haraldwolff
/
CLBlast
mirror of https://github.com/CNugteren/CLBlast.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Updated according to feedback from CNugteren

pull/451/head
Angus, Alexander 2023-01-17 08:35:29 -08:00
parent ff6a5689df
commit 73f49e9b3d
36 changed files with 82 additions and 79 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
Adreno_730_tuning_results.zip
View File

Binary file not shown.

BIN
Adreno_740_tuning_results.zip
View File

Binary file not shown.

5
CHANGELOG
View File

@ -1,5 +1,10 @@
Development version (next version)
- Fixes two small issues in the plotting script
- Modifications to improve performance on Qualcomm Adreno GPUs:
* Unique database entries for specific Adreno devices
* Toggle OpenCL kernel compilation options for Adreno
* New preprocessor directive RELAX_WORKGROUP_SIZE
- Fixed a bug in handling of #undef in CLBlast loop unrolling and array-to-register mapping functions
Version 1.5.3
- Fix a correctness issue with DGEMM on SM 7.5 Turing GPUs

2
README.md
View File

@ -101,8 +101,6 @@ Known performance related issues:
* Severe performance issues with Beignet v1.3.0 due to missing support for local memory. Please downgrade to v1.2.1 or upgrade to v1.3.1 or newer.
* Performance issues on Qualcomm Adreno GPUs.
Other known issues:
* Routines returning an integer are currently not properly tested for half-precision FP16: IHAMAX/IHAMIN/IHMAX/IHMIN

4
src/kernels/level1/xamax.opencl
View File

@ -32,7 +32,7 @@ R"(
// The main reduction kernel, performing the loading and the majority of the operation
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
#endif
void Xamax(const int n,
@ -102,7 +102,7 @@ void Xamax(const int n,
// be launched with a single workgroup only.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
#endif
void XamaxEpilogue(const __global singlereal* restrict maxgm,

4
src/kernels/level1/xasum.opencl
View File

@ -32,7 +32,7 @@ R"(
// The main reduction kernel, performing the loading and the majority of the operation
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
#endif
void Xasum(const int n,
@ -79,7 +79,7 @@ void Xasum(const int n,
// be launched with a single workgroup only.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
#endif
void XasumEpilogue(const __global real* restrict input,

8
src/kernels/level1/xaxpy.opencl
View File

@ -24,7 +24,7 @@ R"(
// Full version of the kernel with offsets and strided accesses
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void Xaxpy(const int n, const real_arg arg_alpha,
@ -43,7 +43,7 @@ void Xaxpy(const int n, const real_arg arg_alpha,
// assumes that 'n' is dividable by 'VW' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XaxpyFaster(const int n, const real_arg arg_alpha,
@ -67,7 +67,7 @@ void XaxpyFaster(const int n, const real_arg arg_alpha,
// dividable by 'VW', 'WGS' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XaxpyFastest(const int n, const real_arg arg_alpha,
@ -89,7 +89,7 @@ void XaxpyFastest(const int n, const real_arg arg_alpha,
// Full version of the kernel with offsets and strided accesses: batched version
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XaxpyBatched(const int n, const __constant real_arg* arg_alphas,

4
src/kernels/level1/xcopy.opencl
View File

@ -24,7 +24,7 @@ R"(
// Full version of the kernel with offsets and strided accesses
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void Xcopy(const int n,
@ -43,7 +43,7 @@ void Xcopy(const int n,
// dividable by 'VW', 'WGS' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XcopyFast(const int n,

4
src/kernels/level1/xdot.opencl
View File

@ -32,7 +32,7 @@ R"(
// The main reduction kernel, performing the multiplication and the majority of the sum operation
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
#endif
void Xdot(const int n,
@ -78,7 +78,7 @@ void Xdot(const int n,
// be launched with a single workgroup only.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
#endif
void XdotEpilogue(const __global real* restrict input,

6
src/kernels/level1/xhad.opencl
View File

@ -68,7 +68,7 @@ INLINE_FUNC realV MultiplyVectorVector(realV cvec, const realV aval, const realV
// Full version of the kernel with offsets and strided accesses
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void Xhad(const int n, const real_arg arg_alpha, const real_arg arg_beta,
@ -96,7 +96,7 @@ void Xhad(const int n, const real_arg arg_alpha, const real_arg arg_beta,
// assumes that 'n' is dividable by 'VW' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XhadFaster(const int n, const real_arg arg_alpha, const real_arg arg_beta,
@ -127,7 +127,7 @@ void XhadFaster(const int n, const real_arg arg_alpha, const real_arg arg_beta,
// dividable by 'VW', 'WGS' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XhadFastest(const int n, const real_arg arg_alpha, const real_arg arg_beta,

