From 844e68853e14e5cf5ba67a8b376bd59d27383290 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sat, 12 Aug 2017 15:38:17 +0200 Subject: Moved some utility functions to a test-specific utility compilation-unit --- CMakeLists.txt | 6 ++- src/utilities/utilities.cpp | 97 +------------------------------------ src/utilities/utilities.hpp | 67 ++------------------------ test/correctness/tester.hpp | 2 +- test/performance/client.hpp | 2 +- test/routines/common.hpp | 1 + test/test_utilities.cpp | 114 ++++++++++++++++++++++++++++++++++++++++++++ test/test_utilities.hpp | 99 ++++++++++++++++++++++++++++++++++++++ 8 files changed, 224 insertions(+), 164 deletions(-) create mode 100644 test/test_utilities.cpp create mode 100644 test/test_utilities.hpp diff --git a/CMakeLists.txt b/CMakeLists.txt index ae259fad..9abdadf3 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -380,7 +380,8 @@ endif() # Section for the performance tests (i.e. the client). These compare against optionally a reference # library, either clBLAS, a CPU BLAS, or CUDA's cuBLAS. if(CLIENTS) - set(CLIENTS_COMMON ${WRAPPERS} test/performance/client.hpp test/routines/common.hpp) + set(CLIENTS_COMMON ${WRAPPERS} test/test_utilities.cpp test/test_utilities.hpp + test/performance/client.hpp test/routines/common.hpp) # Visual Studio requires the sources of non-exported objects/libraries if(MSVC) @@ -431,7 +432,8 @@ endif() # CPU BLAS library, and/or cuBLAS to act as a reference. if(TESTS) enable_testing() - set(TESTS_COMMON ${WRAPPERS} test/correctness/testblas.hpp test/correctness/tester.hpp + set(TESTS_COMMON ${WRAPPERS} test/test_utilities.cpp test/test_utilities.hpp + test/correctness/testblas.hpp test/correctness/tester.hpp test/routines/common.hpp) # Visual Studio requires the sources of non-exported objects/libraries diff --git a/src/utilities/utilities.cpp b/src/utilities/utilities.cpp index 95b70cd5..0cd00438 100644 --- a/src/utilities/utilities.cpp +++ b/src/utilities/utilities.cpp @@ -7,7 +7,7 @@ // Author(s): // Cedric Nugteren // -// This file implements the common (test) utility functions. +// This file implements the common utility functions. // // ================================================================================================= @@ -85,14 +85,6 @@ template <> double AbsoluteValue(const double2 value) { return std::sqrt(value.real() * value.real() + value.imag() * value.imag()); } -// Returns whether a scalar is close to zero -template bool IsCloseToZero(const T value) { return (value > -SmallConstant()) && (value < SmallConstant()); } -template bool IsCloseToZero(const float); -template bool IsCloseToZero(const double); -template <> bool IsCloseToZero(const half value) { return IsCloseToZero(HalfToFloat(value)); } -template <> bool IsCloseToZero(const float2 value) { return IsCloseToZero(value.real()) || IsCloseToZero(value.imag()); } -template <> bool IsCloseToZero(const double2 value) { return IsCloseToZero(value.real()) || IsCloseToZero(value.imag()); } - // ================================================================================================= // Implements the string conversion using std::to_string if possible @@ -319,12 +311,6 @@ bool CheckArgument(const std::vector &arguments, std::string &help, // ================================================================================================= -// Returns a random seed. This used to be implemented using 'std::random_device', but that doesn't -// always work. The chrono-timers are more reliable in that sense, but perhaps less random. -unsigned int GetRandomSeed() { - return static_cast(std::chrono::system_clock::now().time_since_epoch().count()); -} - // Create a random number generator and populates a vector with samples from a random distribution template void PopulateVector(std::vector &vector, std::mt19937 &mt, std::uniform_real_distribution &dist) { @@ -354,87 +340,6 @@ void PopulateVector(std::vector &vector, std::mt19937 &mt, std::uniform_re // ================================================================================================= -template -void