ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

tweak

Browse Source 
[ROCm/composable_kernel commit: adc1008836]
This commit is contained in:
Chao Liu
2019-08-03 15:05:25 -05:00
parent 783cc11cef
commit 0319e69e81
6 changed files with 27 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
View File

@@ -233,8 +233,6 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k1_b0b1_thread_mtx_desc, p_out_thread);
const Float* p_wei_block_on_global = p_wei_global;
for(index_t e_block_data_begin = 0; e_block_data_begin < E; e_block_data_begin += EPerBlock)
{
blockwise_in_copy.Run(p_in_global, p_in_block);
@@ -246,8 +244,8 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw
__syncthreads();
blockwise_in_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_wei_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_in_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
blockwise_wei_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
}
// copy output: register to global memory
@@ -304,8 +302,9 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw
{
threadwise_out_copy.Run(p_out_thread, p_out_global);
threadwise_out_copy.MoveSrcSlicingWindow({0, 0, GemmNPerThreadSubC}, true);
threadwise_out_copy.MoveDstSlicingWindow({0, 0, B1}, true);
threadwise_out_copy.MoveSrcSlicingWindow(Sequence<0, 0, GemmNPerThreadSubC>{},
True);
threadwise_out_copy.MoveDstSlicingWindow(Sequence<0, 0, B1>{}, True);
}
}
}

20
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer.hpp
View File

@@ -233,8 +233,6 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k1_b0b1_thread_mtx_desc, p_out_thread);
const Float* p_wei_block_on_global = p_wei_global;
// LDS double buffer: preload data into LDS
{
blockwise_in_copy.Run(p_in_global, p_in_block_double);
@@ -263,15 +261,14 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
Float p_in_register_buffer[blockwise_in_copy.GetRegisterBufferSize()];
Float p_wei_register_buffer[blockwise_wei_copy.GetRegisterBufferSize()];
blockwise_in_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_wei_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_in_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
blockwise_wei_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterBuffer(p_in_global, p_in_register_buffer);
blockwise_wei_copy.RunLoadRegisterBuffer(p_wei_block_on_global,
p_wei_register_buffer);
blockwise_wei_copy.RunLoadRegisterBuffer(p_wei_global, p_wei_register_buffer);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_now, p_in_block_now, p_out_thread);
@@ -288,14 +285,14 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
Float p_wei_register_buffer[blockwise_wei_copy.GetRegisterBufferSize()];
// even iteration
blockwise_in_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_wei_copy.MoveSrcSlicingWindow({EPerBlock, 0}, true);
blockwise_in_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
blockwise_wei_copy.MoveSrcSlicingWindow(Sequence<EPerBlock, 0>{}, True);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterBuffer(p_in_global, p_in_register_buffer);
blockwise_wei_copy.RunLoadRegisterBuffer(p_wei_block_on_global, p_wei_register_buffer);
blockwise_wei_copy.RunLoadRegisterBuffer(p_wei_global, p_wei_register_buffer);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_double, p_in_block_double, p_out_thread);
@@ -369,8 +366,9 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
{
threadwise_out_copy.Run(p_out_thread, p_out_global);
threadwise_out_copy.MoveSrcSlicingWindow({0, 0, GemmNPerThreadSubC}, true);
threadwise_out_copy.MoveDstSlicingWindow({0, 0, B1}, true);
threadwise_out_copy.MoveSrcSlicingWindow(Sequence<0, 0, GemmNPerThreadSubC>{},
True);
threadwise_out_copy.MoveDstSlicingWindow(Sequence<0, 0, B1>{}, True);
}
}
}

14
composable_kernel/include/tensor_operation/blockwise_generic_tensor_slice_copy.hpp
View File

@@ -447,17 +447,19 @@ struct BlockwiseGenericTensorSliceCopy_v2
}
template <class T, bool PositiveDirection>
__device__ void MoveSrcSlicingWindow(T step_sizes, integral_constant<bool, PositiveDirection>)
__device__ void
MoveSrcSlicingWindow(T step_sizes,
integral_constant<bool, PositiveDirection> positive_direction)
{
mThreadwiseLoad.MoveSrcSlicingWindow(step_sizes,
integral_constant<bool, PositiveDirection>{});
mThreadwiseLoad.MoveSrcSlicingWindow(step_sizes, positive_direction);
}
template <class T, bool PositiveDirection>
__device__ void MoveDstSlicingWindow(T step_sizes, integral_constant<bool, PositiveDirection>)
__device__ void
MoveDstSlicingWindow(T step_sizes,
integral_constant<bool, PositiveDirection> positive_direction)
{
mThreadwiseLoad.MoveDstSlicingWindow(step_sizes,
integral_constant<bool, PositiveDirection>{});
mThreadwiseLoad.MoveDstSlicingWindow(step_sizes, positive_direction);
}
private:

6
composable_kernel/include/tensor_operation/threadwise_generic_tensor_slice_copy.hpp
View File

@@ -227,9 +227,9 @@ struct ThreadwiseGenericTensorSliceCopy_v2
template <class T, bool PositiveDirection>
__device__ void MoveDstSlicingWindow(T step_sizes, integral_constant<bool, PositiveDirection>)
{
static_if<PositiveDirection>([&](auto) { mDstSliceOrigin += step_sizes; }).Else([&](auto) {
mDstSliceOrigin -= step_sizes;
});
static_if<PositiveDirection>{}([&](auto) {
mDstSliceOrigin += step_sizes;
}).Else([&](auto) { mDstSliceOrigin -= step_sizes; });
}
// private: