diff options
| -rw-r--r-- | CHANGELOG | 3 | ||||
| -rw-r--r-- | CMakeLists.txt | 2 | ||||
| -rw-r--r-- | README.md | 10 | ||||
| -rw-r--r-- | include/internal/routines/level2/xtbmv.h | 51 | ||||
| -rw-r--r-- | include/internal/routines/level2/xtpmv.h | 51 | ||||
| -rw-r--r-- | include/internal/routines/level2/xtrmv.h | 51 | ||||
| -rw-r--r-- | scripts/generator/generator.py | 10 | ||||
| -rw-r--r-- | src/clblast.cc | 63 | ||||
| -rw-r--r-- | src/kernels/level2/xgemv.opencl | 67 | ||||
| -rw-r--r-- | src/routines/level2/xtbmv.cc | 81 | ||||
| -rw-r--r-- | src/routines/level2/xtpmv.cc | 81 | ||||
| -rw-r--r-- | src/routines/level2/xtrmv.cc | 81 | ||||
| -rw-r--r-- | test/routines/level2/xtbmv.h | 125 | ||||
| -rw-r--r-- | test/routines/level2/xtpmv.h | 125 | ||||
| -rw-r--r-- | test/routines/level2/xtrmv.h | 125 | ||||
| -rw-r--r-- | test/wrapper_clblas.h | 36 |
16 files changed, 915 insertions, 47 deletions
@@ -14,6 +14,9 @@ Development version (next release) * CHPMV/ZHPMV * SSBMV/DSBMV * SSPMV/DSPMV + * STRMV/DTRMV/CTRMV/ZTRMV + * STBMV/DTBMV/CTBMV/ZTBMV + * STPMV/DTPMV/CTPMV/ZTPMV Version 0.4.0 - Now using the Claduc C++11 interface to OpenCL diff --git a/CMakeLists.txt b/CMakeLists.txt index 7e670664..1ddd2f77 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -106,7 +106,7 @@ set(KERNELS copy pad transpose padtranspose xaxpy xdot xgemv xgemm) set(SAMPLE_PROGRAMS_CPP sgemm) set(SAMPLE_PROGRAMS_C sgemm) set(LEVEL1_ROUTINES xswap xscal xcopy xaxpy xdot xdotu xdotc) -set(LEVEL2_ROUTINES xgemv xgbmv xhemv xhbmv xhpmv xsymv xsbmv xspmv) +set(LEVEL2_ROUTINES xgemv xgbmv xhemv xhbmv xhpmv xsymv xsbmv xspmv xtrmv xtbmv xtpmv) set(LEVEL3_ROUTINES xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm) set(ROUTINES ${LEVEL1_ROUTINES} ${LEVEL2_ROUTINES} ${LEVEL3_ROUTINES}) set(PRECISIONS 32 3232 64 6464) @@ -6,7 +6,7 @@ CLBlast: The tuned OpenCL BLAS library CLBlast is a modern, lightweight, performant and tunable OpenCL BLAS library written in C++11. It is designed to leverage the full performance potential of a wide variety of OpenCL devices from different vendors, including desktop and laptop GPUs, embedded GPUs, and other accelerators. CLBlast implements BLAS routines: basic linear algebra subprograms operating on vectors and matrices. -__Note that the CLBlast library is actively being developed, and is not mature enough for production environments__. This preview-version doesn't support all routines yet: others will be added in due time. It also lacks extensive tuning on some common OpenCL platforms: __out-of-the-box performance on some devices might be poor__. See below for more details. +__Note that the CLBlast library is actively being developed, and is not mature enough for production environments__. This preview-version doesn't support the less commonly used routines yet: they will be added in due time. It also lacks extensive tuning on some common OpenCL platforms: __out-of-the-box performance on some devices might be poor__. See below for more details. Why CLBlast and not clBLAS or cuBLAS? @@ -130,7 +130,7 @@ These graphs can be generated automatically on your own device. First, compile C Supported routines ------------- -CLBlast is in active development and currently does not support the full set of BLAS routines. The currently supported routines are marked with '✔' in the following tables: +CLBlast is in active development but already supports the majority of BLAS routines. The currently supported routines are marked with '✔' in the following tables: | Level-1 | S | D | C | Z | Notes | | ---------|---|---|---|---|---------| @@ -160,9 +160,9 @@ CLBlast is in active development and currently does not support the full set of | xSYMV | ✔ | ✔ | - | - | | | xSBMV | ✔ | ✔ | - | - | | | xSPMV | ✔ | ✔ | - | - | | -| xTRMV | | | | | | -| xTBMV | | | | | | -| xTPMV | | | | | | +| xTRMV | ✔ | ✔ | ✔ | ✔ | | +| xTBMV | ✔ | ✔ | ✔ | ✔ | | +| xTPMV | ✔ | ✔ | ✔ | ✔ | | | xTRSV | | | | | | | xTBSV | | | | | | | xTPSV | | | | | | diff --git a/include/internal/routines/level2/xtbmv.h b/include/internal/routines/level2/xtbmv.h new file mode 100644 index 00000000..89c90193 --- /dev/null +++ b/include/internal/routines/level2/xtbmv.h @@ -0,0 +1,51 @@ + +// ================================================================================================= +// 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 Xtbmv routine. It is based on the generalized mat-vec multiplication +// routine (Xgemv). The Xtbmv class inherits from the templated class Xgemv, allowing it to call the +// "MatVec" function directly. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XTBMV_H_ +#define CLBLAST_ROUTINES_XTBMV_H_ + +#include "internal/routines/level2/xgemv.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class Xtbmv: public Xgemv<T> { + public: + + // Members from the base class + using Routine<T>::queue_; + using Routine<T>::context_; + + // Uses the generic matrix-vector routine + using Xgemv<T>::MatVec; + + // Constructor + Xtbmv(Queue &queue, Event &event, const std::string &name = "TBMV"); + + // Templated-precision implementation of the routine + StatusCode DoTbmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, const size_t k, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XTBMV_H_ +#endif diff --git a/include/internal/routines/level2/xtpmv.h b/include/internal/routines/level2/xtpmv.h new file mode 100644 index 00000000..183d3505 --- /dev/null +++ b/include/internal/routines/level2/xtpmv.h @@ -0,0 +1,51 @@ + +// ================================================================================================= +// 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 Xtpmv routine. It is based on the generalized mat-vec multiplication +// routine (Xgemv). The Xtpmv class inherits from the templated class Xgemv, allowing it to call the +// "MatVec" function directly. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XTPMV_H_ +#define CLBLAST_ROUTINES_XTPMV_H_ + +#include "internal/routines/level2/xgemv.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class Xtpmv: public Xgemv<T> { + public: + + // Members from the base class + using Routine<T>::queue_; + using Routine<T>::context_; + + // Uses the generic matrix-vector routine + using Xgemv<T>::MatVec; + + // Constructor + Xtpmv(Queue &queue, Event &event, const std::string &name = "TPMV"); + + // Templated-precision implementation of the routine + StatusCode DoTpmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const Buffer<T> &ap_buffer, const size_t ap_offset, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XTPMV_H_ +#endif diff --git a/include/internal/routines/level2/xtrmv.h b/include/internal/routines/level2/xtrmv.h new file mode 100644 index 00000000..dadfbc98 --- /dev/null +++ b/include/internal/routines/level2/xtrmv.h @@ -0,0 +1,51 @@ + +// ================================================================================================= +// 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 Xtrmv routine. It is based on the generalized mat-vec multiplication +// routine (Xgemv). The Xtrmv class inherits from the templated class Xgemv, allowing it to call the +// "MatVec" function directly. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XTRMV_H_ +#define CLBLAST_ROUTINES_XTRMV_H_ + +#include "internal/routines/level2/xgemv.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class Xtrmv: public Xgemv<T> { + public: + + // Members from the base class + using Routine<T>::queue_; + using Routine<T>::context_; + + // Uses the generic matrix-vector routine + using Xgemv<T>::MatVec; + + // Constructor + Xtrmv(Queue &queue, Event &event, const std::string &name = "TRMV"); + + // Templated-precision implementation of the routine + StatusCode DoTrmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XTRMV_H_ +#endif diff --git a/scripts/generator/generator.py b/scripts/generator/generator.py index 338c8468..25f02861 100644 --- a/scripts/generator/generator.py +++ b/scripts/generator/generator.py @@ -71,9 +71,9 @@ routines = [ Routine(True, "2a", "symv", T, [S,D], ["n"], ["layout","triangle"], ["a","x"], ["y"], ["alpha","beta"], False, "Symmetric