DeviceToHost(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, - Queue &queue, const std::vector &names) { - for (auto &name: names) { - if (name == kBufVecX) {buffers_host.x_vec = std::vector(args.x_size, static_cast(0)); buffers.x_vec.Read(queue, args.x_size, buffers_host.x_vec); } - else if (name == kBufVecY) { buffers_host.y_vec = std::vector(args.y_size, static_cast(0)); buffers.y_vec.Read(queue, args.y_size, buffers_host.y_vec); } - else if (name == kBufMatA) { buffers_host.a_mat = std::vector(args.a_size, static_cast(0)); buffers.a_mat.Read(queue, args.a_size, buffers_host.a_mat); } - else if (name == kBufMatB) { buffers_host.b_mat = std::vector(args.b_size, static_cast(0)); buffers.b_mat.Read(queue, args.b_size, buffers_host.b_mat); } - else if (name == kBufMatC) { buffers_host.c_mat = std::vector(args.c_size, static_cast(0)); buffers.c_mat.Read(queue, args.c_size, buffers_host.c_mat); } - else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector(args.ap_size, static_cast(0)); buffers.ap_mat.Read(queue, args.ap_size, buffers_host.ap_mat); } - else if (name == kBufScalar) { buffers_host.scalar = std::vector(args.scalar_size, static_cast(0)); buffers.scalar.Read(queue, args.scalar_size, buffers_host.scalar); } - else { throw std::runtime_error("Invalid buffer name"); } - } -} - -template -void HostToDevice(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, - Queue &queue, const std::vector &names) { - for (auto &name: names) { - if (name == kBufVecX) { buffers.x_vec.Write(queue, args.x_size, buffers_host.x_vec); } - else if (name == kBufVecY) { buffers.y_vec.Write(queue, args.y_size, buffers_host.y_vec); } - else if (name == kBufMatA) { buffers.a_mat.Write(queue, args.a_size, buffers_host.a_mat); } - else if (name == kBufMatB) { buffers.b_mat.Write(queue, args.b_size, buffers_host.b_mat); } - else if (name == kBufMatC) { buffers.c_mat.Write(queue, args.c_size, buffers_host.c_mat); } - else if (name == kBufMatAP) { buffers.ap_mat.Write(queue, args.ap_size, buffers_host.ap_mat); } - else if (name == kBufScalar) { buffers.scalar.Write(queue, args.scalar_size, buffers_host.scalar); } - else { throw std::runtime_error("Invalid buffer name"); } - } -} - -// Compiles the above functions -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); -template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); - -// ================================================================================================= - -// Conversion between half and single-precision -std::vector HalfToFloatBuffer(const std::vector& source) { - auto result = std::vector(source.size()); - for (auto i = size_t(0); i < source.size(); ++i) { result[i] = HalfToFloat(source[i]); } - return result; -} -void FloatToHalfBuffer(std::vector& result, const std::vector& source) { - for (auto i = size_t(0); i < source.size(); ++i) { result[i] = FloatToHalf(source[i]); } -} - -// As above, but now for OpenCL data-types instead of std::vectors -Buffer HalfToFloatBuffer(const Buffer& source, cl_command_queue queue_raw) { - const auto size = source.GetSize() / sizeof(half); - auto queue = Queue(queue_raw); - auto context = queue.GetContext(); - auto source_cpu = std::vector(size); - source.Read(queue, size, source_cpu); - auto result_cpu = HalfToFloatBuffer(source_cpu); - auto result = Buffer(context, size); - result.Write(queue, size, result_cpu); - return result; -} -void FloatToHalfBuffer(Buffer& result, const Buffer& source, cl_command_queue queue_raw) { - const auto size = source.GetSize() / sizeof(float); - auto queue = Queue(queue_raw); - auto context = queue.GetContext(); - auto source_cpu = std::vector(size); - source.Read(queue, size, source_cpu); - auto result_cpu = std::vector(size); - FloatToHalfBuffer(result_cpu, source_cpu); - result.Write(queue, size, result_cpu); -} - // Converts a 'real' value to a 'real argument' value to be passed to a kernel. Normally there is // no conversion, but half-precision is not supported as kernel argument so it is converted to float. template <> typename RealArg::Type GetRealArg(const half value) { return