diff options
27 files changed, 1485 insertions, 154 deletions
@@ -10,6 +10,9 @@ Development (next version) - Improved compilation time by splitting the tuning database into multiple compilation units - Various minor fixes and enhancements - Added tuned parameters for various devices (see README) +- Added a strided-batched (not part of the BLAS standard) routine, faster but less generic compared + to the existing xGEMMBATCHED routines: + * SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED Version 1.2.0 - Fixed a bug in the TRSM/TRSV routines due to missing synchronisations after GEMM/GEMV calls diff --git a/CMakeLists.txt b/CMakeLists.txt index 53944b25..64f258c5 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -202,7 +202,7 @@ set(LEVEL1_ROUTINES xswap xscal xcopy xaxpy xdot xdotu xdotc xnrm2 xasum xamax) set(LEVEL2_ROUTINES xgemv xgbmv xhemv xhbmv xhpmv xsymv xsbmv xspmv xtrmv xtbmv xtpmv xtrsv xger xgeru xgerc xher xhpr xher2 xhpr2 xsyr xspr xsyr2 xspr2) set(LEVEL3_ROUTINES xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm xtrsm) -set(LEVELX_ROUTINES xomatcopy xim2col xaxpybatched xgemmbatched) +set(LEVELX_ROUTINES xomatcopy xim2col xaxpybatched xgemmbatched xgemmstridedbatched) set(ROUTINES ${LEVEL1_ROUTINES} ${LEVEL2_ROUTINES} ${LEVEL3_ROUTINES} ${LEVELX_ROUTINES}) set(PRECISIONS 32 64 3232 6464 16) diff --git a/doc/clblast.md b/doc/clblast.md index 5ee601f5..ce6f0906 100644 --- a/doc/clblast.md +++ b/doc/clblast.md @@ -3182,6 +3182,108 @@ Requirements for GEMMBATCHED: +xGEMMSTRIDEDBATCHED: StridedBatched version of GEMM +------------- + +As GEMM, but multiple strided operations are batched together for better performance. + +C++ API: +``` +template <typename T> +StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const T beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +``` + +C API: +``` +CLBlastStatusCode CLBlastSgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const float alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const float beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastDgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const double alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const double beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastCgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_float2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_float2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastZgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_double2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_double2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastHgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_half alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_half beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) +``` + +Arguments to GEMMSTRIDEDBATCHED: + +* `const Layout layout`: Data-layout of the matrices, either `Layout::kRowMajor` (101) for row-major layout or `Layout::kColMajor` (102) for column-major data-layout. +* `const Transpose a_transpose`: Transposing the input matrix A, either `Transpose::kNo` (111), `Transpose::kYes` (112), or `Transpose::kConjugate` (113) for a complex-conjugate transpose. +* `const Transpose b_transpose`: Transposing the input matrix B, either `Transpose::kNo` (111), `Transpose::kYes` (112), or `Transpose::kConjugate` (113) for a complex-conjugate transpose. +* `const size_t m`: Integer size argument. This value must be positive. +* `const size_t n`: Integer size argument. This value must be positive. +* `const size_t k`: Integer size argument. This value must be positive. +* `const T alpha`: Input scalar constant. +* `const cl_mem a_buffer`: OpenCL buffer to store the input A matrix. +* `const size_t a_offset`: The offset in elements from the start of the input A matrix. +* `const size_t a_ld`: Leading dimension of the input A matrix. This value must be greater than 0. +* `const size_t a_stride`: The (fixed) stride between two batches of the A matrix. +* `const cl_mem b_buffer`: OpenCL buffer to store the input B matrix. +* `const size_t b_offset`: The offset in elements from the start of the input B matrix. +* `const size_t b_ld`: Leading dimension of the input B matrix. This value must be greater than 0. +* `const size_t b_stride`: The (fixed) stride between two batches of the B matrix. +* `const T beta`: Input scalar constant. +* `cl_mem c_buffer`: OpenCL buffer to store the output C matrix. +* `const size_t c_offset`: The offset in elements from the start of the output C matrix. +* `const size_t c_ld`: Leading dimension of the output C matrix. This value must be greater than 0. +* `const size_t c_stride`: The (fixed) stride between two batches of the C matrix. +* `const size_t batch_count`: Number of batches. This value must be positive. +* `cl_command_queue* queue`: Pointer to an OpenCL command queue associated with a context and device to execute the routine on. +* `cl_event* event`: Pointer to an OpenCL event to be able to wait for completion of the routine's OpenCL kernel(s). This is an optional argument. + +Requirements for GEMMSTRIDEDBATCHED: + +* When `transpose_a == Transpose::kNo`, then `a_ld` must be at least `m`, otherwise `a_ld` must be at least `k`. +* When `transpose_b == Transpose::kNo`, then `b_ld` must be at least `k`, otherwise `b_ld` must be at least `n`. +* The value of `c_ld` must be at least `m`. + + + ClearCache: Resets the cache of compiled binaries (auxiliary function) ------------- diff --git a/include/clblast.h b/include/clblast.h index a05b487f..8e3e64da 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -647,6 +647,18 @@ StatusCode GemmBatched(const Layout layout, const Transpose a_transpose, const T const size_t batch_count, cl_command_queue* queue, cl_event* event = nullptr); +// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED +template <typename T> +StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const T beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event = nullptr); + // ================================================================================================= // Retrieves the required size of the temporary buffer for the GEMM kernel (optional) diff --git a/include/clblast_c.h b/include/clblast_c.h index bd74fe88..f1fc5371 100644 --- a/include/clblast_c.h +++ b/include/clblast_c.h @@ -1451,6 +1451,53 @@ CLBlastStatusCode PUBLIC_API CLBlastHgemmBatched(const CLBlastLayout layout, con const size_t batch_count, cl_command_queue* queue, cl_event* event); +// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED +CLBlastStatusCode PUBLIC_API CLBlastSgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const float alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const float beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event); +CLBlastStatusCode PUBLIC_API CLBlastDgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const double alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const double beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event); +CLBlastStatusCode PUBLIC_API CLBlastCgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_float2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_float2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event); +CLBlastStatusCode PUBLIC_API CLBlastZgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_double2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_double2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event); +CLBlastStatusCode PUBLIC_API CLBlastHgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_half alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_half beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event); + // ================================================================================================= // CLBlast stores binaries of compiled kernels into a cache in case the same kernel is used later on diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h index e1237936..b0cb9aa8 100644 --- a/include/clblast_cuda.h +++ b/include/clblast_cuda.h @@ -619,6 +619,18 @@ StatusCode GemmBatched(const Layout layout, const Transpose a_transpose, const T const size_t batch_count, const CUcontext context, const CUdevice device); +// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED +template <typename T> +StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, + const T alpha, + const CUdeviceptr a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const CUdeviceptr b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const T beta, + CUdeviceptr c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + const CUcontext context, const CUdevice device); + // ================================================================================================= // Retrieves the required size of the temporary buffer for the GEMM kernel (optional) diff --git a/scripts/generator/generator.py b/scripts/generator/generator.py index 5fbce2c4..528e61dd 100755 --- a/scripts/generator/generator.py +++ b/scripts/generator/generator.py @@ -109,71 +109,72 @@ col = "height * width * channels" im2col_constants = ["channels", "height", "width", "kernel_h", "kernel_w", "pad_h", "pad_w", "stride_h", "stride_w", "dilation_h", "dilation_w"] ROUTINES = [ [ # Level 1: vector-vector - Routine(False, True, False, False, "1", "rotg", T, [S,D], [], [], [], ["sa","sb","sc","ss"], ["1","1","1","1"], [], "", "Generate givens plane rotation", "", []), - Routine(False, True, False, False, "1", "rotmg", T, [S,D], [], [], ["sy1"], ["sd1","sd2","sx1","sparam"], ["1","1","1","1","1"], [], "", "Generate modified givens plane rotation", "", []), - Routine(False, True, False, False, "1", "rot", T, [S,D], ["n"], [], [], ["x","y"], [xn,yn], ["cos","sin"],"", "Apply givens plane rotation", "", []), - Routine(False, True, False, False, "1", "rotm", T, [S,D], ["n"], [], [], ["x","y","sparam"], [xn,yn,"1"], [], "", "Apply modified givens plane rotation", "", []), - Routine(True, True, False, False, "1", "swap", T, [S,D,C,Z,H], ["n"], [], [], ["x","y"], [xn,yn], [], "", "Swap two vectors", "Interchanges _n_ elements of vectors _x_ and _y_.", []), - Routine(True, True, False, False, "1", "scal", T, [S,D,C,Z,H], ["n"], [], [], ["x"], [xn], ["alpha"], "", "Vector scaling", "Multiplies _n_ elements of vector _x_ by a scalar constant _alpha_.", []), - Routine(True, True, False, False, "1", "copy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], [], "", "Vector copy", "Copies the contents of vector _x_ into vector _y_.", []), - Routine(True, True, False, False, "1", "axpy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], ["alpha"], "", "Vector-times-constant plus vector", "Performs the operation _y = alpha * x + y_, in which _x_ and _y_ are vectors and _alpha_ is a scalar constant.", []), - Routine(True, True, False, False, "1", "dot", T, [S,D,H], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two vectors", "Multiplies _n_ elements of the vectors _x_ and _y_ element-wise and accumulates the results. The sum is stored in the _dot_ buffer.", []), - Routine(True, True, False, False, "1", "dotu", T, [C,Z], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two complex vectors", "See the regular xDOT routine.", []), - Routine(True, True, False, False, "1", "dotc", T, [C,Z], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two complex vectors, one conjugated", "See the regular xDOT routine.", []), - Routine(True, True, False, False, "1", "nrm2", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["nrm2"], [xn,"1"], [], "2*n", "Euclidian norm of a vector", "Accumulates the square of _n_ elements in the _x_ vector and takes the square root. The resulting L2 norm is stored in the _nrm2_ buffer.", []), - Routine(True, True, False, False, "1", "asum", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["asum"], [xn,"1"], [], "n", "Absolute sum of values in a vector", "Accumulates the absolute value of _n_ elements in the _x_ vector. The results are stored in the _asum_ buffer.", []), - Routine(True, False, False, False, "1", "sum", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["sum"], [xn,"1"], [], "n", "Sum of values in a vector (non-BLAS function)", "Accumulates the values of _n_ elements in the _x_ vector. The results are stored in the _sum_ buffer. This routine is the non-absolute version of the xASUM BLAS routine.", []), - Routine(True, True, False, False, "1", "amax", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imax"], [xn,"1"], [], "2*n", "Index of absolute maximum value in a vector", "Finds the index of the maximum of the absolute values in the _x_ vector. The resulting integer index is stored in the _imax_ buffer.", []), - Routine(True, False, False, False, "1", "amin", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imin"], [xn,"1"], [], "2*n", "Index of absolute minimum value in a vector (non-BLAS function)", "Finds the index of the minimum of the absolute values in the _x_ vector. The resulting integer index is stored in the _imin_ buffer.", []), - Routine(True, False, False, False, "1", "max", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imax"], [xn,"1"], [], "2*n", "Index of maximum value in a vector (non-BLAS function)", "Finds the index of the maximum of the values in the _x_ vector. The resulting integer index is stored in the _imax_ buffer. This routine is the non-absolute version of the IxAMAX BLAS routine.", []), - Routine(True, False, False, False, "1", "min", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imin"], [xn,"1"], [], "2*n", "Index of minimum value in a vector (non-BLAS function)", "Finds the index of the minimum of the values in the _x_ vector. The resulting integer index is stored in the _imin_ buffer. This routine is the