matrix-vector multiplication"), Routine(True, "2a", "sbmv", T, [S,D], ["n","k"], ["layout","triangle"], ["a","x"], ["y"], ["alpha","beta"], False, "Symmetric banded matrix-vector multiplication"), Routine(True, "2a", "spmv", T, [S,D], ["n"], ["layout","triangle"], ["ap","x"], ["y"], ["alpha","beta"], False, "Symmetric packed matrix-vector multiplication"), - Routine(False, "2a", "trmv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], True, "Triangular matrix-vector multiplication"), - Routine(False, "2a", "tbmv", T, [S,D,C,Z], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], True, "Triangular banded matrix-vector multiplication"), - Routine(False, "2a", "tpmv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [], True, "Triangular packed matrix-vector multiplication"), + Routine(True, "2a", "trmv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], True, "Triangular matrix-vector multiplication"), + Routine(True, "2a", "tbmv", T, [S,D,C,Z], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], True, "Triangular banded matrix-vector multiplication"), + Routine(True, "2a", "tpmv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [], True, "Triangular packed matrix-vector multiplication"), Routine(False, "2a", "trsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], False, "Solves a triangular system of equations"), Routine(False, "2a", "tbsv", T, [S,D,C,Z], ["n","k"], ["layout","triangle","a_transpose","diagonal"], ["a"], ["x"], [], False, "Solves a banded triangular system of equations"), Routine(False, "2a", "tpsv", T, [S,D,C,Z], ["n"], ["layout","triangle","a_transpose","diagonal"], ["ap"], ["x"], [], False, "Solves a packed triangular system of equations"), @@ -213,7 +213,7 @@ def wrapper_clblas(routines): if routine.scratch: result += " auto queue = Queue(queues[0]);\n" result += " auto context = queue.GetContext();\n" - result += " auto scratch_buffer = Buffer<"+flavour.template+">(context, n);\n" + result += " auto scratch_buffer = Buffer<"+flavour.template+">(context, n*x_inc + x_offset);\n" arguments += ["scratch_buffer()"] result += " return clblas"+flavour.name+routine.name+"(" result += (",\n"+indent).join([a for a in arguments]) @@ -237,7 +237,7 @@ files = [ path_clblast+"/src/clblast_c.cc", path_clblast+"/test/wrapper_clblas.h", ] -header_lines = [84, 49, 80, 24, 22] +header_lines = [84, 52, 80, 24, 22] footer_lines = [6, 3, 5, 2, 6] # Checks whether the command-line arguments are valid; exists otherwise diff --git a/src/clblast.cc b/src/clblast.cc index f7baf5e8..77999aaf 100644 --- a/src/clblast.cc +++ b/src/clblast.cc @@ -35,6 +35,9 @@ #include "internal/routines/level2/xsymv.h" #include "internal/routines/level2/xsbmv.h" #include "internal/routines/level2/xspmv.h" +#include "internal/routines/level2/xtrmv.h" +#include "internal/routines/level2/xtbmv.h" +#include "internal/routines/level2/xtpmv.h" // BLAS level-3 includes #include "internal/routines/level3/xgemm.h" @@ -628,12 +631,20 @@ template StatusCode Spmv<double>(const Layout, const Triangle, // Triangular matrix-vector multiplication: STRMV/DTRMV/CTRMV/ZTRMV template <typename T> -StatusCode Trmv(const Layout, const Triangle, const Transpose, const Diagonal, - const size_t, - const cl_mem, const size_t, const size_t, - cl_mem, const size_t, const size_t, - cl_command_queue*, cl_event*) { - return StatusCode::kNotImplemented; +StatusCode Trmv(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + cl_mem x_buffer, const size_t x_offset, const size_t x_inc, + cl_command_queue* queue, cl_event* event) { + auto queue_cpp = Queue(*queue); + auto event_cpp = Event(*event); + auto routine = Xtrmv<T>(queue_cpp, event_cpp); + auto status = routine.SetUp(); + if (status != StatusCode::kSuccess) { return status; } + return routine.DoTrmv(layout, triangle, a_transpose, diagonal, + n, + Buffer<T>(a_buffer), a_offset, a_ld, + Buffer<T>(x_buffer), x_offset, x_inc); } template StatusCode Trmv<float>(const Layout, const Triangle, const Transpose, const Diagonal, const size_t, @@ -658,12 +669,20 @@ template StatusCode Trmv<double2>(const