diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2017-03-19 16:04:04 +0100 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2017-03-19 16:04:04 +0100 |
commit | c27d2f0c1ea69820f39d440f307c7bc3f97472c4 (patch) | |
tree | 772e8166917905ec800586ca6e2b9e120c3ebf78 /src/kernels/level3/xgemm_batched.opencl | |
parent | 2fd04dae83acb01933856e768a938db9ac808ce0 (diff) |
Added an (optional) non-direct implementation of the batched GEMM routine
Diffstat (limited to 'src/kernels/level3/xgemm_batched.opencl')
-rw-r--r-- | src/kernels/level3/xgemm_batched.opencl | 70 |
1 files changed, 70 insertions, 0 deletions
diff --git a/src/kernels/level3/xgemm_batched.opencl b/src/kernels/level3/xgemm_batched.opencl new file mode 100644 index 00000000..c7bf10d5 --- /dev/null +++ b/src/kernels/level3/xgemm_batched.opencl @@ -0,0 +1,70 @@ + +// ================================================================================================= +// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This +// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- +// width of 100 characters per line. +// +// Author(s): +// Cedric Nugteren <www.cedricnugteren.nl> +// +// This file contains the batched version of the non-direct GEMM kernel. See part 1 for information +// about the non-batched version of the kernel. +// +// ================================================================================================= + +// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string +// literal). Comment-out this line for syntax-highlighting when developing. +R"( + +// ================================================================================================= + +// Main entry point of the kernel. This is the regular full version. +__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1))) +void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK, + const __constant real_arg* arg_alphas, + const __constant real_arg* arg_betas, + const __global realM* restrict agm, const int a_one, const int a_two, + const __global realN* restrict bgm, const int b_one, const int b_two, + __global realM* cgm, const int c_one, const int c_two) { + const int batch = get_group_id(2); + const real alpha = GetRealArg(arg_alphas[batch]); + const real beta = GetRealArg(arg_betas[batch]); + + // Sets the offsets + const int a_offset = batch * a_one * a_two; + const int b_offset = batch * b_one * b_two; + const int c_offset = batch * c_one * c_two; + const __global realM* restrict agm_ = &agm[a_offset / VWM]; + const __global realN* restrict bgm_ = &bgm[b_offset / VWN]; + __global realM* restrict cgm_ = &cgm[c_offset / VWM]; + + // Allocates workgroup-private memory (local memory) + #if SA == 1 + __local realM alm[KWG * MWG/VWM]; + #endif + #if SB == 1 + __local realN blm[KWG * NWG/VWN]; + #endif + + // Computes the matrix-multiplication and stores the result in register memory + realM cpm[NWI][MWI/VWM]; + #if SA == 1 && SB == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, cpm, alm, blm); + #elif SA == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, cpm, alm); + #elif SB == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, cpm, blm); + #else + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, cpm); + #endif + + // Stores an MWG * NWG tile of results and performs the multiplication with alpha and beta + StoreResults(cgm_, cpm, kSizeM, alpha, beta); +} + +// ================================================================================================= + +// End of the C++11 raw string literal +)" + +// ================================================================================================= |