4
src/kernels/level1/xnrm2.opencl
View File

@ -32,7 +32,7 @@ R"(
// The main reduction kernel, performing the multiplication and the majority of the operation
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
#endif
void Xnrm2(const int n,
@ -77,7 +77,7 @@ void Xnrm2(const int n,
// be launched with a single workgroup only.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
#endif
void Xnrm2Epilogue(const __global real* restrict input,

4
src/kernels/level1/xscal.opencl
View File

@ -24,7 +24,7 @@ R"(
// Full version of the kernel with offsets and strided accesses
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void Xscal(const int n, const real_arg arg_alpha,
@ -46,7 +46,7 @@ void Xscal(const int n, const real_arg arg_alpha,
// dividable by 'VW', 'WGS' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XscalFast(const int n, const real_arg arg_alpha,

4
src/kernels/level1/xswap.opencl
View File

@ -24,7 +24,7 @@ R"(
// Full version of the kernel with offsets and strided accesses
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void Xswap(const int n,
@ -45,7 +45,7 @@ void Xswap(const int n,
// dividable by 'VW', 'WGS' and 'WPT'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
#endif
void XswapFast(const int n,

2
src/kernels/level2/xgemv.opencl
View File

@ -212,7 +212,7 @@ INLINE_FUNC real LoadMatrixA(const __global real* restrict agm, const int x, con
// Full version of the kernel
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
#endif
void Xgemv(const int m, const int n,

4
src/kernels/level2/xgemv_fast.opencl
View File

@ -90,7 +90,7 @@ INLINE_FUNC realVF LoadMatrixAVF(const __global realVF* restrict agm, const int
// --> 'do_conjugate' is 0
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
#endif
void XgemvFast(const int m, const int n,
@ -197,7 +197,7 @@ void XgemvFast(const int m, const int n,
// --> 'do_conjugate' is 0
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS3, 1, 1)))
#endif
void XgemvFastRot(const int m, const int n,

2
src/kernels/level2/xger.opencl
View File

@ -20,7 +20,7 @@ R"(
// Regular version of the rank-1 matrix update kernel (GER, GERU, GERC)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
#endif
void Xger(const int max1, const int max2,

2
src/kernels/level2/xher.opencl
View File

@ -20,7 +20,7 @@ R"(
// Symmetric version of the rank-1 matrix update kernel (HER, HPR, SYR, SPR)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
#endif
void Xher(const int n,

2
src/kernels/level2/xher2.opencl
View File

@ -20,7 +20,7 @@ R"(
// Symmetric version of the rank-2 matrix update kernel (HER2, HPR2, SYR2, SPR2)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
#endif
void Xher2(const int n,

4
src/kernels/level2/xtrsv.opencl
View File

@ -41,7 +41,7 @@ void FillVector(const int n, const int inc, const int offset,
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
#endif
void trsv_forward(int n,
@ -93,7 +93,7 @@ void trsv_forward(int n,
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
#endif
void trsv_backward(int n,

4
src/kernels/level3/convert_hermitian.opencl
View File

@ -23,7 +23,7 @@ R"(
// stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void HermLowerToSquared(const int src_dim,
@ -66,7 +66,7 @@ void HermLowerToSquared(const int src_dim,
// Same as above, but now the matrix' data is stored in the upper-triangle
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void HermUpperToSquared(const int src_dim,

4
src/kernels/level3/convert_symmetric.opencl
View File

@ -22,7 +22,7 @@ R"(
// stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void SymmLowerToSquared(const int src_dim,
@ -59,7 +59,7 @@ void SymmLowerToSquared(const int src_dim,
// Same as above, but now the matrix' data is stored in the upper-triangle
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void SymmUpperToSquared(const int src_dim,

4
src/kernels/level3/convert_triangular.opencl
View File

@ -22,7 +22,7 @@ R"(
// stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void TriaLowerToSquared(const int src_dim,
@ -61,7 +61,7 @@ void TriaLowerToSquared(const int src_dim,
// Same as above, but now the matrix' data is stored in the upper-triangle
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void TriaUpperToSquared(const int src_dim,

2
src/kernels/level3/copy_fast.opencl
View File

@ -37,7 +37,7 @@ R"(
// COPY_VW. Also requires both matrices to be of the same dimensions and without offset.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
#endif
void CopyMatrixFast(const int ld,