Layout, const Triangle, const Transpose, // Triangular banded matrix-vector multiplication: STBMV/DTBMV/CTBMV/ZTBMV template <typename T> -StatusCode Tbmv(const Layout, const Triangle, const Transpose, const Diagonal, - const size_t, const size_t, - const cl_mem, const size_t, const size_t, - cl_mem, const size_t, const size_t, - cl_command_queue*, cl_event*) { - return StatusCode::kNotImplemented; +StatusCode Tbmv(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Diagonal diagonal, + const size_t n, const size_t k, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + cl_mem x_buffer, const size_t x_offset, const size_t x_inc, + cl_command_queue* queue, cl_event* event) { + auto queue_cpp = Queue(*queue); + auto event_cpp = Event(*event); + auto routine = Xtbmv<T>(queue_cpp, event_cpp); + auto status = routine.SetUp(); + if (status != StatusCode::kSuccess) { return status; } + return routine.DoTbmv(layout, triangle, a_transpose, diagonal, + n, k, + Buffer<T>(a_buffer), a_offset, a_ld, + Buffer<T>(x_buffer), x_offset, x_inc); } template StatusCode Tbmv<float>(const Layout, const Triangle, const Transpose, const Diagonal, const size_t, const size_t, @@ -688,12 +707,20 @@ template StatusCode Tbmv<double2>(const Layout, const Triangle, const Transpose, // Triangular packed matrix-vector multiplication: STPMV/DTPMV/CTPMV/ZTPMV template <typename T> -StatusCode Tpmv(const Layout, const Triangle, const Transpose, const Diagonal, - const size_t, - const cl_mem, const size_t, - cl_mem, const size_t, const size_t, - cl_command_queue*, cl_event*) { - return StatusCode::kNotImplemented; +StatusCode Tpmv(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const cl_mem ap_buffer, const size_t ap_offset, + cl_mem x_buffer, const size_t x_offset, const size_t x_inc, + cl_command_queue* queue, cl_event* event) { + auto queue_cpp = Queue(*queue); + auto event_cpp = Event(*event); + auto routine = Xtpmv<T>(queue_cpp, event_cpp); + auto status = routine.SetUp(); + if (status != StatusCode::kSuccess) { return status; } + return routine.DoTpmv(layout, triangle, a_transpose, diagonal, + n, + Buffer<T>(ap_buffer), ap_offset, + Buffer<T>(x_buffer), x_offset, x_inc); } template StatusCode Tpmv<float>(const Layout, const Triangle, const Transpose, const Diagonal, const size_t, diff --git a/src/kernels/level2/xgemv.opencl b/src/kernels/level2/xgemv.opencl index ab7802e5..8ed0e9e4 100644 --- a/src/kernels/level2/xgemv.opencl +++ b/src/kernels/level2/xgemv.opencl @@ -107,6 +107,19 @@ inline real LoadMatrixA(const __global real* restrict agm, const int x, const in #endif } + // For triangular matrices + #elif defined(ROUTINE_TRMV) + if (((parameter == 0 || parameter == 2) && y <= x) || + ((parameter == 1 || parameter == 3) && x <= y)) { + result = agm[a_ld*y + x + a_offset]; + if (parameter >= 2 && y == x) { + SetToOne(result); + } + } + else { + SetToZero(result); + } + // For symmetric/hermitian banded matrices #elif defined(ROUTINE_HBMV) || defined(ROUTINE_SBMV) if (parameter == 1) { @@ -146,6 +159,35 @@ inline real LoadMatrixA(const __global real* restrict agm, const int x, const in } } + // For triangular banded matrices + #elif defined(ROUTINE_TBMV) + if (parameter == 1 || parameter == 3) { + if (x <= y) { + const int m = kl - y; + if (x >= y-kl && x <= y) { result = agm[a_ld*y + m + x + a_offset]; } + else { SetToZero(result); } + if (parameter >= 2 && y == x) { + SetToOne(result); + } + } + else { + SetToZero(result); + } + } + else { + if (x >= y) { + const int m = -y; + if (x >= y && x < y+kl+1) { result = agm[a_ld*y + m + x + a_offset]; } + else { SetToZero(result); } + if (parameter >= 2 && y == x) { + SetToOne(result); + } + } + else { + SetToZero(result); + } + } + // For symmetric/hermitian packed matrices #elif defined(ROUTINE_HPMV) || defined(ROUTINE_SPMV) if (parameter == 1) { @@ -177,6 +219,31 @@ inline real LoadMatrixA(const __global real* restrict agm, const int x, const in } } + // For triangular packed matrices + #elif defined(ROUTINE_TPMV) + if (parameter == 1 || parameter == 3) { + if (x <= y) { + result = agm[((y+1)*y)/2 + x + a_offset]; + if (parameter >= 2 && y == x) { + SetToOne(result); + } + } + else { + SetToZero(result); + } + } + else { + if (x >= y) { + result = agm[((2*a_ld-(y+1))*y)/2 + x + a_offset]; + if (parameter >= 2 && y == x) { + SetToOne(result); + } + } + else { + SetToZero(result); + } + } + // For general matrices #else result = agm[a_ld*y + x + a_offset]; diff --git a/src/routines/level2/xtbmv.cc b/src/routines/level2/xtbmv.cc new file mode 100644 index 00000000..2e1aebff --- /dev/null +++ b/src/routines/level2/xtbmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 Xtbmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtbmv.h" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xtbmv<T>::Xtbmv(Queue &queue, Event &event, const std::string &name): + Xgemv<T>(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template <typename T> +StatusCode Xtbmv<T>::DoTbmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, const size_t k, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer<T>(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular banded matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TBMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast<T>(1), + a_buffer, a_offset, a_ld, + scratch_buffer, x_offset, x_inc, static_cast<T>(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, false, k, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtbmv<float>; +template class Xtbmv<double>; +template class Xtbmv<float2>; +template class Xtbmv<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level2/xtpmv.cc b/src/routines/level2/xtpmv.cc new file mode 100644 index 00000000..aa0e099b --- /dev/null +++ b/src/routines/level2/xtpmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 Xtpmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtpmv.h" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xtpmv<T>::Xtpmv(Queue &queue, Event &event, const std::string &name): + Xgemv<T>(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template <typename T> +StatusCode Xtpmv<T>::DoTpmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const Buffer<T> &ap_buffer, const size_t ap_offset, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer<T>(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular packed matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TPMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast<T>(1), + ap_buffer, ap_offset, n, + scratch_buffer, x_offset, x_inc, static_cast<T>(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, true, 0, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtpmv<float>; +template class Xtpmv<double>; +template class Xtpmv<float2>; +template class Xtpmv<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level2/xtrmv.cc b/src/routines/level2/xtrmv.cc new file mode 100644 index 00000000..94424743 --- /dev/null +++ b/src/routines/level2/xtrmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 Xtrmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtrmv.h" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xtrmv<T>::Xtrmv(Queue &queue, Event &event, const std::string &name): + Xgemv<T>(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template <typename T> +StatusCode Xtrmv<T>::DoTrmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer<T>(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TRMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast<T>(1), + a_buffer, a_offset, a_ld, + scratch_buffer, x_offset, x_inc, static_cast<T>(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, false, 0, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtrmv<float>; +template class Xtrmv<double>; +template class Xtrmv<float2>; +template class Xtrmv<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/test/routines/level2/xtbmv.h b/test/routines/level2/xtbmv.h new file mode 100644 index 00000000..dbdddb65 --- /dev/null +++ b/test/routines/level2/xtbmv.h @@ -0,0 +1,125 @@ + +// ================================================================================================= +// 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 