non-absolute minimum version of the IxAMAX BLAS routine.", []), + Routine(False, True, 0, False, "1", "rotg", T, [S,D], [], [], [], ["sa","sb","sc","ss"], ["1","1","1","1"], [], "", "Generate givens plane rotation", "", []), + Routine(False, True, 0, False, "1", "rotmg", T, [S,D], [], [], ["sy1"], ["sd1","sd2","sx1","sparam"], ["1","1","1","1","1"], [], "", "Generate modified givens plane rotation", "", []), + Routine(False, True, 0, False, "1", "rot", T, [S,D], ["n"], [], [], ["x","y"], [xn,yn], ["cos","sin"],"", "Apply givens plane rotation", "", []), + Routine(False, True, 0, False, "1", "rotm", T, [S,D], ["n"], [], [], ["x","y","sparam"], [xn,yn,"1"], [], "", "Apply modified givens plane rotation", "", []), + Routine(True, True, 0, False, "1", "swap", T, [S,D,C,Z,H], ["n"], [], [], ["x","y"], [xn,yn], [], "", "Swap two vectors", "Interchanges _n_ elements of vectors _x_ and _y_.", []), + Routine(True, True, 0, False, "1", "scal", T, [S,D,C,Z,H], ["n"], [], [], ["x"], [xn], ["alpha"], "", "Vector scaling", "Multiplies _n_ elements of vector _x_ by a scalar constant _alpha_.", []), + Routine(True, True, 0, False, "1", "copy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], [], "", "Vector copy", "Copies the contents of vector _x_ into vector _y_.", []), + Routine(True, True, 0, False, "1", "axpy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], ["alpha"], "", "Vector-times-constant plus vector", "Performs the operation _y = alpha * x + y_, in which _x_ and _y_ are vectors and _alpha_ is a scalar constant.", []), + Routine(True, True, 0, False, "1", "dot", T, [S,D,H], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two vectors", "Multiplies _n_ elements of the vectors _x_ and _y_ element-wise and accumulates the results. The sum is stored in the _dot_ buffer.", []), + Routine(True, True, 0, False, "1", "dotu", T, [C,Z], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two complex vectors", "See the regular xDOT routine.", []), + Routine(True, True, 0, False, "1", "dotc", T, [C,Z], ["n"], [], ["x","y"], ["dot"], [xn,yn,"1"], [], "n", "Dot product of two complex vectors, one conjugated", "See the regular xDOT routine.", []), + Routine(True, True, 0, False, "1", "nrm2", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["nrm2"], [xn,"1"], [], "2*n", "Euclidian norm of a vector", "Accumulates the square of _n_ elements in the _x_ vector and takes the square root. The resulting L2 norm is stored in the _nrm2_ buffer.", []), + Routine(True, True, 0, False, "1", "asum", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["asum"], [xn,"1"], [], "n", "Absolute sum of values in a vector", "Accumulates the absolute value of _n_ elements in the _x_ vector. The results are stored in the _asum_ buffer.", []), + Routine(True, False, 0, False, "1", "sum", T, [S,D,Sc,Dz,H], ["n"], [], ["x"], ["sum"], [xn,"1"], [], "n", "Sum of values in a vector (non-BLAS function)", "Accumulates the values of _n_ elements in the _x_ vector. The results are stored in the _sum_ buffer. This routine is the non-absolute version of the xASUM BLAS routine.", []), + Routine(True, True, 0, False, "1", "amax", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imax"], [xn,"1"], [], "2*n", "Index of absolute maximum value in a vector", "Finds the index of the maximum of the absolute values in the _x_ vector. The resulting integer index is stored in the _imax_ buffer.", []), + Routine(True, False, 0, False, "1", "amin", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imin"], [xn,"1"], [], "2*n", "Index of absolute minimum value in a vector (non-BLAS function)", "Finds the index of the minimum of the absolute values in the _x_ vector. The resulting integer index is stored in the _imin_ buffer.", []), + Routine(True, False, 0, False, "1", "max", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imax"], [xn,"1"], [], "2*n", "Index of maximum value in a vector (non-BLAS function)", "Finds the index of the maximum of the values in the _x_ vector. The resulting integer index is stored in the _imax_ buffer. This routine is the non-absolute version of the IxAMAX BLAS routine.", []), + Routine(True, False, 0, False, "1", "min", T, [iS,iD,iC,iZ,iH], ["n"], [], ["x"], ["imin"], [xn,"1"], [], "2*n", "Index of minimum value in a vector (non-BLAS function)", "Finds the index of the minimum of the values in the _x_ vector. The resulting integer index is stored in the _imin_ buffer. This routine is the non-absolute minimum version of the IxAMAX BLAS routine.", []), ], [ # Level 2: matrix-vector - Routine(True, True, False, False, "2a", "gemv", T, [S,D,C,Z,H], ["m","n"], ["layout","a_transpose"], ["a","x"], ["y"], [amn,xmn,ynm], ["alpha","beta"], "", "General matrix-vector multiplication", "Performs the operation _y = alpha * A * x + beta * y_, in which _x_ is an input vector, _y_ is an input and output vector, _A_ is an input matrix, and _alpha_ and _beta_ are scalars. The matrix _A_ can optionally be transposed before performing the operation.", [ald_m]), - Routine(True, True, False, False, "2a", "gbmv", T, [S,D,C,Z,H], ["m","n","kl","ku"], ["layout","a_transpose"], ["a","x"], ["y"], [amn,xmn,ynm], ["alpha","beta"], "", "General banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is banded instead.", [ald_kl_ku_one]), - Routine(True, True, False, False, "2a", "hemv", T, [C,Z], ["n"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Hermitian matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian matrix instead.", [ald_n]), - Routine(True, True, False, False, "2a", "hbmv", T, [C,Z], ["n","k"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Hermitian banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian banded matrix instead.", [ald_k_one]), - Routine(True, True, False, False, "2a", "hpmv", T, [C,Z], ["n"], ["layout","triangle"], ["ap","x"], ["y"], [apn,xn,yn], ["alpha","beta"], "", "Hermitian packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), - Routine(True, True, False, False, "2a", "symv", T, [S,D,H], ["n"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Symmetric matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is symmetric instead.", [ald_n]), - Routine(True, True, False, False, "2a", "sbmv", T, [S,D,H], ["n","k"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Symmetric banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is symmetric and banded instead.", [ald_k_one]), - Routine(True, True, False, False, "2a", "spmv", T, [S,D,H], ["n"], ["layout","triangle"], ["ap","x"], ["y"], [apn,xn,yn], ["alpha","beta"], "", "Symmetric packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), - Routine(True, True, False, False, "2a", "trmv", T, [S,D,C,Z,H], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "n", "Triangular matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is triangular instead.", [ald_n]), - Routine(True, True, False, False, "2a", "tbmv", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "n", "Triangular banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is triangular and banded instead.", [ald_k_one]), - Routine(True, True, False, False, "2a", "tpmv", T, [S,D,C,Z,H], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [apn,xn], [], "n", "Triangular packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is a triangular packed matrix instead and repreented as _AP_.", []), - Routine(True, True, False, False, "2a", "trsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "", "Solves a triangular system of equations", "", []), - Routine(False, True, False, False, "2a", "tbsv", T, [S,D,C,Z], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "", "Solves a banded triangular system of equations", "", [ald_k_one]), - Routine(False, True, False, False, "2a", "tpsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [apn,xn], [], "", "Solves a packed triangular system of equations", "", []), + Routine(True, True, 0, False, "2a", "gemv", T, [S,D,C,Z,H], ["m","n"], ["layout","a_transpose"], ["a","x"], ["y"], [amn,xmn,ynm], ["alpha","beta"], "", "General matrix-vector multiplication", "Performs the operation _y = alpha * A * x + beta * y_, in which _x_ is an input vector, _y_ is an input and output vector, _A_ is an input matrix, and _alpha_ and _beta_ are scalars. The matrix _A_ can optionally be transposed before performing the operation.", [ald_m]), + Routine(True, True, 0, False, "2a", "gbmv", T, [S,D,C,Z,H], ["m","n","kl","ku"], ["layout","a_transpose"], ["a","x"], ["y"], [amn,xmn,ynm], ["alpha","beta"], "", "General banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is banded instead.", [ald_kl_ku_one]), + Routine(True, True, 0, False, "2a", "hemv", T, [C,Z], ["n"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Hermitian matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian matrix instead.", [ald_n]), + Routine(True, True, 0, False, "2a", "hbmv", T, [C,Z], ["n","k"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Hermitian banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian banded matrix instead.", [ald_k_one]), + Routine(True, True, 0, False, "2a", "hpmv", T, [C,Z], ["n"], ["layout","triangle"], ["ap","x"], ["y"], [apn,xn,yn], ["alpha","beta"], "", "Hermitian packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2a", "symv", T, [S,D,H], ["n"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Symmetric matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is symmetric instead.", [ald_n]), + Routine(True, True, 0, False, "2a", "sbmv", T, [S,D,H], ["n","k"], ["layout","triangle"], ["a","x"], ["y"], [an,xn,yn], ["alpha","beta"], "", "Symmetric banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is symmetric and banded instead.", [ald_k_one]), + Routine(True, True, 0, False, "2a", "spmv", T, [S,D,H], ["n"], ["layout","triangle"], ["ap","x"], ["y"], [apn,xn,yn], ["alpha","beta"], "", "Symmetric packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2a", "trmv", T, [S,D,C,Z,H], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "n", "Triangular matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is triangular instead.", [ald_n]), + Routine(True, True, 0, False, "2a", "tbmv", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "n", "Triangular banded matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is triangular and banded instead.", [ald_k_one]), + Routine(True, True, 0, False, "2a", "tpmv", T, [S,D,C,Z,H], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [apn,xn], [], "n", "Triangular packed matrix-vector multiplication", "Same operation as xGEMV, but matrix _A_ is a triangular packed matrix instead and repreented as _AP_.", []), + Routine(True, True, 0, False, "2a", "trsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "", "Solves a triangular system of equations", "", []), + Routine(False, True, 0, False, "2a", "tbsv", T, [S,D,C,Z], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [an,xn], [], "", "Solves a banded triangular system of equations", "", [ald_k_one]), + Routine(False, True, 0, False, "2a", "tpsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [apn,xn], [], "", "Solves a packed triangular system of equations", "", []), # Level 2: matrix update - Routine(True, True, False, False, "2b", "ger", T, [S,D,H], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 matrix update", "Performs the operation _A = alpha * x * y^T + A_, in which _x_ is an input vector, _y^T_ is the transpose of the input vector _y_, _A_ is the matrix to be updated, and _alpha_ is a scalar value.", [ald_m]), - Routine(True, True, False, False, "2b", "geru", T, [C,Z], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 complex matrix update", "Same operation as xGER, but with complex data-types.", [ald_m]), - Routine(True, True, False, False, "2b", "gerc", T, [C,Z], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 complex conjugated matrix update", "Same operation as xGERU, but the update is done based on the complex conjugate of the input vectors.", [ald_m]), - Routine(True, True, False, False, "2b", "her", Tc, [Css,Zdd], ["n"], ["layout","triangle"], ["x"], ["a"], [xn,an], ["alpha"], "", "Hermitian rank-1 matrix update", "Performs the operation _A = alpha * x * x^T + A_, in which x is an input vector, x^T is the transpose of this vector, _A_ is the triangular Hermetian matrix to be updated, and alpha is a scalar value.", [ald_n]), - Routine(True, True, False, False, "2b", "hpr", Tc, [Css,Zdd], ["n"], ["layout","triangle"], ["x"], ["ap"], [xn,apn], ["alpha"], "", "Hermitian packed rank-1 matrix update", "Same operation as xHER, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), - Routine(True, True, False, False, "2b", "her2", T, [C,Z], ["n"], ["layout","triangle"], ["x","y"], ["a"], [xn,yn,an], ["alpha"], "", "Hermitian rank-2 matrix update", "Performs the operation _A = alpha * x * y^T + conj(alpha) * y * x^T + A_, in which _x_ is an input vector and _x^T_ its transpose, _y_ is an input vector and _y^T_ its transpose, _A_ is the triangular Hermetian matrix to be updated, _alpha_ is a scalar value and _conj(alpha)_ its complex conjugate.", [ald_n]), - Routine(True, True, False, False, "2b", "hpr2", T, [C,Z], ["n"], ["layout","triangle"], ["x","y"], ["ap"], [xn,yn,apn], ["alpha"], "", "Hermitian packed rank-2 matrix update", "Same operation as xHER2, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), - Routine(True, True, False, False, "2b", "syr", T, [S,D,H], ["n"], ["layout","triangle"], ["x"], ["a"], [xn,an], ["alpha"], "", "Symmetric rank-1 matrix update", "Same operation as xHER, but matrix A is a symmetric matrix instead.", [ald_n]), - Routine(True, True, False, False, "2b", "spr", T, [S,D,H], ["n"], ["layout","triangle"], ["x"], ["ap"], [xn,apn], ["alpha"], "", "Symmetric packed rank-1 matrix update", "Same operation