12
src/kernels/level3/copy_pad.opencl
View File

@ -61,7 +61,7 @@ INLINE_FUNC void _CopyPadMatrix(const int src_one, const int src_two,
// Interface to the above function
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyPadMatrix(const int src_one, const int src_two,
@ -124,7 +124,7 @@ INLINE_FUNC void _CopyMatrix(const int src_one, const int src_two,
// Interface to the above function
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyMatrix(const int src_one, const int src_two,
@ -148,7 +148,7 @@ void CopyMatrix(const int src_one, const int src_two,
// Batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyPadMatrixBatched(const int src_one, const int src_two,
@ -170,7 +170,7 @@ void CopyPadMatrixBatched(const int src_one, const int src_two,
// Batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyMatrixBatched(const int src_one, const int src_two,
@ -195,7 +195,7 @@ void CopyMatrixBatched(const int src_one, const int src_two,
// Strided-batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyPadMatrixStridedBatched(const int src_one, const int src_two,
@ -217,7 +217,7 @@ void CopyPadMatrixStridedBatched(const int src_one, const int src_two,
// Strided-batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
#endif
void CopyMatrixStridedBatched(const int src_one, const int src_two,

2
src/kernels/level3/invert_diagonal_blocks_part1.opencl
View File

@ -84,7 +84,7 @@ R"(
// Inverts a diagonal block of INTERNAL_BLOCK_SIZE by INTERNAL_BLOCK_SIZE elements in a larger matrix
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(INTERNAL_BLOCK_SIZE, 1, 1)))
#endif
void InvertDiagonalBlock(const int n, __global const real* restrict src, const int src_offset, const int src_ld,

2
src/kernels/level3/transpose_fast.opencl
View File

@ -38,7 +38,7 @@ R"(
// offset. A more general version is available in 'padtranspose.opencl'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(TRA_DIM, TRA_DIM, 1)))
#endif
void TransposeMatrixFast(const int ld,

12
src/kernels/level3/transpose_pad.opencl
View File

@ -86,7 +86,7 @@ INLINE_FUNC void _TransposePadMatrix(LOCAL_PTR real* tile,
// Interface to the above function
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposePadMatrix(const int src_one, const int src_two,
@ -178,7 +178,7 @@ INLINE_FUNC void _TransposeMatrix(LOCAL_PTR real* tile,
// Interface to the above function
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposeMatrix(const int src_one, const int src_two,
@ -203,7 +203,7 @@ void TransposeMatrix(const int src_one, const int src_two,
// Batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposePadMatrixBatched(const int src_one, const int src_two,
@ -226,7 +226,7 @@ void TransposePadMatrixBatched(const int src_one, const int src_two,
// Batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposeMatrixBatched(const int src_one, const int src_two,
@ -252,7 +252,7 @@ void TransposeMatrixBatched(const int src_one, const int src_two,
// Strided-batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposePadMatrixStridedBatched(const int src_one, const int src_two,
@ -275,7 +275,7 @@ void TransposePadMatrixStridedBatched(const int src_one, const int src_two,
// Strided-batched version of the above
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
#endif
void TransposeMatrixStridedBatched(const int src_one, const int src_two,

4
src/kernels/level3/xgemm_batched.opencl
View File

@ -21,7 +21,7 @@ R"(
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
#endif
void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK,
@ -68,7 +68,7 @@ void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK,
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
#endif
void XgemmStridedBatched(const int kSizeM, const int kSizeN, const int kSizeK,

16
src/kernels/level3/xgemm_direct_batched.opencl
View File

@ -22,7 +22,7 @@ R"(
// Direct version of the batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -47,7 +47,7 @@ void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the batched GEMM kernel with [A, B] = [non-transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
@ -72,7 +72,7 @@ void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the batched GEMM kernel with [A, B] = [transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -97,7 +97,7 @@ void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the batched GEMM kernel with [A, B] = [transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
@ -126,7 +126,7 @@ void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -149,7 +149,7 @@ void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int k
// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
@ -172,7 +172,7 @@ void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int k
// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -195,7 +195,7 @@ void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int k
// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectStridedBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,