Xtbmv 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_XTBMV_H_ +#define CLBLAST_TEST_ROUTINES_XTBMV_H_ + +#include <vector> +#include <string> + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXtbmv { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgN, kArgKL, + kArgLayout, kArgTriangle, kArgATransp, kArgDiagonal, + kArgALeadDim, kArgXInc, + kArgAOffset, kArgXOffset}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments<T> &args) { + return args.n * args.x_inc + args.x_offset; + } + static size_t GetSizeA(const Arguments<T> &args) { + return args.n * args.a_ld + args.a_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args) { + args.a_size = GetSizeA(args); + args.x_size = GetSizeX(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDB(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine + + // 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 &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Tbmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, + args.n, args.kl, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXtbmv<T>(static_cast<clblasOrder>(args.layout), + static_cast<clblasUplo>(args.triangle), + static_cast<clblasTranspose>(args.a_transpose), + static_cast<clblasDiag>(args.diagonal), + args.n, args.kl, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast<StatusCode>(status); + } + + // 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.x_size, static_cast<T>(0)); + buffers.x_vec.Read(queue, args.x_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { + return args.n; + } + static size_t ResultID2(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t) { + return id1*args.x_inc + args.x_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return 2 * args.n * args.n; + } + static size_t GetBytes(const Arguments<T> &args) { + return ((args.kl+args.kl+1)*args.n + 2*args.n + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XTBMV_H_ +#endif diff --git a/test/routines/level2/xtpmv.h b/test/routines/level2/xtpmv.h new file mode 100644 index 00000000..4425765e --- /dev/null +++ b/test/routines/level2/xtpmv.h @@ -0,0 +1,125 @@ + +// ================================================================================================= +// 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 Xtpmv 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_XTPMV_H_ +#define CLBLAST_TEST_ROUTINES_XTPMV_H_ + +#include <vector> +#include <string> + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXtpmv { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgN, + kArgLayout, kArgTriangle, kArgATransp, kArgDiagonal, + kArgXInc, + kArgAPOffset, kArgXOffset}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments<T> &args) { + return args.n * args.x_inc + args.x_offset; + } + static size_t GetSizeAP(const Arguments<T> &args) { + return ((args.n*(args.n+1)) / 2) + args.ap_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args) { + args.ap_size = GetSizeAP(args); + args.x_size = GetSizeX(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDB(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine + + // 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 &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Tpmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, + args.n, + buffers.ap_mat(), args.ap_offset, + buffers.x_vec(), args.x_offset, args.x_inc, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXtpmv<T>(static_cast<clblasOrder>(args.layout), + static_cast<clblasUplo>(args.triangle), + static_cast<clblasTranspose>(args.a_transpose), + static_cast<clblasDiag>(args.diagonal), + args.n, + buffers.ap_mat(), args.ap_offset, + buffers.x_vec(), args.x_offset, args.x_inc, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast<StatusCode>(status); + } + + // 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.x_size, static_cast<T>(0)); + buffers.x_vec.Read(queue, args.x_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { + return args.n; + } + static size_t ResultID2(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t) { + return