as xSPR, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), - Routine(True, True, False, False, "2b", "syr2", T, [S,D,H], ["n"], ["layout","triangle"], ["x","y"], ["a"], [xn,yn,an], ["alpha"], "", "Symmetric rank-2 matrix update", "Same operation as xHER2, but matrix _A_ is a symmetric matrix instead.", [ald_n]), - Routine(True, True, False, False, "2b", "spr2", T, [S,D,H], ["n"], ["layout","triangle"], ["x","y"], ["ap"], [xn,yn,apn], ["alpha"], "", "Symmetric packed rank-2 matrix update", "Same operation as xSPR2, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2b", "ger", T, [S,D,H], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 matrix update", "Performs the operation _A = alpha * x * y^T + A_, in which _x_ is an input vector, _y^T_ is the transpose of the input vector _y_, _A_ is the matrix to be updated, and _alpha_ is a scalar value.", [ald_m]), + Routine(True, True, 0, False, "2b", "geru", T, [C,Z], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 complex matrix update", "Same operation as xGER, but with complex data-types.", [ald_m]), + Routine(True, True, 0, False, "2b", "gerc", T, [C,Z], ["m","n"], ["layout"], ["x","y"], ["a"], [xm,yn,amn], ["alpha"], "", "General rank-1 complex conjugated matrix update", "Same operation as xGERU, but the update is done based on the complex conjugate of the input vectors.", [ald_m]), + Routine(True, True, 0, False, "2b", "her", Tc, [Css,Zdd], ["n"], ["layout","triangle"], ["x"], ["a"], [xn,an], ["alpha"], "", "Hermitian rank-1 matrix update", "Performs the operation _A = alpha * x * x^T + A_, in which x is an input vector, x^T is the transpose of this vector, _A_ is the triangular Hermetian matrix to be updated, and alpha is a scalar value.", [ald_n]), + Routine(True, True, 0, False, "2b", "hpr", Tc, [Css,Zdd], ["n"], ["layout","triangle"], ["x"], ["ap"], [xn,apn], ["alpha"], "", "Hermitian packed rank-1 matrix update", "Same operation as xHER, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2b", "her2", T, [C,Z], ["n"], ["layout","triangle"], ["x","y"], ["a"], [xn,yn,an], ["alpha"], "", "Hermitian rank-2 matrix update", "Performs the operation _A = alpha * x * y^T + conj(alpha) * y * x^T + A_, in which _x_ is an input vector and _x^T_ its transpose, _y_ is an input vector and _y^T_ its transpose, _A_ is the triangular Hermetian matrix to be updated, _alpha_ is a scalar value and _conj(alpha)_ its complex conjugate.", [ald_n]), + Routine(True, True, 0, False, "2b", "hpr2", T, [C,Z], ["n"], ["layout","triangle"], ["x","y"], ["ap"], [xn,yn,apn], ["alpha"], "", "Hermitian packed rank-2 matrix update", "Same operation as xHER2, but matrix _A_ is an Hermitian packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2b", "syr", T, [S,D,H], ["n"], ["layout","triangle"], ["x"], ["a"], [xn,an], ["alpha"], "", "Symmetric rank-1 matrix update", "Same operation as xHER, but matrix A is a symmetric matrix instead.", [ald_n]), + Routine(True, True, 0, False, "2b", "spr", T, [S,D,H], ["n"], ["layout","triangle"], ["x"], ["ap"], [xn,apn], ["alpha"], "", "Symmetric packed rank-1 matrix update", "Same operation as xSPR, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), + Routine(True, True, 0, False, "2b", "syr2", T, [S,D,H], ["n"], ["layout","triangle"], ["x","y"], ["a"], [xn,yn,an], ["alpha"], "", "Symmetric rank-2 matrix update", "Same operation as xHER2, but matrix _A_ is a symmetric matrix instead.", [ald_n]), + Routine(True, True, 0, False, "2b", "spr2", T, [S,D,H], ["n"], ["layout","triangle"], ["x","y"], ["ap"], [xn,yn,apn], ["alpha"], "", "Symmetric packed rank-2 matrix update", "Same operation as xSPR2, but matrix _A_ is a symmetric packed matrix instead and represented as _AP_.", []), ], [ # Level 3: matrix-matrix - Routine(True, True, False, True, "3", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "General matrix-matrix multiplication", "Performs the matrix product _C = alpha * A * B + beta * C_, in which _A_ (_m_ by _k_) and _B_ (_k_ by _n_) are two general rectangular input matrices, _C_ (_m_ by _n_) is the matrix to be updated, and _alpha_ and _beta_ are scalar values. The matrices _A_ and/or _B_ can optionally be transposed before performing the operation.", [ald_transa_m_k, bld_transb_k_n, cld_m]), - Routine(True, True, False, False, "3", "symm", T, [S,D,C,Z,H], ["m","n"], ["layout","side","triangle"], ["a","b"], ["c"], [ammn,bmnn,cmn], ["alpha","beta"], "", "Symmetric matrix-matrix multiplication", "Same operation as xGEMM, but _A_ is symmetric instead. In case of `side == kLeft`, _A_ is a symmetric _m_ by _m_ matrix and _C = alpha * A * B + beta * C_ is performed. Otherwise, in case of `side == kRight`, _A_ is a symmtric _n_ by _n_ matrix and _C = alpha * B * A + beta * C_ is performed.", [ald_side_m_n, bld_m, cld_m]), - Routine(True, True, False, False, "3", "hemm", T, [C,Z], ["m","n"], ["layout","side","triangle"], ["a","b"], ["c"], [ammn,bmnn,cmn], ["alpha","beta"], "", "Hermitian matrix-matrix multiplication", "Same operation as xSYMM, but _A_ is an Hermitian matrix instead.", [ald_side_m_n, bld_m, cld_m]), - Routine(True, True, False, False, "3", "syrk", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","a_transpose"], ["a"], ["c"], [ank,cn], ["alpha","beta"], "", "Rank-K update of a symmetric matrix", "Performs the matrix product _C = alpha * A * A^T + beta * C_ or _C = alpha * A^T * A + beta * C_, in which _A_ is a general matrix and _A^T_ is its transpose, _C_ (_n_ by _n_) is the symmetric matrix to be updated, and _alpha_ and _beta_ are scalar values.", [ald_trans_n_k, cld_m]), - Routine(True, True, False, False, "3", "herk", Tc, [Css,Zdd], ["n","k"], ["layout","triangle","a_transpose"], ["a"], ["c"], [ank,cn], ["alpha","beta"], "", "Rank-K update of a hermitian matrix", "Same operation as xSYRK, but _C_ is an Hermitian matrix instead.", [ald_trans_n_k, cld_m]), - Routine(True, True, False, False, "3", "syr2k", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","ab_transpose"], ["a","b"], ["c"], [ankab,bnkab,cn],["alpha","beta"], "", "Rank-2K update of a symmetric matrix", "Performs the matrix product _C = alpha * A * B^T + alpha * B * A^T + beta * C_ or _C = alpha * A^T * B + alpha * B^T * A + beta * C_, in which _A_ and _B_ are general matrices and _A^T_ and _B^T_ are their transposed versions, _C_ (_n_ by _n_) is the symmetric matrix to be updated, and _alpha_ and _beta_ are scalar values.", [ald_trans_n_k, bld_trans_n_k, cld_n]), - Routine(True, True, False, False, "3", "her2k", TU, [Ccs,Zzd], ["n","k"], ["layout","triangle","ab_transpose"], ["a","b"], ["c"], [ankab,bnkab,cn],["alpha","beta"], "", "Rank-2K update of a hermitian matrix", "Same operation as xSYR2K, but _C_ is an Hermitian matrix instead.", [ald_trans_n_k, bld_trans_n_k, cld_n]), - Routine(True, True, False, False, "3", "trmm", T, [S,D,C,Z,H], ["m","n"], ["layout","side","triangle","a_transpose","diagonal"], ["a"], ["b"], [amns,bmn], ["alpha"], "", "Triangular matrix-matrix multiplication", "Performs the matrix product _B = alpha * A * B_ or _B = alpha * B * A_, in which _A_ is a unit or non-unit triangular matrix, _B_ (_m_ by _n_) is the general matrix to be updated, and _alpha_ is a scalar value.", [ald_side_m_n, bld_m]), - Routine(True, True, False, False, "3", "trsm", T, [S,D,C,Z], ["m","n"], ["layout","side","triangle","a_transpose","diagonal"], ["a"], ["b"], [amns,bmn], ["alpha"], "", "Solves a triangular system of equations", "Solves the equation _A * X = alpha * B_ for the unknown _m_ by _n_ matrix X, in which _A_ is an _n_ by _n_ unit or non-unit triangular matrix and B is an _m_ by _n_ matrix. The matrix _B_ is overwritten by the solution _X_.", []), + Routine(True, True, 0, True, "3", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "General matrix-matrix multiplication", "Performs the matrix product _C = alpha * A * B + beta * C_, in which _A_ (_m_ by _k_) and _B_ (_k_ by _n_) are two general rectangular input matrices, _C_ (_m_ by _n_) is the matrix to be updated, and _alpha_ and _beta_ are scalar values. The matrices _A_ and/or _B_ can optionally be transposed before performing the operation.", [ald_transa_m_k, bld_transb_k_n, cld_m]), + Routine(True, True, 0, False, "3", "symm", T, [S,D,C,Z,H], ["m","n"], ["layout","side","triangle"], ["a","b"], ["c"], [ammn,bmnn,cmn], ["alpha","beta"], "", "Symmetric matrix-matrix multiplication", "Same operation as xGEMM, but _A_ is symmetric instead. In case of `side == kLeft`, _A_ is a symmetric _m_ by _m_ matrix and _C = alpha * A * B + beta * C_ is performed. Otherwise, in case of `side == kRight`, _A_ is a symmtric _n_ by _n_ matrix and _C = alpha * B * A + beta * C_ is performed.", [ald_side_m_n, bld_m, cld_m]), + Routine(True, True, 0, False, "3", "hemm", T, [C,Z], ["m","n"], ["layout","side","triangle"], ["a","b"], ["c"], [ammn,bmnn,cmn], ["alpha","beta"], "", "Hermitian matrix-matrix multiplication", "Same operation as xSYMM, but _A_ is an Hermitian matrix instead.", [ald_side_m_n, bld_m, cld_m]), + Routine(True, True, 0, False, "3", "syrk", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","a_transpose"], ["a"], ["c"], [ank,cn], ["alpha","beta"], "", "Rank-K update of a symmetric matrix", "Performs the matrix product _C = alpha * A * A^T + beta * C_ or _C = alpha * A^T * A + beta * C_, in which _A_ is a general matrix and _A^T_ is its transpose, _C_ (_n_ by _n_) is the symmetric matrix to be updated, and _alpha_ and _beta_ are scalar values.", [ald_trans_n_k, cld_m]), + Routine(True, True, 0, False, "3", "herk", Tc, [Css,Zdd], ["n","k"], ["layout","triangle","a_transpose"], ["a"], ["c"], [ank,cn], ["alpha","beta"], "", "Rank-K update of a hermitian matrix", "Same operation as xSYRK, but _C_ is an Hermitian matrix instead.", [ald_trans_n_k, cld_m]), + Routine(True, True, 0, False, "3", "syr2k", T, [S,D,C,Z,H], ["n","k"], ["layout","triangle","ab_transpose"], ["a","b"], ["c"], [ankab,bnkab,cn],["alpha","beta"], "", "Rank-2K update of a symmetric matrix", "Performs the matrix product _C = alpha * A * B^T + alpha * B * A^T + beta * C_ or _C = alpha * A^T * B + alpha * B^T * A + beta * C_, in which _A_ and _B_ are general matrices and _A^T_ and _B^T_ are their transposed versions, _C_ (_n_ by _n_) is the symmetric matrix to be updated, and _alpha_ and _beta_ are scalar values.", [ald_trans_n_k, bld_trans_n_k, cld_n]), + Routine(True, True, 0, False, "3", "her2k", TU, [Ccs,Zzd], ["n","k"], ["layout","triangle","ab_transpose"], ["a","b"], ["c"], [ankab,bnkab,cn],["alpha","beta"], "", "Rank-2K update of a hermitian matrix", "Same operation as xSYR2K, but _C_ is an Hermitian matrix instead.", [ald_trans_n_k, bld_trans_n_k, cld_n]), + Routine(True, True, 0, False, "3", "trmm", T, [S,D,C,Z,H], ["m","n"], ["layout","side","triangle","a_transpose","diagonal"], ["a"], ["b"], [amns,bmn], ["alpha"], "", "Triangular matrix-matrix multiplication", "Performs the matrix product _B = alpha * A * B_ or _B = alpha * B * A_, in which _A_ is a unit or non-unit triangular matrix, _B_ (_m_ by _n_) is the general matrix to be updated, and _alpha_ is a scalar value.", [ald_side_m_n, bld_m]), + Routine(True, True, 0, False, "3", "trsm", T, [S,D,C,Z], ["m","n"], ["layout","side","triangle","a_transpose","diagonal"], ["a"], ["b"], [amns,bmn], ["alpha"], "", "Solves a triangular system of equations", "Solves the equation _A * X = alpha * B_ for the unknown _m_ by _n_ matrix X, in which _A_ is an _n_ by _n_ unit or non-unit triangular matrix and B is an _m_ by _n_ matrix. The matrix _B_ is overwritten by the solution _X_.", []), ], [ # Level X: extra routines (not part of BLAS) # Special routines: - Routine(True, True, False, False, "x", "omatcopy", T, [S,D,C,Z,H], ["m","n"], ["layout","a_transpose"], ["a"], ["b"], [amn,bnma], ["alpha"], "", "Scaling and out-place transpose/copy (non-BLAS function)", "Performs scaling and out-of-place transposition/copying of matrices according to _B = alpha*op(A)_, in which _A_ is an input matrix (_m_ rows by _n_ columns), _B_ an output matrix, and _alpha_ a scalar value. The operation _op_ can be a normal matrix copy, a transposition or a conjugate transposition.", [ald_m, bld_n]), - Routine(True, True, False, False, "x", "im2col", T, [S,D,C,Z,H], im2col_constants, [], ["im"], ["col"], [im,col], [""], "", "Im2col function (non-BLAS function)", "Performs the im2col algorithm, in which _im_ is the input matrix and _col_ is the output matrix.", []), + Routine(True, True, 0, False, "x", "omatcopy", T, [S,D,C,Z,H], ["m","n"], ["layout","a_transpose"], ["a"], ["b"], [amn,bnma], ["alpha"], "", "Scaling and out-place transpose/copy (non-BLAS function)", "Performs scaling and out-of-place transposition/copying of matrices according to _B = alpha*op(A)_, in which _A_ is an input matrix (_m_ rows by _n_ columns), _B_ an output matrix, and _alpha_ a scalar value. The operation _op_ can be a normal matrix copy, a transposition or a conjugate transposition.", [ald_m, bld_n]), + Routine(True, True, 0, False, "x", "im2col", T, [S,D,C,Z,H], im2col_constants, [], ["im"], ["col"], [im,col], [""], "", "Im2col function (non-BLAS function)", "Performs the im2col algorithm, in which _im_ is the input matrix and _col_ is the output matrix.", []), # Batched routines: - Routine(True, True, True, False, "x", "axpy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], ["alpha"], "", "Batched version of AXPY", "As AXPY, but multiple operations are batched together for better performance.", []), - Routine(True, True, True, False, "x", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "Batched version of GEMM", "As GEMM, but multiple operations are batched together for better performance.", [ald_transa_m_k, bld_transb_k_n, cld_m]), + Routine(True, True, 