ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

tidy up

Browse Source
This commit is contained in:
Chao Liu
2019-04-08 10:48:29 -05:00
parent c9fa46af0b
commit 268d1c717c
14 changed files with 120 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/include/blockwise_4d_tensor_op.hip.hpp
View File

@@ -340,10 +340,11 @@ struct BlockwiseChwnTensorCopyPadded
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
const Float* p_src_tmp =
p_src + src_desc.Get1dIndex(c_block_data_begin,
(ho_block_data_begin + h_block_pad_low) - h_global_pad_low,
(wo_block_data_begin + w_block_pad_low) - w_global_pad_low,
n_block_data_begin);
p_src +
src_desc.Get1dIndex(c_block_data_begin,
(ho_block_data_begin + h_block_pad_low) - h_global_pad_low,
(wo_block_data_begin + w_block_pad_low) - w_global_pad_low,
n_block_data_begin);
#if 0
if(get_thread_local_1d_id() == 0)

5
src/include/blockwise_batched_gemm.hip.hpp
View File

@@ -329,8 +329,9 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
{
threadwise_matrix_copy(
c_thread_sub_mtx,
p_c_thread + c_thread_sub_mtx.Get1dIndex(m_repeat * MPerLevel1Cluster,
n_repeat * NPerLevel1Cluster),
p_c_thread +
c_thread_sub_mtx.Get1dIndex(m_repeat * MPerLevel1Cluster,
n_repeat * NPerLevel1Cluster),
c_block_mtx,
p_c_block +
c_block_mtx.Get1dIndex(m_repeat * MPerLevel1Cluster,

27
src/include/blockwise_direct_convolution.hip.hpp
View File

@@ -93,10 +93,11 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
Float p_out_thread[out_thread_desc.GetElementSpace()];
threadwise_4d_tensor_copy(out_block_desc,
p_out_block + out_block_desc.Get1dIndex(n_thread_data_begin,
k_thread_data_begin,
ho_thread_data_begin,
wo_thread_data_begin),
p_out_block +
out_block_desc.Get1dIndex(n_thread_data_begin,
k_thread_data_begin,
ho_thread_data_begin,
wo_thread_data_begin),
out_thread_desc,
p_out_thread,
out_thread_desc.GetLengths());
@@ -107,10 +108,11 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
// threadwise convolution
threadwise_direct_convolution_2(
in_thread_block_desc,
p_in_block + in_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data_begin,
hi_thread_data_begin,
wi_thread_data_begin),
p_in_block +
in_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data_begin,
hi_thread_data_begin,
wi_thread_data_begin),
wei_thread_block_desc,
p_wei_block +
wei_block_desc.Get1dIndex(k_thread_data_begin, c_thread_data_begin, 0, 0),
@@ -122,10 +124,11 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
threadwise_4d_tensor_copy(out_thread_desc,
p_out_thread,
out_block_desc,
p_out_block + out_block_desc.Get1dIndex(n_thread_data_begin,
k_thread_data_begin,
ho_thread_data_begin,
wo_thread_data_begin),
p_out_block +
out_block_desc.Get1dIndex(n_thread_data_begin,
k_thread_data_begin,
ho_thread_data_begin,
wo_thread_data_begin),
out_thread_desc.GetLengths());
}
}

8
src/include/gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hip.hpp
View File

@@ -183,8 +183,9 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn
threadwise_4d_tensor_set_zero(out_khwn_thread_desc, p_out_thread);
const Float* p_in_global_block_begin =
p_in_global + in_chwn_global_desc.Get1dIndex(
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
p_in_global +
in_chwn_global_desc.Get1dIndex(
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_begin =
p_wei_global + wei_cyxk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
@@ -267,7 +268,8 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 = (GemmNLevel0Cluster * GemmNLevel1Cluster) / (NPerBlock / GemmNPerThreadSubC);
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / (NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;

6
src/include/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hip.hpp
View File

@@ -238,9 +238,9 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn
#elif 1
blockwise_gemm.Run_asm
#endif
(p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block + y * Wi + x,
p_out_thread);
(p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block + y * Wi + x,
p_out_thread);
}
}
}

15
src/include/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp
View File

@@ -387,13 +387,14 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn_lds_double_buffer
constexpr auto out_kb_global_desc = make_ConstantTensorDescriptor(Sequence<K, B>{});
threadwise_6d_tensor_copy(out_6d_thread_desc,
p_out_thread,
out_6d_global_desc,
p_out_global + out_kb_global_desc.Get1dIndex(
k_thread_data_begin, b_thread_data_begin),
out_6d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite>{});
threadwise_6d_tensor_copy(
out_6d_thread_desc,
p_out_thread,
out_6d_global_desc,
p_out_global +
out_kb_global_desc.Get1dIndex(k_thread_data_begin, b_thread_data_begin),
out_6d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite>{});
}
else
#endif

