockWise : public DeviceReduce<InElementwiseOperation, AccEl
return (avg_time); return (avg_time);
}; };
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}; };
}; };
......
include/ck/tensor_operation/gpu/device/device_reduce_blockwise_second_call.hpp
View file @ e00a943e
...@@ -182,7 +182,7 @@ struct DeviceReduceBlockWiseSecondCall ...@@ -182,7 +182,7 @@ struct DeviceReduceBlockWiseSecondCall
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto in_grid_desc_m_k = DeviceReduceBlockWiseSecondCall::MakeSrc2dDescriptor( const auto in_grid_desc_m_k = DeviceReduceBlockWiseSecondCall::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_); arg.inLengths_, arg.inStrides_);
...@@ -224,8 +224,8 @@ struct DeviceReduceBlockWiseSecondCall ...@@ -224,8 +224,8 @@ struct DeviceReduceBlockWiseSecondCall
InElementwiseOperation, InElementwiseOperation,
AccElementwiseOperation>; AccElementwiseOperation>;
avg_time = launch_and_time_kernel(kernel, avg_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.gridSize), dim3(arg.gridSize),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -243,10 +243,11 @@ struct DeviceReduceBlockWiseSecondCall ...@@ -243,10 +243,11 @@ struct DeviceReduceBlockWiseSecondCall
return (avg_time); return (avg_time);
}; };
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}; }
}; };
bool IsSupportedArgument(const BaseArgument* p_arg) override bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/device/device_reduce_multiblock_atomic_add.hpp
View file @ e00a943e
...@@ -245,7 +245,7 @@ struct DeviceReduceMultiBlockAtomicAdd ...@@ -245,7 +245,7 @@ struct DeviceReduceMultiBlockAtomicAdd
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto in_grid_desc_m_k = DeviceReduceMultiBlockAtomicAdd::MakeSrc2dDescriptor( const auto in_grid_desc_m_k = DeviceReduceMultiBlockAtomicAdd::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.kBlockTileIterations); arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.kBlockTileIterations);
...@@ -275,8 +275,6 @@ struct DeviceReduceMultiBlockAtomicAdd ...@@ -275,8 +275,6 @@ struct DeviceReduceMultiBlockAtomicAdd
float avg_time = 0; float avg_time = 0;
KernelTimer timer;
const auto kernel_pre = kernel_buffer_set_value<BlockSize, OutDataType, OutGridDesc_M>; const auto kernel_pre = kernel_buffer_set_value<BlockSize, OutDataType, OutGridDesc_M>;
const auto kernel_main = kernel_reduce_multiblock_atocmi_add<GridwiseReduce, const auto kernel_main = kernel_reduce_multiblock_atocmi_add<GridwiseReduce,
InDataType, InDataType,
...@@ -287,50 +285,38 @@ struct DeviceReduceMultiBlockAtomicAdd ...@@ -287,50 +285,38 @@ struct DeviceReduceMultiBlockAtomicAdd
InElementwiseOperation, InElementwiseOperation,
AccElementwiseOperation>; AccElementwiseOperation>;
printf("launch_and_time_kernel: grid_dim {%ld, 1, 1}, block_dim {%d, 1, 1} \n", avg_time += launch_and_time_kernel(stream_config,
arg.gridSize, kernel_pre,
BlockSize); dim3(arg.gridSize_pre),
printf("Warm up\n"); dim3(BlockSize),
0,
for(int i = 0; i < nrepeat + 1; i++) out_grid_desc_m,
{ arg.out_dev_,
if(i == 1) static_cast<OutDataType>(0.0f));
timer.Start();
avg_time += launch_and_time_kernel(stream_config,
launch_kernel(kernel_pre, kernel_main,
dim3(arg.gridSize_pre), dim3(arg.gridSize),
dim3(BlockSize), dim3(BlockSize),
0, 0,
out_grid_desc_m, in_grid_desc_m_k,
arg.out_dev_, out_grid_desc_m,
static_cast<OutDataType>(0.0f)); arg.in_elementwise_op_,
arg.acc_elementwise_op_,
launch_kernel(kernel_main, arg.blkGroupSize,
dim3(arg.gridSize), arg.kBlockTileIterations,
dim3(BlockSize), arg.alpha_,
0, arg.in_dev_,
in_grid_desc_m_k, arg.out_dev_);
out_grid_desc_m,
arg.in_elementwise_op_, return avg_time;
arg.acc_elementwise_op_, }
arg.blkGroupSize,
arg.kBlockTileIterations,
arg.alpha_,
arg.in_dev_,
arg.out_dev_);
};
timer.End();
avg_time = timer.GetElapsedTime() / nrepeat;
return (avg_time);
};
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}; }
}; };
bool IsSupportedArgument(const BaseArgument* p_arg) override bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/device/device_reduce_multiblock_partial_reduce.hpp
View file @ e00a943e
...@@ -273,7 +273,7 @@ struct DeviceReduceMultiBlockPartialReduce ...@@ -273,7 +273,7 @@ struct DeviceReduceMultiBlockPartialReduce
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto in_grid_desc_m_k = DeviceReduceMultiBlockPartialReduce::MakeSrc2dDescriptor( const auto in_grid_desc_m_k = DeviceReduceMultiBlockPartialReduce::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.kBlockTileIterations); arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.kBlockTileIterations);
...@@ -313,8 +313,8 @@ struct DeviceReduceMultiBlockPartialReduce ...@@ -313,8 +313,8 @@ struct DeviceReduceMultiBlockPartialReduce
InElementwiseOperation, InElementwiseOperation,
AccElementwiseOperation>; AccElementwiseOperation>;
avg_time = launch_and_time_kernel(kernel, avg_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.gridSize), dim3(arg.gridSize),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -331,10 +331,11 @@ struct DeviceReduceMultiBlockPartialReduce ...@@ -331,10 +331,11 @@ struct DeviceReduceMultiBlockPartialReduce
return (avg_time); return (avg_time);
}; };