1, False, "x", "axpy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], ["alpha"], "", "Batched version of AXPY", "As AXPY, but multiple operations are batched together for better performance.", []), + Routine(True, True, 1, False, "x", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "Batched version of GEMM", "As GEMM, but multiple operations are batched together for better performance.", [ald_transa_m_k, bld_transb_k_n, cld_m]), + Routine(True, True, 2, False, "x", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "StridedBatched version of GEMM", "As GEMM, but multiple strided operations are batched together for better performance.", [ald_transa_m_k, bld_transb_k_n, cld_m]), ]] @@ -223,10 +224,10 @@ def main(argv): if i == 6: body += cpp.wrapper_cublas(routine) if i == 7: - if not routine.batched: + if routine.batched == 0: body += cpp.clblast_netlib_c_h(routine) if i == 8: - if not routine.batched: + if routine.batched == 0: body += cpp.clblast_netlib_c_cc(routine) if i == 9: body += cpp.clblast_h(routine, cuda=True) diff --git a/scripts/generator/generator/cpp.py b/scripts/generator/generator/cpp.py index 656253d7..51ca047c 100644 --- a/scripts/generator/generator/cpp.py +++ b/scripts/generator/generator/cpp.py @@ -58,11 +58,12 @@ def clblast_cc(routine, cuda=False): result += " auto queue_cpp = Queue(*queue);" + NL event = "nullptr" if cuda else "event" result += " auto routine = X" + routine.plain_name() + "<" + routine.template.template + ">(queue_cpp, " + event + ");" + NL - if routine.batched: + if routine.batched == 1: result += " " + (NL + " ").join(routine.batched_transform_to_cpp()) + NL if routine.temp_buffer: - result += " const auto temp_buffer_provided = temp_buffer != nullptr;\n" - result += " auto temp_buffer_cpp = temp_buffer_provided ? Buffer<T>(temp_buffer) : Buffer<T>(nullptr);\n" + null = "0" if cuda else "nullptr" + result += " const auto temp_buffer_provided = temp_buffer != " + null + ";\n" + result += " auto temp_buffer_cpp = temp_buffer_provided ? Buffer<T>(temp_buffer) : Buffer<T>(" + null + ");\n" result += " routine.Do" + routine.capitalized_name() + "(" result += ("," + NL + indent1).join([a for a in routine.arguments_clcudaapi()]) if routine.temp_buffer: @@ -109,7 +110,7 @@ def clblast_c_cc(routine): template = "<" + flavour.template + ">" if routine.no_scalars() else "" indent = " " * (16 + routine.length() + len(template)) result += routine.routine_header_c(flavour, 27, "") + " {" + NL - if routine.batched: + if routine.batched == 1: result += " " + (NL + " ").join(routine.batched_transform_to_complex(flavour)) + NL result += " try {" + NL result += " return static_cast<CLBlastStatusCode>(" + NL @@ -387,7 +388,7 @@ def performance_test(routine, level_string): found = False for flavour in routine.flavours: if flavour.precision_name == precision: - extra_template_argument = "0, " if routine.name == "gemm" and not routine.batched else "" + extra_template_argument = "0, " if routine.name == "gemm" and routine.batched == 0 else "" result += NL + " clblast::RunClient<clblast::TestX" + routine.plain_name() result += flavour.test_template(extra_template_argument) result += ">(argc, argv); break;" + NL @@ -409,7 +410,7 @@ def correctness_test(routine, level_string): result += "int main(int argc, char *argv[]) {" + NL result += " auto errors = size_t{0};" + NL not_first = "false" - extra_template_arguments = ["1, ", "2, "] if routine.name == "gemm" and not routine.batched else [""] + extra_template_arguments = ["1, ", "2, "] if routine.name == "gemm" and routine.batched == 0 else [""] for extra_template_argument in extra_template_arguments: for flavour in routine.flavours: result += " errors += clblast::RunTests<clblast::TestX" + routine.plain_name() diff --git a/scripts/generator/generator/routine.py b/scripts/generator/generator/routine.py index 22be02b0..f7c2a701 100644 --- a/scripts/generator/generator/routine.py +++ b/scripts/generator/generator/routine.py @@ -12,12 +12,12 @@ import generator.convert as convert class Routine: """Class holding routine-specific information (e.g. name, which arguments, which precisions)""" - def __init__(self, implemented, has_tests, batched, temp_buffer, level, name, template, flavours, sizes, options, + def __init__(self, implemented, has_tests, batched_strided, temp_buffer, level, name, template, flavours, sizes, options, inputs, outputs, buffer_sizes, scalars, scratch, description, details, requirements): self.implemented = implemented self.has_tests = has_tests - self.batched = batched + self.batched = batched_strided self.temp_buffer = temp_buffer self.level = level self.name = name @@ -35,38 +35,42 @@ class Routine: self.requirements = requirements def lowercase_name(self): - postfix = "batched" if self.batched else "" + postfix = "strided" if self.batched == 2 else "" + postfix += "batched" if self.batched != 0 else "" return self.name + postfix def plain_name(self): - postfix = "Batched" if self.batched else "" + postfix = "Strided" if self.batched == 2 else "" + postfix += "Batched" if self.batched != 0 else "" return self.name + postfix def capitalized_name(self): - postfix = "Batched" if self.batched else "" + postfix = "Strided" if self.batched == 2 else "" + postfix += "Batched" if self.batched != 0 else "" return self.name.capitalize() + postfix def upper_name(self): - postfix = "BATCHED" if self.batched else "" + postfix = "STRIDED" if self.batched == 2 else "" + postfix += "BATCHED" if self.batched != 0 else "" return self.name.upper() + postfix def b_star(self): - return "*" if self.batched else "" + return "*" if self.batched == 1 else "" def b_s(self): - return "s" if self.batched else "" + return "s" if self.batched == 1 else "" def batch_count_def(self): - return ["const size_t batch_count"] if self.batched else [] + return ["const size_t batch_count"] if self.batched != 0 else [] def batch_count_list(self): - return ["batch_count"] if self.batched else [] + return ["batch_count"] if self.batched != 0 else [] def batch_count_type(self): - return ["const size_t"] if self.batched else [] + return ["const size_t"] if self.batched != 0 else [] def batch_count_doc(self): - return ["`const size_t batch_count`: Number of batches. This value must be positive."] if self.batched else [] + return ["`const size_t batch_count`: Number of batches. This value must be positive."] if self.batched != 0 else [] def batched_transform_to_cpp(self): result = [] @@ -230,6 +234,8 @@ class Routine: a = [name + "_buffer"] b = [name + "_offset" + self.b_s()] c = [name + "_" + self.postfix(name)] if (name not in self.buffers_without_ld_inc()) else [] + if self.batched == 2: + c += [name + "_stride"] return [", ".join(a + b + c)] return [] @@ -239,6 +245,8 @@ class Routine: a = [name + "_buffer_bis"] b = [name + "_offset"] c = [name + "_" + self.postfix(name)] if name not in self.buffers_without_ld_inc() else [] + if self.batched == 2: + c += [name + "_stride"] return [", ".join(a + b + c)] return [] @@ -258,6 +266,8 @@ class Routine: a = [prefix + "cl_mem " + name + "_buffer"] b = ["const size_t " + self.b_star() + name + "_offset" + self.b_s()] c = ["const size_t " + name + "_" + self.postfix(name)] if name not in self.buffers_without_ld_inc() else [] + if self.batched == 2: + c += ["const size_t " + name + "_stride"] return [", ".join(a + b + c)] return [] @@ -307,8 +317,10 @@ class Routine: if name in self.inputs or name in self.outputs: buffer_type = "unsigned int" if (name in self.index_buffers()) else self.template.buffer_type a = ["Buffer<" + buffer_type + ">(" + name + "_buffer)"] - b = [name + "_offsets_cpp"] if self.batched else [name + "_offset"] + b = [name + "_offsets_cpp"] if self.batched == 1 else [name + "_offset"] c = [name + "_" + self.postfix(name)] if (name not in self.buffers_without_ld_inc()) else [] + if self.batched == 2: + c += [name + "_stride"] return [", ".join(a + b + c)] return [] @@ -375,6 +387,8 @@ class Routine: a = [prefix + "cl_mem"] b = ["const size_t" + self.b_star()] c = ["const size_t"] if (name not in self.buffers_without_ld_inc()) else [] + if self.batched == 2: + c += ["const size_t"] return [", ".join(a + b + c)] return [] @@ -391,13 +405,15 @@ class Routine: if name not in self.buffers_without_ld_inc(): c = ["`const size_t " + name + "_" + self.postfix(name) + "`: " + inc_ld_description + "of the " + inout + " " + math_name + ". This value must be greater than 0."] + if self.batched == 2: + c += ["`const size_t " + name + "_stride`: The (fixed) stride between two batches of the " + name.upper() + " matrix."] return a + b + c return [] def scalar(self, name): """Retrieves the name of a scalar (alpha/beta)""" if name in self.scalars: - if self.batched: + if self.batched == 1: return [name + "s_cpp"] return [name] return [] @@ -418,11 +434,11 @@ class Routine: """Retrieves the use of a scalar (alpha/beta)""" if name in self.scalars: if name == "alpha": - if self.batched: + if self.batched == 1: return ["alphas_cpp.data()"] return [flavour.use_alpha()] elif name == "beta": - if self.batched: + if self.batched == 1: return ["betas_cpp.data()"] return [flavour.use_beta()] return [name] @@ -822,7 +838,7 @@ class Routine: if self.temp_buffer: result += ",\n" + indent + mem_type + " temp_buffer" if not implementation: - result += " = nullptr" + result += " = 0" if cuda else " = nullptr" result += ")" return result @@ -866,7 +882,7 @@ class Routine: if self.name in self.routines_scalar_no_return(): routine_name += "_sub" indent += " " - if self.batched: + if self.batched != 0: routine_name += "batched" result = return_type + extra_qualifier + " cblas_" + flavour.name.lower() + routine_name + "(" result += (",\n" + indent).join([a for a in self.arguments_def_netlib(flavour)]) + ")" diff --git a/src/clblast.cpp b/src/clblast.cpp index f5e2f1be..c4c51538 100644 --- a/src/clblast.cpp +++ b/src/clblast.cpp @@ -2336,6 +2336,77 @@ template StatusCode PUBLIC_API GemmBatched<half>(const Layout, const Transpose, const size_t, cl_command_queue*, cl_event*); +// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED +template <typename T> +StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const T beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + auto queue_cpp = Queue(*queue); + auto routine = XgemmStridedBatched<T>(queue_cpp, event); + routine.DoGemmStridedBatched(layout, a_transpose, b_transpose, + m, n, k, + alpha, + Buffer<T>(a_buffer), a_offset, a_ld, a_stride, + Buffer<T>(b_buffer), b_offset, b_ld, b_stride, + beta, + Buffer<T>(c_buffer), c_offset, c_ld, c_stride, + batch_count); + return StatusCode::kSuccess; + } catch (...) { return DispatchException(); } +} +template StatusCode PUBLIC_API GemmStridedBatched<float>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const float, + const cl_mem, const size_t, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const size_t, + const float, + cl_mem, const size_t, const size_t, const size_t, + const size_t, + cl_command_queue*, cl_event*); +template StatusCode PUBLIC_API GemmStridedBatched<double>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const double, + const cl_mem, const size_t, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const size_t, + const double, + cl_mem, const size_t, const size_t, const size_t, + const size_t, + cl_command_queue*, cl_event*); +template StatusCode PUBLIC_API GemmStridedBatched<float2>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const float2, + const cl_mem, const size_t, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const size_t, + const float2, + cl_mem, const size_t, const size_t, const size_t, + const size_t, + cl_command_queue*, cl_event*); +template StatusCode PUBLIC_API GemmStridedBatched<double2>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const double2, + const cl_mem, const size_t, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const size_t, + const double2, + cl_mem, const size_t, const size_t, const size_t, + const size_t, + cl_command_queue*, cl_event*); +template StatusCode PUBLIC_API GemmStridedBatched<half>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const half, + const cl_mem, const size_t, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const size_t, + const half, + cl_mem, const size_t, const size_t, const size_t, + const size_t, + cl_command_queue*, cl_event*); + // ================================================================================================= // Retrieves the required size of the temporary buffer for the GEMM kernel (optional) diff --git a/src/clblast_c.cpp b/src/clblast_c.cpp index 24697779..aa52cbca 100644 --- a/src/clblast_c.cpp +++ b/src/clblast_c.cpp @@ -3846,6 +3846,133 @@ CLBlastStatusCode CLBlastHgemmBatched(const CLBlastLayout layout, const CLBlastT } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } } +// GEMM +CLBlastStatusCode CLBlastSgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const float alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const float beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + return static_cast<CLBlastStatusCode>( + clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout), + static_cast<clblast::Transpose>(a_transpose), + static_cast<clblast::Transpose>(b_transpose), + m, n, k, + alpha, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, + beta, + c_buffer, c_offset, c_ld, c_stride, + batch_count, + queue, event) + ); + } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } +} +CLBlastStatusCode CLBlastDgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const double alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const double beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + return static_cast<CLBlastStatusCode>( + clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout), + static_cast<clblast::Transpose>(a_transpose), + static_cast<clblast::Transpose>(b_transpose), + m, n, k, + alpha, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, + beta, + c_buffer, c_offset, c_ld, c_stride, + batch_count, + queue, event) + ); + } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } +} +CLBlastStatusCode CLBlastCgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_float2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_float2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + return static_cast<CLBlastStatusCode>( + clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout), + static_cast<clblast::Transpose>(a_transpose), + static_cast<clblast::Transpose>(b_transpose), + m, n, k, + float2{alpha.s[0], alpha.s[1]}, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, + float2{beta.s[0], beta.s[1]}, + c_buffer, c_offset, c_ld, c_stride, + batch_count, + queue, event) + ); + } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } +} +CLBlastStatusCode CLBlastZgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_double2 alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_double2 beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + return static_cast<CLBlastStatusCode>( + clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout), + static_cast<clblast::Transpose>(a_transpose), + static_cast<clblast::Transpose>(b_transpose), + m, n, k, + double2{alpha.s[0], alpha.s[1]}, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, + double2{beta.s[0], beta.s[1]}, + c_buffer, c_offset, c_ld, c_stride, + batch_count, + queue, event) + ); + } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } +} +CLBlastStatusCode CLBlastHgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose, + const size_t m, const size_t n, const size_t k, + const cl_half alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const cl_half beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + cl_command_queue* queue, cl_event* event) { + try { + return static_cast<CLBlastStatusCode>( + clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout), + static_cast<clblast::Transpose>(a_transpose), + static_cast<clblast::Transpose>(b_transpose), + m, n, k, + alpha, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, + beta, + c_buffer, c_offset, c_ld, c_stride, + batch_count, + queue, event) + ); + } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); } +} + // ================================================================================================= // Clears the cache of stored binaries diff --git a/src/clblast_cuda.cpp b/src/clblast_cuda.cpp index 348ff3f5..0aa57087 100644 --- a/src/clblast_cuda.cpp +++ b/src/clblast_cuda.cpp @@ -2436,6 +2436,79 @@ template StatusCode PUBLIC_API GemmBatched<half>(const Layout, const Transpose, const size_t, const CUcontext, const CUdevice); +// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED +template <typename T> +StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, + const T alpha, + const CUdeviceptr a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const CUdeviceptr b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, + const T beta, + CUdeviceptr c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count, + const CUcontext context, const CUdevice device) { + try { + const auto context_cpp = Context(context); + const auto device_cpp = Device(device); + auto queue_cpp = Queue(context_cpp, device_cpp); + auto routine = XgemmStridedBatched<T>(queue_cpp, nullptr); + routine.DoGemmStridedBatched(layout, a_transpose, b_transpose, + m, n, k, + alpha, + Buffer<T>(a_buffer), a_offset, a_ld, a_stride, + Buffer<T>(b_buffer), b_offset, b_ld, b_stride, + beta, + Buffer<T>(c_buffer), c_offset, c_ld, c_stride, + batch_count); + return StatusCode::kSuccess; + } catch (...) { return DispatchException(); } +} +template StatusCode PUBLIC_API GemmStridedBatched<float>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const float, + const CUdeviceptr, const size_t, const size_t, const size_t, + const CUdeviceptr, const size_t, const size_t, const size_t, + const float, + CUdeviceptr, const size_t, const size_t, const size_t, + const size_t, + const CUcontext, const CUdevice); +template StatusCode PUBLIC_API GemmStridedBatched<double>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const double, + const CUdeviceptr, const size_t, const size_t, const size_t, + const CUdeviceptr, const size_t, const size_t, const size_t, + const double, + CUdeviceptr, const size_t, const size_t, const size_t, + const size_t, + const CUcontext, const CUdevice); +template StatusCode PUBLIC_API GemmStridedBatched<float2>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const float2, + const CUdeviceptr, const size_t, const size_t, const size_t, + const CUdeviceptr, const size_t, const size_t, const size_t, + const float2, + CUdeviceptr, const size_t, const size_t, const size_t, + const size_t, + const CUcontext, const CUdevice); +template StatusCode PUBLIC_API GemmStridedBatched<double2>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const double2, + const CUdeviceptr, const size_t, const size_t, const size_t, + const CUdeviceptr, const size_t, const size_t, const size_t, + const double2, + CUdeviceptr, const size_t, const size_t, const size_t, + const size_t, + const CUcontext, const CUdevice); +template StatusCode PUBLIC_API GemmStridedBatched<half>(const Layout, const Transpose, const Transpose, + const size_t, const size_t, const size_t, + const half, + const CUdeviceptr, const size_t, const size_t, const size_t, + const CUdeviceptr, const size_t, const size_t, const size_t, + const half, + CUdeviceptr, const size_t, const size_t, const size_t, + const size_t, + const CUcontext, const CUdevice); + // ================================================================================================= // Retrieves the required size of the temporary buffer for the GEMM kernel (optional) diff --git a/src/kernels/level3/copy_pad.opencl b/src/kernels/level3/copy_pad.opencl index 2e191514..3d389b74 100644 --- a/src/kernels/level3/copy_pad.opencl +++ b/src/kernels/level3/copy_pad.opencl @@ -174,6 +174,45 @@ void CopyMatrixBatched(const int src_one, const int src_two, #endif // ================================================================================================= +#if defined(ROUTINE_GEMMSTRIDEDBATCHED) + +// Strided-batched version of the above +__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +void CopyPadMatrixStridedBatched(const int src_one, const int src_two, + const int src_ld, const int src_offset, + const int src_stride, __global const real* restrict src, + const int dest_one, const int dest_two, + const int dest_ld, const int dest_offset, + const int dest_stride, __global real* dest, + const int do_conjugate) { + const int batch = get_group_id(2); + const int src_offset_batch = src_offset + src_stride * batch; + const int dest_offset_batch = dest_offset + dest_stride * batch; + real alpha; SetToOne(alpha); + _CopyPadMatrix(src_one, src_two, src_ld, src_offset_batch, src, + dest_one, dest_two, dest_ld, dest_offset_batch, dest, + alpha, do_conjugate); +} + +// Strided-batched version of the above +__kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +void CopyMatrixStridedBatched(const int src_one, const int src_two, + const int src_ld, const int src_offset, + const int src_stride, __global const real* restrict src, + const int dest_one, const int dest_two, + const int dest_ld, const int dest_offset, + const int dest_stride, __global real* dest) { + const int batch = get_group_id(2); + const int src_offset_batch = src_offset + src_stride * batch; + const int dest_offset_batch = dest_offset + dest_stride * batch; + real alpha; SetToOne(alpha); + _CopyMatrix(src_one, src_two, src_ld, src_offset_batch, src, + dest_one, dest_two, dest_ld, dest_offset_batch, dest, + alpha, 0, 0, 0); +} + +#endif +// ================================================================================================= // End of the C++11 raw string literal )" diff --git a/src/kernels/level3/transpose_pad.opencl b/src/kernels/level3/transpose_pad.opencl index 67c2bf72..e55a8b7c 100644 --- a/src/kernels/level3/transpose_pad.opencl +++ b/src/kernels/level3/transpose_pad.opencl @@ -231,6 +231,47 @@ void TransposeMatrixBatched(const int src_one, const int src_two, #endif // ================================================================================================= +#if defined(ROUTINE_GEMMSTRIDEDBATCHED) + +// Strided-batched version of the above +__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +void TransposePadMatrixStridedBatched(const int src_one, const int src_two, + const int src_ld, const int src_offset, + const int src_stride, __global const real* restrict src, + const int dest_one, const int dest_two, + const int dest_ld, const int dest_offset, + const int dest_stride, __global real* dest, + const int do_conjugate) { + const int batch = get_group_id(2); + const int src_offset_batch = src_offset + src_stride * batch; + const int dest_offset_batch = dest_offset + dest_stride * batch; + real alpha; SetToOne(alpha); + __local real tile[(PADTRA_WPT*PADTRA_TILE) * (PADTRA_WPT*PADTRA_TILE + PADTRA_PAD)]; + _TransposePadMatrix(tile, src_one, src_two, src_ld, src_offset_batch, src, + dest_one, dest_two, dest_ld, dest_offset_batch, dest, + alpha, do_conjugate); +} + +// Strided-batched version of the above +__kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1))) +void TransposeMatrixStridedBatched(const int src_one, const int src_two, + const int src_ld, const int src_offset, + const int src_stride, __global const real* restrict src, + const int dest_one, const int dest_two, + const int dest_ld, const int dest_offset, + const int dest_stride, __global real* dest) { + const int batch = get_group_id(2); + const int src_offset_batch = src_offset + src_stride * batch; + const int dest_offset_batch = dest_offset + dest_stride * batch; + real alpha; SetToOne(alpha); + __local real tile[(PADTRA_WPT*PADTRA_TILE) * (PADTRA_WPT*PADTRA_TILE + PADTRA_PAD)]; + _TransposeMatrix(tile, src_one, src_two, src_ld, src_offset_batch, src, + dest_one, dest_two, dest_ld, dest_offset_batch, dest, + alpha, 0, 0, 0); +} + +#endif +// ================================================================================================= // End of the C++11 raw string literal )" diff --git a/src/kernels/level3/xgemm_batched.opencl b/src/kernels/level3/xgemm_batched.opencl index 372f910b..b51e6298 100644 --- a/src/kernels/level3/xgemm_batched.opencl +++ b/src/kernels/level3/xgemm_batched.opencl @@ -17,8 +17,8 @@ R"( // ================================================================================================= +#if defined(ROUTINE_GEMMBATCHED) -// 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, @@ -58,6 +58,49 @@ void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK, #endif } +#endif +// ================================================================================================= +#if defined(ROUTINE_GEMMSTRIDEDBATCHED) + +__kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1))) +void XgemmStridedBatched(const int kSizeM, const int kSizeN, const int kSizeK, + const real_arg arg_alpha, const real_arg arg_beta, + 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_alpha); + const real beta = GetRealArg(arg_beta); + + // 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 global memory + #if SA == 1 && SB == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, alpha, beta, alm, blm); + #elif SA == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, alpha, beta, alm); + #elif SB == 1 + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, alpha, beta, blm); + #else + XgemmBody(kSizeM, kSizeN, kSizeK, agm_, bgm_, cgm_, alpha, beta); + #endif +} + +#endif // ================================================================================================= // End of the C++11 raw string literal diff --git a/src/kernels/level3/xgemm_direct_batched.opencl b/src/kernels/level3/xgemm_direct_batched.opencl index d946a056..d15ed31e 100644 --- a/src/kernels/level3/xgemm_direct_batched.opencl +++ b/src/kernels/level3/xgemm_direct_batched.opencl @@ -17,15 +17,16 @@ R"( // ================================================================================================= +#if defined(ROUTINE_GEMMBATCHED) // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, non-transposed] __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK, - const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, - const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, - const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, - __global real* cgm, const __constant int* c_offsets, const int c_ld, - const int c_transpose, const int a_conjugate, const int b_conjugate) { + const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, + const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, + const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, + __global real* cgm, const __constant int* c_offsets, const int c_ld, + const int c_transpose, const int a_conjugate, const int b_conjugate) { const int batch = get_group_id(2); const real_arg arg_alpha = arg_alphas[batch]; const real_arg arg_beta = arg_betas[batch]; @@ -42,11 +43,11 @@ void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK, // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, transposed] __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK, - const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, - const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, - const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, - __global real* cgm, const __constant int* c_offsets, const int c_ld, - const int