19
src/include/gridwise_direct_convolution_1.hip.hpp
View File

@@ -113,10 +113,11 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
c_block_work_begin += CPerBlock)
{
// copy input tensor to LDS
blockwise_in_copy.Run(p_in_global + in_global_desc.Get1dIndex(n_block_work_begin,
c_block_work_begin,
hi_block_work_begin,
wi_block_work_begin),
blockwise_in_copy.Run(p_in_global +
in_global_desc.Get1dIndex(n_block_work_begin,
c_block_work_begin,
hi_block_work_begin,
wi_block_work_begin),
p_in_block);
// copy weight tensor to LDS
@@ -143,9 +144,9 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
}
// copy output tensor from LDS to device mem
blockwise_out_copy.Run(p_out_block,
p_out_global + out_global_desc.Get1dIndex(n_block_work_begin,
k_block_work_begin,
ho_block_work_begin,
wo_block_work_begin));
blockwise_out_copy.Run(
p_out_block,
p_out_global +
out_global_desc.Get1dIndex(
n_block_work_begin, k_block_work_begin, ho_block_work_begin, wo_block_work_begin));
}

43
src/include/gridwise_direct_convolution_2_nchw_kcyx_nkhw.hip.hpp
View File

@@ -175,16 +175,18 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
c_block_data_begin += CPerBlock, __syncthreads())
{
// copy input tensor to LDS
blockwise_in_copy.Run(p_in_global + in_nchw_global_desc.Get1dIndex(n_block_data_begin,
c_block_data_begin,
hi_block_data_begin,
wi_block_data_begin),
blockwise_in_copy.Run(p_in_global +
in_nchw_global_desc.Get1dIndex(n_block_data_begin,
c_block_data_begin,
hi_block_data_begin,
wi_block_data_begin),
p_in_block);
// copy weight tensor to LDS
blockwise_wei_copy.Run(p_wei_global + wei_kcyx_global_desc.Get1dIndex(
k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_block);
blockwise_wei_copy.Run(
p_wei_global +
wei_kcyx_global_desc.Get1dIndex(k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_block);
__syncthreads();
@@ -194,10 +196,11 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
#if 1
threadwise_direct_convolution_2(
in_nchw_thread_block_desc,
p_in_block + in_nchw_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
p_in_block +
in_nchw_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_thread_block_desc,
p_wei_block +
wei_kcyx_block_desc.Get1dIndex(k_thread_data_begin, c_thread_data, 0, 0),
@@ -206,10 +209,11 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
#elif 0
threadwise_direct_convolution_3(
in_nchw_thread_block_desc,
p_in_block + in_nchw_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
p_in_block +
in_nchw_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_thread_block_desc,
p_wei_block +
wei_kcyx_block_desc.Get1dIndex(k_thread_data_begin, c_thread_data, 0, 0),
@@ -224,9 +228,10 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
out_nkhw_thread_desc,
p_out_thread,
out_nkhw_global_desc,
p_out_global + out_nkhw_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
p_out_global +
out_nkhw_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_nkhw_thread_desc.GetLengths());
}

34
src/include/gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hip.hpp
View File

@@ -198,9 +198,10 @@ __global__ void gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw(
p_in_vec_block);
// copy weight tensor to LDS
blockwise_wei_copy.Run(p_wei_vec_global + wei_kcyx_vec_global_desc.Get1dIndex(
k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_vec_block);
blockwise_wei_copy.Run(
p_wei_vec_global +
wei_kcyx_vec_global_desc.Get1dIndex(k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_vec_block);
__syncthreads();
@@ -210,10 +211,11 @@ __global__ void gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw(
#if 1
threadwise_direct_convolution_2(
in_nchw_vec_thread_block_desc,
p_in_vec_block + in_nchw_vec_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
p_in_vec_block +
in_nchw_vec_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_vec_thread_block_desc,
p_wei_vec_block +
wei_kcyx_vec_block_desc.Get1dIndex(k_thread_data_begin, c_thread_data, 0, 0),
@@ -222,10 +224,11 @@ __global__ void gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw(
#elif 0
threadwise_direct_convolution_3(
in_nchw_vec_thread_block_desc,
p_in_vec_block + in_nchw_vec_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
p_in_vec_block +
in_nchw_vec_block_desc.Get1dIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_vec_thread_block_desc,
p_wei_vec_block +
wei_kcyx_vec_block_desc.Get1dIndex(k_thread_data_begin, c_thread_data, 0, 0),
@@ -240,9 +243,10 @@ __global__ void gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw(
out_nkhw_thread_desc,
p_out_thread,
out_nkhw_global_desc,
p_out_global + out_nkhw_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
p_out_global +
out_nkhw_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_nkhw_thread_desc.GetLengths());
}

9
src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hip.hpp
View File

@@ -283,10 +283,11 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(
out_hkwn_thread_desc,
p_out_thread,
out_khwn_global_desc,
p_out_global + out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
p_out_global +
out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_hkwn_thread_desc.GetLengths(),
reorder_khwn_from_hkwn);
}

3
src/include/tensor.hpp
View File

@@ -22,8 +22,7 @@ std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
return os;
}
typedef enum
{
typedef enum {
Half = 0,
Float = 1,
} DataType_t;