id1*args.x_inc + args.x_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return 2 * args.n * args.n; + } + static size_t GetBytes(const Arguments<T> &args) { + return (((args.n*(args.n+1)) / 2) + 2*args.n + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XTPMV_H_ +#endif diff --git a/test/routines/level2/xtrmv.h b/test/routines/level2/xtrmv.h new file mode 100644 index 00000000..1c0c6fd8 --- /dev/null +++ b/test/routines/level2/xtrmv.h @@ -0,0 +1,125 @@ + +// ================================================================================================= +// 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 Xtrmv 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_XTRMV_H_ +#define CLBLAST_TEST_ROUTINES_XTRMV_H_ + +#include <vector> +#include <string> + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXtrmv { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgN, + kArgLayout, kArgTriangle, kArgATransp, kArgDiagonal, + kArgALeadDim, kArgXInc, + kArgAOffset, kArgXOffset}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments<T> &args) { + return args.n * args.x_inc + args.x_offset; + } + static size_t GetSizeA(const Arguments<T> &args) { + return args.n * args.a_ld + args.a_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args) { + args.a_size = GetSizeA(args); + args.x_size = GetSizeX(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDB(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine + + // 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 &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Trmv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, + args.n, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXtrmv<T>(static_cast<clblasOrder>(args.layout), + static_cast<clblasUplo>(args.triangle), + static_cast<clblasTranspose>(args.a_transpose), + static_cast<clblasDiag>(args.diagonal), + args.n, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast<StatusCode>(status); + } + + // 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.x_size, static_cast<T>(0)); + buffers.x_vec.Read(queue, args.x_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { + return args.n; + } + static size_t ResultID2(const Arguments<T> &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t) { + return id1*args.x_inc + args.x_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return 2 * args.n * args.n; + } + static size_t GetBytes(const Arguments<T> &args) { + return (args.n*args.n + 2*args.n + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XTRMV_H_ +#endif diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h index 10c7dd47..23a02a45 100644 --- a/test/wrapper_clblas.h +++ b/test/wrapper_clblas.h @@ -238,7 +238,7 @@ clblasStatus clblasXdot<float>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float>(context, n); + auto scratch_buffer = Buffer<float>(context, n*x_inc + x_offset); return clblasSdot(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -255,7 +255,7 @@ clblasStatus clblasXdot<double>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double>(context, n); + auto scratch_buffer = Buffer<double>(context, n*x_inc + x_offset); return clblasDdot(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -281,7 +281,7 @@ clblasStatus clblasXdotu<float2>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float2>(context, n); + auto scratch_buffer = Buffer<float2>(context, n*x_inc + x_offset); return clblasCdotu(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -298,7 +298,7 @@ clblasStatus clblasXdotu<double2>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double2>(context, n); + auto scratch_buffer = Buffer<double2>(context, n*x_inc + x_offset); return clblasZdotu(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -324,7 +324,7 @@ clblasStatus clblasXdotc<float2>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float2>(context, n); + auto scratch_buffer = Buffer<float2>(context, n*x_inc + x_offset); return clblasCdotc(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -341,7 +341,7 @@ clblasStatus clblasXdotc<double2>(const size_t n, cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double2>(context, n); + auto scratch_buffer = Buffer<double2>(context, n*x_inc + x_offset); return clblasZdotc(n, dot_buffer, dot_offset, x_buffer, x_offset, static_cast<int>(x_inc), @@ -747,7 +747,7 @@ clblasStatus clblasXtrmv<float>(const clblasOrder layout, const clblasUplo trian cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float>(context, n); + auto scratch_buffer = Buffer<float>(context, n*x_inc + x_offset); return clblasStrmv(layout, triangle, a_transpose, diagonal, n, a_buffer, a_offset, a_ld, @@ -764,7 +764,7 @@ clblasStatus clblasXtrmv<double>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double>(context, n); + auto scratch_buffer = Buffer<double>(context, n*x_inc + x_offset); return clblasDtrmv(layout, triangle, a_transpose, diagonal, n, a_buffer, a_offset, a_ld, @@ -781,7 +781,7 @@ clblasStatus clblasXtrmv<float2>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float2>(context, n); + auto scratch_buffer = Buffer<float2>(context, n*x_inc + x_offset); return clblasCtrmv(layout, triangle, a_transpose, diagonal, n, a_buffer, a_offset, a_ld, @@ -798,7 +798,7 @@ clblasStatus clblasXtrmv<double2>(const clblasOrder layout, const clblasUplo tri cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double2>(context, n); + auto scratch_buffer = Buffer<double2>(context, n*x_inc + x_offset); return clblasZtrmv(layout, triangle, a_transpose, diagonal, n, a_buffer, a_offset, a_ld, @@ -824,7 +824,7 @@ clblasStatus clblasXtbmv<float>(const clblasOrder layout, const clblasUplo trian cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float>(context, n); + auto scratch_buffer = Buffer<float>(context, n*x_inc + x_offset); return clblasStbmv(layout, triangle, a_transpose, diagonal, n, k, a_buffer, a_offset, a_ld, @@ -841,7 +841,7 @@ clblasStatus clblasXtbmv<double>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double>(context, n); + auto scratch_buffer = Buffer<double>(context, n*x_inc + x_offset); return clblasDtbmv(layout, triangle, a_transpose, diagonal, n, k, a_buffer, a_offset, a_ld, @@ -858,7 +858,7 @@ clblasStatus clblasXtbmv<float2>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float2>(context, n); + auto scratch_buffer = Buffer<float2>(context, n*x_inc + x_offset); return clblasCtbmv(layout, triangle, a_transpose, diagonal, n, k, a_buffer, a_offset, a_ld, @@ -875,7 +875,7 @@ clblasStatus clblasXtbmv<double2>(const clblasOrder layout, const clblasUplo tri cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double2>(context, n); + auto scratch_buffer = Buffer<double2>(context, n*x_inc + x_offset); return clblasZtbmv(layout, triangle, a_transpose, diagonal, n, k, a_buffer, a_offset, a_ld, @@ -901,7 +901,7 @@ clblasStatus clblasXtpmv<float>(const clblasOrder layout, const clblasUplo trian cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float>(context, n); + auto scratch_buffer = Buffer<float>(context, n*x_inc + x_offset); return clblasStpmv(layout, triangle, a_transpose, diagonal, n, ap_buffer, ap_offset, @@ -918,7 +918,7 @@ clblasStatus clblasXtpmv<double>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double>(context, n); + auto scratch_buffer = Buffer<double>(context, n*x_inc + x_offset); return clblasDtpmv(layout, triangle, a_transpose, diagonal, n, ap_buffer, ap_offset, @@ -935,7 +935,7 @@ clblasStatus clblasXtpmv<float2>(const clblasOrder layout, const clblasUplo tria cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<float2>(context, n); + auto scratch_buffer = Buffer<float2>(context, n*x_inc + x_offset); return clblasCtpmv(layout, triangle, a_transpose, diagonal, n, ap_buffer, ap_offset, @@ -952,7 +952,7 @@ clblasStatus clblasXtpmv<double2>(const clblasOrder layout, const clblasUplo tri cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { auto queue = Queue(queues[0]); auto context = queue.GetContext(); - auto scratch_buffer = Buffer<double2>(context, n); + auto scratch_buffer = Buffer<double2>(context, n*x_inc + x_offset); return clblasZtpmv(layout, triangle, a_transpose, diagonal, n, ap_buffer, ap_offset, |