c_transpose, const int a_conjugate, const int b_conjugate) { + const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, + const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, + const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, + __global real* cgm, const __constant int* c_offsets, const int c_ld, + const int c_transpose, const int a_conjugate, const int b_conjugate) { const int batch = get_group_id(2); const real_arg arg_alpha = arg_alphas[batch]; const real_arg arg_beta = arg_betas[batch]; @@ -63,11 +64,11 @@ void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK, // Direct version of the batched GEMM kernel with [A, B] = [transposed, non-transposed] __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK, - const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, - const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, - const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, - __global real* cgm, const __constant int* c_offsets, const int c_ld, - const int c_transpose, const int a_conjugate, const int b_conjugate) { + const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, + const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, + const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, + __global real* cgm, const __constant int* c_offsets, const int c_ld, + const int c_transpose, const int a_conjugate, const int b_conjugate) { const int batch = get_group_id(2); const real_arg arg_alpha = arg_alphas[batch]; const real_arg arg_beta = arg_betas[batch]; @@ -84,11 +85,11 @@ void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK, // Direct version of the batched GEMM kernel with [A, B] = [transposed, transposed] __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK, - const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, - const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, - const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, - __global real* cgm, const __constant int* c_offsets, const int c_ld, - const int c_transpose, const int a_conjugate, const int b_conjugate) { + const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas, + const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld, + const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld, + __global real* cgm, const __constant int* c_offsets, const int c_ld, + const int c_transpose, const int a_conjugate, const int b_conjugate) { const int batch = get_group_id(2); const real_arg arg_alpha = arg_alphas[batch]; const real_arg arg_beta = arg_betas[batch]; @@ -102,6 +103,87 @@ void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK, alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate); } +#endif +// ================================================================================================= +#if defined(ROUTINE_GEMMSTRIDEDBATCHED) + +// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, non-transposed] +__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK, + const real_arg arg_alpha, const real_arg arg_beta, + const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride, + const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride, + __global real* cgm, const int c_offset, const int c_ld, const int c_stride, + const int c_transpose, const int a_conjugate, const int b_conjugate) { + const int batch = get_group_id(2); + const int a_offset_batch = a_offset + a_stride * batch; + const int b_offset_batch = b_offset + b_stride * batch; + const int c_offset_batch = c_offset + c_stride * batch; + __local real alm[WGD * (WGD + PADA)]; + __local real blm[WGD * (WGD + PADB)]; + XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta, + agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld, + alm, blm, 0, 0, c_transpose, a_conjugate, b_conjugate); +} + +// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, transposed] +__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK, + const real_arg arg_alpha, const real_arg arg_beta, + const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride, + const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride, + __global real* cgm, const int c_offset, const int c_ld, const int c_stride, + const int c_transpose, const int a_conjugate, const int b_conjugate) { + const int batch = get_group_id(2); + const int a_offset_batch = a_offset + a_stride * batch; + const int b_offset_batch = b_offset + b_stride * batch; + const int c_offset_batch = c_offset + c_stride * batch; + __local real alm[WGD * (WGD + PADA)]; + __local real blm[WGD * (WGD + PADB)]; + XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta, + agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld, + alm, blm, 0, 1, c_transpose, a_conjugate, b_conjugate); +} + +// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, non-transposed] +__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK, + const real_arg arg_alpha, const real_arg arg_beta, + const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride, + const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride, + __global real* cgm, const int c_offset, const int c_ld, const int c_stride, + const int c_transpose, const int a_conjugate, const int b_conjugate) { + const int batch = get_group_id(2); + const int a_offset_batch = a_offset + a_stride * batch; + const int b_offset_batch = b_offset + b_stride * batch; + const int c_offset_batch = c_offset + c_stride * batch; + __local real alm[WGD * (WGD + PADA)]; + __local real blm[WGD * (WGD + PADB)]; + XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta, + agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld, + alm, blm, 1, 0, c_transpose, a_conjugate, b_conjugate); +} + +// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, transposed] +__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1))) +void XgemmDirectStridedBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK, + const real_arg arg_alpha, const real_arg arg_beta, + const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride, + const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride, + __global real* cgm, const int c_offset, const int c_ld, const int c_stride, + const int c_transpose, const int a_conjugate, const int b_conjugate) { + const int batch = get_group_id(2); + const int a_offset_batch = a_offset + a_stride * batch; + const int b_offset_batch = b_offset + b_stride * batch; + const int c_offset_batch = c_offset + c_stride * batch; + __local real alm[WGD * (WGD + PADA)]; + __local real blm[WGD * (WGD + PADB)]; + XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta, + agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld, + alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate); +} + +#endif // ================================================================================================= // End of the C++11 raw string literal diff --git a/src/routines/common.hpp b/src/routines/common.hpp index 06d001d9..6cbe1e1b 100644 --- a/src/routines/common.hpp +++ b/src/routines/common.hpp @@ -239,6 +239,72 @@ void PadCopyTransposeMatrixBatched(Queue &queue, const Device &device, } } +// Batched version of the above +template <typename T> +void PadCopyTransposeMatrixStridedBatched(Queue &queue, const Device &device, + const Databases &db, + EventPointer event, const std::vector<Event> &waitForEvents, + const size_t src_one, const size_t src_two, + const size_t src_ld, const size_t src_offset, + const size_t src_stride, const Buffer<T> &src, + const size_t dest_one, const size_t dest_two, + const size_t dest_ld, const size_t dest_offset, + const size_t dest_stride, const Buffer<T> &dest, + const Program &program, const bool do_pad, + const bool do_transpose, const bool do_conjugate, + const size_t batch_count) { + + // Determines the right kernel + auto kernel_name = std::string{}; + if (do_transpose) { + kernel_name = (do_pad) ? "TransposePadMatrixStridedBatched" : "TransposeMatrixStridedBatched"; + } + else { + kernel_name = (do_pad) ? "CopyPadMatrixStridedBatched" : "CopyMatrixStridedBatched"; + } + + // Retrieves the kernel from the compiled binary + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast<int>(src_one)); + kernel.SetArgument(1, static_cast<int>(src_two)); + kernel.SetArgument(2, static_cast<int>(src_ld)); + kernel.SetArgument(3, static_cast<int>(src_offset)); + kernel.SetArgument(4, static_cast<int>(src_stride)); + kernel.SetArgument(5, src()); + kernel.SetArgument(6, static_cast<int>(dest_one)); + kernel.SetArgument(7, static_cast<int>(dest_two)); + kernel.SetArgument(8, static_cast<int>(dest_ld)); + kernel.SetArgument(9, static_cast<int>(dest_offset)); + kernel.SetArgument(10, static_cast<int>(dest_stride)); + kernel.SetArgument(11, dest()); + if (do_pad) { + kernel.SetArgument(12, static_cast<int>(do_conjugate)); + } + + // Launches the kernel and returns the error code. Uses global and local thread sizes based on + // parameters in the database. + if (do_transpose) { + const auto global = std::vector<size_t>{ + Ceil(CeilDiv(dest_one, db["PADTRA_WPT"]), db["PADTRA_TILE"]), + Ceil(CeilDiv(dest_two, db["PADTRA_WPT"]), db["PADTRA_TILE"]), + batch_count + }; + const auto local = std::vector<size_t>{db["PADTRA_TILE"], db["PADTRA_TILE"], 1}; + RunKernel(kernel, queue, device, global, local, event, waitForEvents); + } + else { + const auto global = std::vector<size_t>{ + Ceil(CeilDiv(dest_one, db["PAD_WPTX"]), db["PAD_DIMX"]), + Ceil(CeilDiv(dest_two, db["PAD_WPTY"]), db["PAD_DIMY"]), + batch_count + }; + const auto local = std::vector<size_t>{db["PAD_DIMX"], db["PAD_DIMY"], 1}; + RunKernel(kernel, queue, device, global, local, event, waitForEvents); + } +} + // ================================================================================================= } // namespace clblast diff --git a/src/routines/levelx/xgemmbatched.cpp b/src/routines/levelx/xgemmbatched.cpp index 8ce2dedc..1c0953e8 100644 --- a/src/routines/levelx/xgemmbatched.cpp +++ b/src/routines/levelx/xgemmbatched.cpp @@ -12,6 +12,7 @@ // ================================================================================================= #include "routines/levelx/xgemmbatched.hpp" +#include "routines/level3/xgemm.hpp" #include <string> #include <vector> @@ -64,34 +65,12 @@ void XgemmBatched<T>::DoGemmBatched(const Layout layout, const Transpose a_trans throw BLASError(StatusCode::kInvalidBatchCount); } - // Makes sure all dimensions are larger than zero - if ((m == 0) || (n == 0) || (k == 0)) { throw BLASError(StatusCode::kInvalidDimension); } - - // Computes whether or not the matrices are transposed in memory. See GEMM routine for details. - const auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && a_transpose == Transpose::kNo); - const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && b_transpose == Transpose::kNo); - const auto c_rotated = (layout == Layout::kRowMajor); - static const auto a_want_rotated = false; - static const auto b_want_rotated = true; - static const auto c_want_rotated = false; - const auto a_do_transpose = a_rotated != a_want_rotated; - const auto b_do_transpose = b_rotated != b_want_rotated; - const auto c_do_transpose = c_rotated != c_want_rotated; - - // In case of complex data-types, the transpose can also become a conjugate transpose - const auto a_conjugate = (a_transpose == Transpose::kConjugate); - const auto b_conjugate = (b_transpose == Transpose::kConjugate); - - // Computes the first and second dimensions of the 3 matrices taking into account whether the - // matrices are rotated or not - const auto a_one = (a_rotated) ? k : m; - const auto a_two = (a_rotated) ? m : k; - const auto b_one = (b_rotated) ? n : k; - const auto b_two = (b_rotated) ? k : n; - const auto c_one = (c_rotated) ? n : m; - const auto c_two = (c_rotated) ? m : n; + // Computes the transpose/conjugate options and sets the a/b/c sizes based on that + bool a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate; + size_t a_one, a_two, b_one, b_two, c_one, c_two; + Xgemm<T>::ProcessArguments(layout, a_transpose, b_transpose, m, n, k, + a_one, a_two, b_one, b_two, c_one, c_two, + a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate); // Tests the matrices for validity for (auto batch = size_t{0}; batch < batch_count; ++batch) { @@ -130,10 +109,7 @@ void XgemmBatched<T>::DoGemmBatched(const Layout layout, const Transpose a_trans a_buffer, a_offsets_int, a_ld, b_buffer, b_offsets_int, b_ld, betas_device, c_buffer, c_offsets_int, c_ld, a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate, - a_one, a_two, a_want_rotated, - b_one, b_two, b_want_rotated, - c_one, c_two, c_want_rotated, - batch_count); + a_one, a_two, b_one, b_two, c_one, c_two, batch_count); } } @@ -152,9 +128,9 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const const Buffer<T> &c_buffer, const std::vector<int> &c_offsets, const size_t c_ld, const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, const bool a_conjugate, const bool b_conjugate, - const size_t a_one, const size_t a_two, const bool a_want_rotated, - const size_t b_one, const size_t b_two, const bool b_want_rotated, - const size_t c_one, const size_t c_two, const bool c_want_rotated, + const size_t a_one, const size_t a_two, + const size_t b_one, const size_t b_two, + const size_t c_one, const size_t c_two, const size_t batch_count) { // Calculates the ceiled versions of m, n, and k const auto m_ceiled = Ceil(Ceil(m, db_["MWG"]), db_["VWM"]); @@ -163,12 +139,9 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const // Computes the first and second "internal" (ceiled) dimensions of the 3 matrices taking into account // whether the matrices need to be rotated or not for the kernel. - const auto a_one_i = (a_want_rotated) ? k_ceiled : m_ceiled; - const auto a_two_i = (a_want_rotated) ? m_ceiled : k_ceiled; - const auto b_one_i = (b_want_rotated) ? n_ceiled : k_ceiled; - const auto b_two_i = (b_want_rotated) ? k_ceiled : n_ceiled; - const auto c_one_i = (c_want_rotated) ? n_ceiled : m_ceiled; - const auto c_two_i = (c_want_rotated) ? m_ceiled : n_ceiled; + size_t a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i; + Xgemm<T>::CalculateInternalDimensions(m, n, k, db_["MWG"], db_["NWG"], db_["KWG"], + a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i); // Sets the "internal" offsets, i.e. the perfect offsets auto a_offsets_i = std::vector<int>(batch_count); @@ -182,11 +155,11 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const // Determines whether or not temporary matrices are needed auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offsets == a_offsets_i && - a_do_transpose == false && a_conjugate == false; + !a_do_transpose && !a_conjugate; auto b_no_temp = b_one == b_one_i && b_two == b_two_i && b_ld == b_one && b_offsets == b_offsets_i && - b_do_transpose == false && b_conjugate == false; + !b_do_transpose && !b_conjugate; auto c_no_temp = c_one == c_one_i && c_two == c_two_i && c_ld == c_one && c_offsets == c_offsets_i && - c_do_transpose == false; + !c_do_transpose; // Creates the temporary matrices const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, batch_count * a_one_i * a_two_i); diff --git a/src/routines/levelx/xgemmbatched.hpp b/src/routines/levelx/xgemmbatched.hpp index 6136dd5f..989f3815 100644 --- a/src/routines/levelx/xgemmbatched.hpp +++ b/src/routines/levelx/xgemmbatched.hpp @@ -48,9 +48,9 @@ class XgemmBatched: public Routine { const Buffer<T> &c_buffer, const std::vector<int> &c_offsets, const size_t c_ld, const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, const bool a_conjugate, const bool b_conjugate, - const size_t a_one, const size_t a_two, const bool a_want_rotated, - const size_t b_one, const size_t b_two, const bool b_want_rotated, - const size_t c_one, const size_t c_two, const bool c_want_rotated, + const size_t a_one, const size_t a_two, + const size_t b_one, const size_t b_two, + const size_t c_one, const size_t c_two, const size_t batch_count); // Direct version of batched GEMM (no pre and post-processing kernels) diff --git a/src/routines/levelx/xgemmstridedbatched.cpp b/src/routines/levelx/xgemmstridedbatched.cpp new file mode 100644 index 00000000..affbceee --- /dev/null +++ b/src/routines/levelx/xgemmstridedbatched.cpp @@ -0,0 +1,278 @@ + +// ================================================================================================= +// 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 implements the XgemmStridedBatched class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/levelx/xgemmstridedbatched.hpp" +#include "routines/level3/xgemm.hpp" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +XgemmStridedBatched<T>::XgemmStridedBatched(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm","XgemmDirect","GemmRoutine"}, + PrecisionValue<T>(), {}, { + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + , // separated in multiple parts to prevent C1091 in MSVC 2013 + #include "../../kernels/level3/xgemm_direct_part1.opencl" + #include "../../kernels/level3/xgemm_direct_part2.opencl" + #include "../../kernels/level3/xgemm_direct_part3.opencl" + , // separated in multiple parts to prevent C1091 in MSVC 2013 + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + #include "../../kernels/level3/xgemm_part3.opencl" + #include "../../kernels/level3/xgemm_part4.opencl" + , // separated in multiple parts to prevent C1091 in MSVC 2013 + #include "../../kernels/level3/xgemm_batched.opencl" + #include "../../kernels/level3/xgemm_direct_batched.opencl" + }) { +} + +// ================================================================================================= + +// The main routine +template <typename T> +void XgemmStridedBatched<T>::DoGemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count) { + + // Tests for a valid batch count + if (batch_count < 1) { + throw BLASError(StatusCode::kInvalidBatchCount); + } + + // Computes the transpose/conjugate options and sets the a/b/c sizes based on that + bool a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate; + size_t a_one, a_two, b_one, b_two, c_one, c_two; + Xgemm<T>::ProcessArguments(layout, a_transpose, b_transpose, m, n, k, + a_one, a_two, b_one, b_two, c_one, c_two, + a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate); + + // Tests the matrices for validity + for (auto batch = size_t{0}; batch < batch_count; ++batch) { + TestMatrixA(a_one, a_two, a_buffer, a_offset + a_stride * batch, a_ld); + TestMatrixB(b_one, b_two, b_buffer, b_offset + b_stride * batch, b_ld); + TestMatrixC(c_one, c_two, c_buffer, c_offset + c_stride * batch, c_ld); + } + + // Selects which version of the batched GEMM to run + const auto do_gemm_direct = true; + if (do_gemm_direct) { // single generic kernel + BatchedGemmDirect(m, n, k, alpha, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, beta, + c_buffer, c_offset, c_ld, c_stride, + a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate, + batch_count); + } + else { // pre/post-processing plus a very fast kernel + BatchedGemmIndirect(m, n, k, alpha, + a_buffer, a_offset, a_ld, a_stride, + b_buffer, b_offset, b_ld, b_stride, beta, + c_buffer, c_offset, c_ld, c_stride, + a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate, + a_one, a_two, b_one, b_two, c_one, c_two, batch_count); + } +} + + +// ================================================================================================= + +// The indirect version of batched GEMM. This uses the faster but non-general kernel. It has specific +// requirements, but several pre and post-processing kernels take care of those. However, the +// overhead of these extra kernels might not be ideal for certain devices/arguments. +template <typename T> +void XgemmStridedBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, + const bool a_conjugate, const bool b_conjugate, + const size_t a_one, const size_t a_two, + const size_t b_one, const size_t b_two, + const size_t c_one, const size_t c_two, + const size_t batch_count) { + + // Calculates the ceiled versions of m, n, and k + const auto m_ceiled = Ceil(Ceil(m, db_["MWG"]), db_["VWM"]); + const auto n_ceiled = Ceil(Ceil(n, db_["NWG"]), db_["VWN"]); + const auto k_ceiled = Ceil(Ceil(k, db_["KWG"]), db_["VWM"]); + + // Computes the first and second "internal" (ceiled) dimensions of the 3 matrices taking into account + // whether the matrices need to be rotated or not for the kernel. + size_t a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i; + Xgemm<T>::CalculateInternalDimensions(m, n, k, db_["MWG"], db_["NWG"], db_["KWG"], + a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && !a_do_transpose && !a_conjugate; + auto b_no_temp = b_one == b_one_i && b_two == b_two_i && b_ld == b_one && !b_do_transpose && !b_conjugate; + auto c_no_temp = c_one == c_one_i && c_two == c_two_i && c_ld == c_one && !c_do_transpose; + + // Creates the temporary matrices + const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, batch_count * a_one_i * a_two_i); + const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, batch_count * b_one_i * b_two_i); + const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, batch_count * c_one_i * c_two_i); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector<Event>(); + auto emptyEventList = std::vector<Event>(); + + // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. + if (!a_no_temp) { + auto eventProcessA = Event(); + PadCopyTransposeMatrixStridedBatched(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_stride, a_buffer, + a_one_i, a_two_i, a_one_i, 0, a_one_i * a_two_i, a_temp, + program_, true, a_do_transpose, a_conjugate, batch_count); + eventWaitList.push_back(eventProcessA); + } + + // As above, but now for matrix B + if (!b_no_temp) { + auto eventProcessB = Event(); + PadCopyTransposeMatrixStridedBatched(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, + b_one, b_two, b_ld, b_offset, b_stride, b_buffer, + b_one_i, b_two_i, b_one_i, 0, b_one_i * b_two_i, b_temp, + program_, true, b_do_transpose, b_conjugate, batch_count); + eventWaitList.push_back(eventProcessB); + } + + // As above, but now for matrix C + if (!c_no_temp) { + auto eventProcessC = Event(); + PadCopyTransposeMatrixStridedBatched(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, + c_one, c_two, c_ld, c_offset, c_stride, c_buffer, + c_one_i, c_two_i, c_one_i, 0, c_one_i * c_two_i, c_temp, + program_, true, c_do_transpose, false, batch_count); + eventWaitList.push_back(eventProcessC); + } + + // Retrieves the Xgemm kernel from the compiled binary + auto kernel = Kernel(program_, "XgemmStridedBatched"); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast<int>(m_ceiled)); + kernel.SetArgument(1, static_cast<int>(n_ceiled)); + kernel.SetArgument(2, static_cast<int>(k_ceiled)); + kernel.SetArgument(3, GetRealArg(alpha)); + kernel.SetArgument(4, GetRealArg(beta)); + kernel.SetArgument(5, a_temp()); + kernel.SetArgument(6, static_cast<int>(a_one_i)); + kernel.SetArgument(7, static_cast<int>(a_two_i)); + kernel.SetArgument(8, b_temp()); + kernel.SetArgument(9, static_cast<int>(b_one_i)); + kernel.SetArgument(10, static_cast<int>(b_two_i)); + kernel.SetArgument(11, c_temp()); + kernel.SetArgument(12, static_cast<int>(c_one_i)); + kernel.SetArgument(13, static_cast<int>(c_two_i)); + + // Computes the global and local thread sizes + const auto global = std::vector<size_t>{ + (c_one_i * db_["MDIMC"]) / db_["MWG"], + (c_two_i * db_["NDIMC"]) / db_["NWG"], + batch_count + }; + const auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"], 1}; + + // Launches the kernel + auto eventKernel = Event(); + auto eventPointer = eventKernel.pointer(); + RunKernel(kernel, queue_, device_, global, local, eventPointer, eventWaitList); + + // Runs the post-processing kernel if needed + if (!c_no_temp) { + eventWaitList.push_back(eventKernel); + PadCopyTransposeMatrixStridedBatched(queue_, device_, db_, event_, eventWaitList, + c_one_i, c_two_i, c_one_i, 0, c_one_i * c_two_i, c_temp, + c_one, c_two, c_ld, c_offset, c_stride, c_buffer, + program_, false, c_do_transpose, false, batch_count); + } +} + +// ================================================================================================= + +// The direct version of batched GEMM, requiring just one kernel, no pre or post-processing kernels. +template <typename T> +void XgemmStridedBatched<T>::BatchedGemmDirect(const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, + const bool a_conjugate, const bool b_conjugate, + const size_t batch_count) { + + // Retrieves the proper XgemmDirect kernel from the compiled binary + const auto name = (a_do_transpose) ? (b_do_transpose ? "XgemmDirectStridedBatchedTT" : "XgemmDirectStridedBatchedTN") : + (b_do_transpose ? "XgemmDirectStridedBatchedNT" : "XgemmDirectStridedBatchedNN"); + auto kernel = Kernel(program_, name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast<int>(m)); + kernel.SetArgument(1, static_cast<int>(n)); + kernel.SetArgument(2, static_cast<int>(k)); + kernel.SetArgument(3, GetRealArg(alpha)); + kernel.SetArgument(4, GetRealArg(beta)); + kernel.SetArgument(5, a_buffer()); + kernel.SetArgument(6, static_cast<int>(a_offset)); + kernel.SetArgument(7, static_cast<int>(a_ld)); + kernel.SetArgument(8, static_cast<int>(a_stride)); + kernel.SetArgument(9, b_buffer()); + kernel.SetArgument(10, static_cast<int>(b_offset)); + kernel.SetArgument(11, static_cast<int>(b_ld)); + kernel.SetArgument(12, static_cast<int>(b_stride)); + kernel.SetArgument(13, c_buffer()); + kernel.SetArgument(14, static_cast<int>(c_offset)); + kernel.SetArgument(15, static_cast<int>(c_ld)); + kernel.SetArgument(16, static_cast<int>(c_stride)); + kernel.SetArgument(17, static_cast<int>(c_do_transpose)); + kernel.SetArgument(18, static_cast<int>(a_conjugate)); + kernel.SetArgument(19, static_cast<int>(b_conjugate)); + + // Computes the global and local thread sizes + const auto m_ceiled = Ceil(m, db_["WGD"]); + const auto n_ceiled = Ceil(n, db_["WGD"]); + const auto global = std::vector<size_t>{ + (m_ceiled * db_["MDIMCD"]) / db_["WGD"], + (n_ceiled * db_["NDIMCD"]) / db_["WGD"], + batch_count + }; + const auto local = std::vector<size_t>{db_["MDIMCD"], db_["NDIMCD"], 1}; + + // Launches the kernel + RunKernel(kernel, queue_, device_, global, local, event_); +} + +// ================================================================================================= + +// Compiles the templated class +template class XgemmStridedBatched<half>; +template class XgemmStridedBatched<float>; +template class XgemmStridedBatched<double>; +template class XgemmStridedBatched<float2>; +template class XgemmStridedBatched<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/levelx/xgemmstridedbatched.hpp b/src/routines/levelx/xgemmstridedbatched.hpp new file mode 100644 index 00000000..0dbbcb10 --- /dev/null +++ b/src/routines/levelx/xgemmstridedbatched.hpp @@ -0,0 +1,66 @@ + +// ================================================================================================= +// 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 implements the XgemmStridedBatched routine. This is a non-blas batched version of GEMM. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ +#define CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ + +#include <vector> + +#include "routine.