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}; }
}; };
bool IsSupportedArgument(const BaseArgument* p_arg) override bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/device/device_reduce_threadwise.hpp
View file @ e00a943e
...@@ -212,7 +212,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE ...@@ -212,7 +212,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, int nrepeat = 1) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto in_grid_desc_m_k = const auto in_grid_desc_m_k =
DeviceReduceThreadWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_); DeviceReduceThreadWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_);
...@@ -254,8 +254,8 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE ...@@ -254,8 +254,8 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE
InElementwiseOperation, InElementwiseOperation,
OutElementwiseOperation>; OutElementwiseOperation>;
avg_time = launch_and_time_kernel(kernel, avg_time = launch_and_time_kernel(stream_config,
nrepeat, kernel,
dim3(arg.gridSize), dim3(arg.gridSize),
dim3(BlockSize), dim3(BlockSize),
0, 0,
...@@ -272,10 +272,11 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE ...@@ -272,10 +272,11 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, OutE
return (avg_time); return (avg_time);
}; };
float Run(const BaseArgument* p_arg, int nrepeat = 1) override float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{ {
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat); return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}; }
}; };
bool IsSupportedArgument(const BaseArgument* p_arg) override bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
View file @ e00a943e
#pragma once #pragma once
#include "common_header.hpp" #include "common_header.hpp"
#include "tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
namespace ck { namespace ck {
...@@ -248,4 +249,116 @@ struct GridwiseGemmPipeline_v1<2> ...@@ -248,4 +249,116 @@ struct GridwiseGemmPipeline_v1<2>
} }
}; };
template <index_t NumPrefetch>
struct GridwiseGemmPipelineInterwave_v1;
template <>
struct GridwiseGemmPipelineInterwave_v1<1>
{
__host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
__host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
{
return num_loop > 1;
}
template <bool HasMainLoop,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename BlockwiseGemm,
typename CThreadBuffer>
static __device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
const BlockwiseGemm& blockwise_gemm,
CThreadBuffer& c_thread_buf,
index_t num_loop)
{
// preload data into LDS
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
block_sync_lds();
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
// block_sync_lds(); // moved into blockwise_gemm
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
++i;
} while(i < (num_loop - 1));
}
// tail
{
block_sync_lds();
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
}
};
// Note: 2 stage prefetch not optimized for inter-wave loop scheduler
template <>
struct GridwiseGemmPipelineInterwave_v1<2> : public GridwiseGemmPipeline_v1<2>
{
};
template <index_t NumPrefetch, LoopScheduler LoopSched>
constexpr auto GridwiseGemmPipeline_v1_Selector()
{
if constexpr(LoopSched == LoopScheduler::Default)
{
return GridwiseGemmPipeline_v1<NumPrefetch>{};
}
else if constexpr(LoopSched == LoopScheduler::Interwave)
{
return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
}
}
} // namespace ck } // namespace ck
library/include/ck/library/host/host_interface.hpp 0 → 100644
View file @ e00a943e
#pragma once
#include <memory>
#include <string>
#include "stream_config.hpp"
#include "config.hpp"
#include "device_base.hpp"
struct DeviceConvFwdPtr_t
{
using BaseArgument = ck::tensor_operation::device::BaseArgument;
using BaseInvoker = ck::tensor_operation::device::BaseInvoker;
struct DeviceConvFwdPtrImpl;
std::unique_ptr<DeviceConvFwdPtrImpl> pImpl;
DeviceConvFwdPtr_t();
~DeviceConvFwdPtr_t();
DeviceConvFwdPtr_t(DeviceConvFwdPtr_t&&);
DeviceConvFwdPtr_t(DeviceConvFwdPtrImpl&);
DeviceConvFwdPtr_t& operator=(DeviceConvFwdPtr_t&) = delete;
DeviceConvFwdPtr_t& operator=(const DeviceConvFwdPtr_t&) = delete;
std::unique_ptr<BaseArgument>
MakeArgumentPointer(void* in_ptr,
void* wei_ptr,
void* out_ptr,
size_t N,
size_t K,
size_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
const; // in,wei and out element ops are ignored for now since even if we change them, they
// cant be linked
std::unique_ptr<BaseInvoker>
MakeInvokerPointer() const; // requires including BaseInvoker headers
std::string GetTypeString();
bool IsSupportedArgument(const BaseArgument* arg_ptr);
};
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances_t(
std::vector<DeviceConvFwdPtr_t>& instances);
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances_t(
std::vector<DeviceConvFwdPtr_t>& instances);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances_t(
std::vector<DeviceConvFwdPtr_t>& instances);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances_t(
std::vector<DeviceConvFwdPtr_t>& instances);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances_t(
std::vector<DeviceConvFwdPtr_t>& instances);
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