8
src/kernels/level3/xgemm_direct_part3.opencl
View File

@ -220,7 +220,7 @@ INLINE_FUNC void XgemmDirect(const int kSizeM, const int kSizeN, const int kSize
// Direct version of the GEMM kernel with [A, B] = [non-transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -239,7 +239,7 @@ void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the GEMM kernel with [A, B] = [non-transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK,
@ -258,7 +258,7 @@ void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the GEMM kernel with [A, B] = [transposed, non-transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK,
@ -277,7 +277,7 @@ void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK,
// Direct version of the GEMM kernel with [A, B] = [transposed, transposed]
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XgemmDirectTT(const int kSizeM, const int kSizeN, const int kSizeK,

6
src/kernels/level3/xgemm_part4.opencl
View File

@ -21,7 +21,7 @@ R"(
// Main entry point of the kernel. This is the upper-triangular version.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
#endif
void XgemmUpper(const int kSizeN, const int kSizeK,
@ -61,7 +61,7 @@ void XgemmUpper(const int kSizeN, const int kSizeK,
// Main entry point of the kernel. This is the lower-triangular version.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
#endif
void XgemmLower(const int kSizeN, const int kSizeK,
@ -105,7 +105,7 @@ void XgemmLower(const int kSizeN, const int kSizeK,
// Main entry point of the kernel. This is the regular full version.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
#endif
void Xgemm(const int kSizeM, const int kSizeN, const int kSizeK,

4
src/kernels/levelx/col2im.opencl
View File

@ -94,7 +94,7 @@ INLINE_FUNC void Xcol2im(const int input_h, const int input_w, const int channel
// Kernel flip version of the Xcol2im kernel (for convolution)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
#endif
void Xcol2imKernelFlip(const int input_h, const int input_w, const int channels,
@ -119,7 +119,7 @@ void Xcol2imKernelFlip(const int input_h, const int input_w, const int channels,
// Normal version of the Xcol2im kernel (for cross-correlation)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
#endif
void Xcol2imKernelNormal(const int input_h, const int input_w, const int channels,

4
src/kernels/levelx/im2col.opencl
View File

@ -76,7 +76,7 @@ INLINE_FUNC void Xim2col(const int input_h, const int input_w, const int channel
// Kernel flip version of the Xim2col kernel (for convolution)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
#endif
void Xim2colKernelFlip(const int input_h, const int input_w, const int channels,
@ -97,7 +97,7 @@ void Xim2colKernelFlip(const int input_h, const int input_w, const int channels,
// Normal version of the Xim2col kernel (for cross-correlation)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
#endif
void Xim2colKernelNormal(const int input_h, const int input_w, const int channels,

6
src/kernels/levelx/xconvgemm_part2.opencl
View File

@ -25,7 +25,7 @@ R"(
#if defined(CONVGEMM_WITH_IM2COL)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size,
@ -291,7 +291,7 @@ INLINE_FUNC void Xconvgemm(const int num_patches, const int num_kernels, const i
#if !defined(CONVGEMM_WITH_IM2COL)
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch_size,
@ -316,7 +316,7 @@ void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#elif
#else
__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
#endif
void XconvgemmNormal(const int num_patches, const int num_kernels, const int patch_size,

4
src/utilities/compile.cpp
View File

@ -43,7 +43,7 @@ std::shared_ptr<Program> CompileFromSource(
// For specific devices, use the non-IEE754 compliant OpenCL mad() instruction. This can improve
// performance, but might result in a reduced accuracy.
if ((device.IsAMD() && device.IsGPU()) || device.IsQualcomm()) {
if ((device.IsAMD() && device.IsGPU()) || (device.IsQualcomm() && device.IsGPU())) {
header_string += "#define USE_CL_MAD 1\n";
}
@ -54,7 +54,7 @@ std::shared_ptr<Program> CompileFromSource(
// For specific devices add a global synchronisation barrier to the GEMM kernel to optimize
// performance through better cache behaviour
if ((device.IsARM() && device.IsGPU()) || device.IsQualcomm()) {
if ((device.IsARM() && device.IsGPU()) || (device.IsQualcomm() && device.IsGPU())) {
header_string += "#define GLOBAL_MEM_FENCE 1\n";
}

2
src/utilities/utilities.cpp
View File

@ -463,7 +463,7 @@ std::string GetDeviceArchitecture(const Device& device) {
else if (device.HasExtension(kKhronosAttributesAMD)) {
device_architecture = device.Name(); // Name is architecture for AMD APP and AMD ROCm
}
else if (device.IsQualcomm()) { // queries the Adreno GPU architecture version
else if ((device.IsQualcomm() && device.IsGPU())) { // queries the Adreno GPU architecture version
device_architecture = device.AdrenoVersion();
}
// Note: no else - 'device_architecture' might be the empty string