ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor

Browse Source
This commit is contained in:
Chao Liu
2019-01-16 15:45:02 -06:00
parent 2b52fbd24a
commit 5872b710df
6 changed files with 171 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/include/blockwise_tensor_op.cuh
View File

@@ -135,20 +135,6 @@ __device__ void blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_ds
const unsigned bindex = dst_desc.Get1dIndex(did[IR0], did[IR1], did[IR2], did[IR3]);
#if 0
printf("did %u %u %u %u, did_IR %u %u %u %u, index %u %u\n",
did[0],
did[1],
did[2],
did[3],
did[IR0],
did[IR1],
did[IR2],
did[IR3],
aindex,
bindex);
#endif
f(p_src[aindex], p_dst[bindex]);
}

9
src/include/gridwise_implicit_gemm_convolution_nchw_kcsr.cuh
View File

@@ -157,7 +157,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
c_block_data_begin += CPerBlock, __syncthreads())
{
// input: global mem to LDS,
// convert 4d-tensor in[N,C,Hi,Wi] to matrix in_matrix[C,Hi*Wi*N]
// convert [N,C,Hi,Wi] to [C,Hi,Wi,N]
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>(
in_nchw_global_desc,
p_in_global + in_nchw_global_desc.Get1dIndex(n_block_data_begin,
@@ -170,6 +170,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
reorder_chwn_from_nchw);
// weight: global mem to LDS,
// convert [K,C,S,R] to [S,R,C,K]
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>(
wei_kcsr_global_desc,
p_wei_global +
@@ -217,10 +218,4 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
wo_block_data_begin + wo_thread_data_begin),
out_hkwn_thread_desc.GetLengths(),
reorder_nkhw_from_hkwn);
// printf("%f %f %f %f\n", p_out_thread[0], p_out_thread[1], p_out_thread[2], p_out_thread[3]);
// printf("%u %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(),
// matrix_c_index.batch_begin, matrix_c_index.row_begin, matrix_c_index.col_begin); printf("%u
// %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(), ho_thread_data_begin,
// k_thread_data_begin, wo_thread_data_begin);
}

9
src/include/gridwise_implicit_gemm_convolution_nchw_srck.cuh
View File

@@ -153,7 +153,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_srck(InGlobalDesc,
c_block_data_begin += CPerBlock, __syncthreads())
{
// input: global mem to LDS,
// convert 4d-tensor in[N,C,Hi,Wi] to matrix in_matrix[C,Hi*Wi*N]
// convert [N,C,Hi,Wi] to [C,Hi,Wi,N]
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>(
in_nchw_global_desc,
p_in_global + in_nchw_global_desc.Get1dIndex(n_block_data_begin,
@@ -166,6 +166,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_srck(InGlobalDesc,
reorder_chwn_from_nchw);
// weight: global mem to LDS,
// format is [S,R,C,K], no conversion needed
blockwise_4d_tensor_copy<BlockSize>(
wei_srck_global_desc,
p_wei_global +
@@ -212,10 +213,4 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_srck(InGlobalDesc,
wo_block_data_begin + wo_thread_data_begin),
out_hkwn_thread_desc.GetLengths(),
reorder_nkhw_from_hkwn);
// printf("%f %f %f %f\n", p_out_thread[0], p_out_thread[1], p_out_thread[2], p_out_thread[3]);
// printf("%u %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(),
// matrix_c_index.batch_begin, matrix_c_index.row_begin, matrix_c_index.col_begin); printf("%u
// %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(), ho_thread_data_begin,
// k_thread_data_begin, wo_thread_data_begin);
}