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class XgemmStridedBatched: public Routine { +public: + + // Constructor + XgemmStridedBatched(Queue &queue, EventPointer event, const std::string &name = "GEMMSTRIDEDBATCHED"); + + // Templated-precision implementation of the routine + void DoGemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose, + const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const size_t batch_count); + + // Indirect version of strided batched GEMM (with pre and post-processing kernels) + void BatchedGemmIndirect(const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, + const bool a_conjugate, const bool b_conjugate, + const size_t a_one, const size_t a_two, + const size_t b_one, const size_t b_two, + const size_t c_one, const size_t c_two, + const size_t batch_count); + + // Direct version of strided batched GEMM (no pre and post-processing kernels) + void BatchedGemmDirect(const size_t m, const size_t n, const size_t k, const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride, + const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta, + const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride, + const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose, + const bool a_conjugate, const bool b_conjugate, + const size_t batch_count); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ +#endif diff --git a/src/routines/routines.hpp b/src/routines/routines.hpp index 9e7768b9..0aeff707 100644 --- a/src/routines/routines.hpp +++ b/src/routines/routines.hpp @@ -71,6 +71,7 @@ #include "routines/levelx/xim2col.hpp" #include "routines/levelx/xaxpybatched.hpp" #include "routines/levelx/xgemmbatched.hpp" +#include "routines/levelx/xgemmstridedbatched.hpp" // CLBLAST_ROUTINES_ROUTINES_H_ #endif diff --git a/test/correctness/routines/levelx/xgemmstridedbatched.cpp b/test/correctness/routines/levelx/xgemmstridedbatched.cpp new file mode 100644 index 00000000..d2ea19d0 --- /dev/null +++ b/test/correctness/routines/levelx/xgemmstridedbatched.cpp @@ -0,0 +1,26 @@ + +// ================================================================================================= +// 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> +// +// ================================================================================================= + +#include "test/correctness/testblas.hpp" +#include "test/routines/levelx/xgemmstridedbatched.hpp" + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + auto errors = size_t{0}; + errors += clblast::RunTests<clblast::TestXgemmStridedBatched<float>, float, float>(argc, argv, false, "SGEMMSTRIDEDBATCHED"); + errors += clblast::RunTests<clblast::TestXgemmStridedBatched<double>, double, double>(argc, argv, true, "DGEMMSTRIDEDBATCHED"); + errors += clblast::RunTests<clblast::TestXgemmStridedBatched<clblast::float2>, clblast::float2, clblast::float2>(argc, argv, true, "CGEMMSTRIDEDBATCHED"); + errors += clblast::RunTests<clblast::TestXgemmStridedBatched<clblast::double2>, clblast::double2, clblast::double2>(argc, argv, true, "ZGEMMSTRIDEDBATCHED"); + errors += clblast::RunTests<clblast::TestXgemmStridedBatched<clblast::half>, clblast::half, clblast::half>(argc, argv, true, "HGEMMSTRIDEDBATCHED"); + if (errors > 0) { return 1; } else { return 0; } +} + +// ================================================================================================= diff --git a/test/performance/routines/levelx/xgemmstridedbatched.cpp b/test/performance/routines/levelx/xgemmstridedbatched.cpp new file mode 100644 index 00000000..5358e466 --- /dev/null +++ b/test/performance/routines/levelx/xgemmstridedbatched.cpp @@ -0,0 +1,33 @@ + +// ================================================================================================= +// 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> +// +// ================================================================================================= + +#include "test/performance/client.hpp" +#include "test/routines/levelx/xgemmstridedbatched.hpp" + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv); + switch(clblast::GetPrecision(command_line_args, clblast::Precision::kSingle)) { + case clblast::Precision::kHalf: + clblast::RunClient<clblast::TestXgemmStridedBatched<clblast::half>, clblast::half, clblast::half>(argc, argv); break; + case clblast::Precision::kSingle: + clblast::RunClient<clblast::TestXgemmStridedBatched<float>, float, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient<clblast::TestXgemmStridedBatched<double>, double, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + clblast::RunClient<clblast::TestXgemmStridedBatched<clblast::float2>, clblast::float2, clblast::float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient<clblast::TestXgemmStridedBatched<clblast::double2>, clblast::double2, clblast::double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/levelx/xaxpybatched.hpp b/test/routines/levelx/xaxpybatched.hpp index 9a09b47f..c7f6e40d 100644 --- a/test/routines/levelx/xaxpybatched.hpp +++ b/test/routines/levelx/xaxpybatched.hpp @@ -62,7 +62,7 @@ class TestXaxpyBatched { for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { args.x_offsets[batch] = batch * PerBatchSizeX(args) + args.x_offset; args.y_offsets[batch] = batch * PerBatchSizeY(args) + args.y_offset; - args.alphas[batch] = args.alpha + Constant<T>(static_cast<double>(batch)); + args.alphas[batch] = args.alpha + Constant<T>(static_cast<double>(batch + 1)); } } diff --git a/test/routines/levelx/xgemmbatched.hpp b/test/routines/levelx/xgemmbatched.hpp index e237a018..b787ca27 100644 --- a/test/routines/levelx/xgemmbatched.hpp +++ b/test/routines/levelx/xgemmbatched.hpp @@ -86,8 +86,8 @@ class TestXgemmBatched { args.a_offsets[batch] = batch * PerBatchSizeA(args) + args.a_offset; args.b_offsets[batch] = batch * PerBatchSizeB(args) + args.b_offset; args.c_offsets[batch] = batch * PerBatchSizeC(args) + args.c_offset; - args.alphas[batch] = args.alpha + Constant<T>(static_cast<double>(batch)); - args.betas[batch] = args.beta + Constant<T>(static_cast<double>(batch)); + args.alphas[batch] = args.alpha + Constant<T>(static_cast<double>(batch + 1)); + args.betas[batch] = args.beta + Constant<T>(static_cast<double>(batch + 1)); } } diff --git a/test/routines/levelx/xgemmstridedbatched.hpp b/test/routines/levelx/xgemmstridedbatched.hpp new file mode 100644 index 00000000..ddb32997 --- /dev/null +++ b/test/routines/levelx/xgemmstridedbatched.hpp @@ -0,0 +1,218 @@ + +// ================================================================================================= +// 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 implements a class with static methods to describe the XgemmStridedBatched routine. Examples of +// such 'descriptions' are how to calculate the size a of buffer or how to run the routine. These +// static methods are used by the correctness tester and the performance tester. +// +// ================================================================================================= + +#ifndef CLBLAST_TEST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ +#define CLBLAST_TEST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ + +#include "test/routines/common.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXgemmStridedBatched { +public: + + // Although it is a non-BLAS routine, it can still be tested against level-3 routines in a loop + static size_t BLASLevel() { return 3; } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgM, kArgN, kArgK, + kArgLayout, kArgATransp, kArgBTransp, + kArgALeadDim, kArgBLeadDim, kArgCLeadDim, + kArgAOffset, kArgBOffset, kArgCOffset, + kArgBatchCount, kArgAlpha, kArgBeta}; + } + static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; } + static std::vector<std::string> BuffersOut() { return {kBufMatC}; } + + // Helper for the sizes per batch + static size_t PerBatchSizeA(const Arguments<T> &args) { + auto a_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + auto a_two = (a_rotated) ? args.m : args.k; + return a_two * args.a_ld; + } + static size_t PerBatchSizeB(const Arguments<T> &args) { + auto b_rotated = (args.layout == Layout::kColMajor && args.b_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.b_transpose == Transpose::kNo); + auto b_two = (b_rotated) ? args.k : args.n; + return b_two * args.b_ld; + } + static size_t PerBatchSizeC(const Arguments<T> &args) { + auto c_rotated = (args.layout == Layout::kRowMajor); + auto c_two = (c_rotated) ? args.m : args.n; + return c_two * args.c_ld; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeA(const Arguments<T> &args) { + return PerBatchSizeA(args) * args.batch_count + args.a_offset; + } + static size_t GetSizeB(const Arguments<T> &args) { + return PerBatchSizeB(args) * args.batch_count + args.b_offset; + } + static size_t GetSizeC(const Arguments<T> &args) { + return PerBatchSizeC(args) * args.batch_count + args.c_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args, Queue&) { + args.a_size = GetSizeA(args); + args.b_size = GetSizeB(args); + args.c_size = GetSizeC(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.k; } + static size_t DefaultLDB(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDC(const Arguments<T> &args) { return args.n; } + + // Describes which transpose options are relevant for this routine + using Transposes = std::vector<Transpose>; + static Transposes GetATransposes(const Transposes &all) { return all; } + static Transposes GetBTransposes(const Transposes &all) { return all; } + + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + #ifdef OPENCL_API + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = GemmStridedBatched(args.layout, args.a_transpose, args.b_transpose, + args.m, args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, PerBatchSizeA(args), + buffers.b_mat(), args.b_offset, args.b_ld, PerBatchSizeB(args), args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, PerBatchSizeC(args), + args.batch_count, + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } + #elif CUDA_API + auto status = GemmStridedBatched(args.layout, args.a_transpose, args.b_transpose, + args.m, args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, PerBatchSizeA(args), + buffers.b_mat(), args.b_offset, args.b_ld, PerBatchSizeB(args), args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, PerBatchSizeC(args), + args.batch_count, + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + #ifdef CLBLAST_REF_CLBLAS + static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { + const auto a_batch_offset = args.a_offset + PerBatchSizeA(args) * batch; + const auto b_batch_offset = args.c_offset + PerBatchSizeB(args) * batch; + const auto c_batch_offset = args.b_offset + PerBatchSizeC(args) * batch; + auto event = cl_event{}; + auto status = clblasXgemm(convertToCLBLAS(args.layout), + convertToCLBLAS(args.a_transpose), + convertToCLBLAS(args.b_transpose), + args.m, args.n, args.k, args.alpha, + buffers.a_mat, a_batch_offset, args.a_ld, + buffers.b_mat, b_batch_offset, args.b_ld, args.beta, + buffers.c_mat, c_batch_offset, args.c_ld, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + if (static_cast<StatusCode>(status) != StatusCode::kSuccess) { + return static_cast<StatusCode>(status); + } + } + return StatusCode::kSuccess; + } + #endif + + // Describes how to run the CPU BLAS routine (for correctness/performance comparison) + #ifdef CLBLAST_REF_CBLAS + static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) { + for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { + const auto a_batch_offset = args.a_offset + PerBatchSizeA(args) * batch; + const auto b_batch_offset = args.c_offset + PerBatchSizeB(args) * batch; + const auto c_batch_offset = args.b_offset + PerBatchSizeC(args) * batch; + cblasXgemm(convertToCBLAS(args.layout), + convertToCBLAS(args.a_transpose), + convertToCBLAS(args.b_transpose), + args.m, args.n, args.k, args.alpha, + buffers_host.a_mat, a_batch_offset, args.a_ld, + buffers_host.b_mat, b_batch_offset, args.b_ld, args.beta, + buffers_host.c_mat, c_batch_offset, args.c_ld); + } + return StatusCode::kSuccess; + } + #endif + + // Describes how to run the cuBLAS routine (for correctness/performance comparison) + #ifdef CLBLAST_REF_CUBLAS + static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { + for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { + const auto a_batch_offset = args.a_offset + PerBatchSizeA(args) * batch; + const auto b_batch_offset = args.c_offset + PerBatchSizeB(args) * batch; + const auto c_batch_offset = args.b_offset + PerBatchSizeC(args) * batch; + auto status = cublasXgemm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, + convertToCUBLAS(args.a_transpose), + convertToCUBLAS(args.b_transpose), + args.m, args.n, args.k, args.alpha, + buffers.a_mat, a_batch_offset, args.a_ld, + buffers.b_mat, b_batch_offset, args.b_ld, args.beta, + buffers.c_mat, c_batch_offset, args.c_ld); + if (status != CUBLAS_STATUS_SUCCESS) { return StatusCode::kUnknownError; } + } + return StatusCode::kSuccess; + } + #endif + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + std::vector<T> result(args.c_size, static_cast<T>(0)); + buffers.c_mat.Read(queue, args.c_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { return args.m; } + static size_t ResultID2(const Arguments<T> &args) { return args.n * args.batch_count; } + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t id2_3) { + const size_t id2 = id2_3 % args.n; + const size_t id3 = id2_3 / args.n; + const auto c_batch_offset = args.c_offset + PerBatchSizeC(args) * id3; + return (args.layout == Layout::kRowMajor) ? + id1*args.c_ld + id2 + c_batch_offset: + id2*args.c_ld + id1 + c_batch_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return args.batch_count * (2 * args.m * args.n * args.k); + } + static size_t GetBytes(const Arguments<T> &args) { + return args.batch_count * (args.m*args.k + args.k*args.n + 2*args.m*args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XGEMMSTRIDEDBATCHED_H_ +#endif |