HalfToFloat(value); } diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp index 03051354..72997d7f 100644 --- a/src/utilities/utilities.hpp +++ b/src/utilities/utilities.hpp @@ -7,10 +7,9 @@ // Author(s): // Cedric Nugteren // -// This file provides declarations for the common (test) utility functions such as a command-line +// This file provides declarations for the common utility functions such as a command-line // argument parser. On top of this, it serves as the 'common' header, including the C++ OpenCL -// wrapper. These utilities are not only used for CLBlast, but also included as part of the tuners, -// the performance client and the correctness testers. +// wrapper. // // ================================================================================================= @@ -81,19 +80,6 @@ constexpr auto kArgBatchCount = "batch_num"; // The tuner-specific arguments in string form constexpr auto kArgFraction = "fraction"; -// The client-specific arguments in string form -constexpr auto kArgCompareclblas = "clblas"; -constexpr auto kArgComparecblas = "cblas"; -constexpr auto kArgComparecublas = "cublas"; -constexpr auto kArgStepSize = "step"; -constexpr auto kArgNumSteps = "num_steps"; -constexpr auto kArgNumRuns = "runs"; -constexpr auto kArgWarmUp = "warm_up"; - -// The test-specific arguments in string form -constexpr auto kArgFullTest = "full_test"; -constexpr auto kArgVerbose = "verbose"; - // The common arguments in string form constexpr auto kArgPlatform = "platform"; constexpr auto kArgDevice = "device"; @@ -101,6 +87,7 @@ constexpr auto kArgPrecision = "precision"; constexpr auto kArgHelp = "h"; constexpr auto kArgQuiet = "q"; constexpr auto kArgNoAbbreviations = "no_abbrv"; +constexpr auto kArgNumRuns = "runs"; // The buffer names constexpr auto kBufVecX = "X"; @@ -133,9 +120,6 @@ template T SmallConstant(); // Returns the absolute value of a scalar (modulus in case of complex numbers) template typename BaseType::Type AbsoluteValue(const T value); -// Returns whether a scalar is close to zero -template bool IsCloseToZero(const T value); - // ================================================================================================= // Structure containing all possible arguments for test clients, including their default values @@ -208,28 +192,6 @@ struct Arguments { bool no_abbrv = false; }; -// Structure containing all possible buffers for test clients -template -struct Buffers { - Buffer x_vec; - Buffer y_vec; - Buffer a_mat; - Buffer b_mat; - Buffer c_mat; - Buffer ap_mat; - Buffer scalar; -}; -template -struct BuffersHost { - std::vector x_vec; - std::vector y_vec; - std::vector a_mat; - std::vector b_mat; - std::vector c_mat; - std::vector ap_mat; - std::vector scalar; -}; - // ================================================================================================= // Converts a value (e.g. an integer) to a string. This also covers special cases for CLBlast @@ -264,9 +226,6 @@ bool CheckArgument(const std::vector &arguments, std::string &help, // ================================================================================================= -// Returns a random number to be used as a seed -unsigned int GetRandomSeed(); - // Test/example data lower and upper limit constexpr auto kTestDataLowerLimit = -2.0; constexpr auto kTestDataUpperLimit = 2.0; @@ -277,26 +236,6 @@ void PopulateVector(std::vector &vector, std::mt19937 &mt, std::uniform_real_ // ================================================================================================= -// Copies buffers from the OpenCL device to the host -template -void DeviceToHost(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, - Queue &queue, const std::vector &names); - -// Copies buffers from the host to the OpenCL device -template -void HostToDevice(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, - Queue &queue, const std::vector &names); - -// ================================================================================================= - -// Conversion between half and single-precision -std::vector HalfToFloatBuffer(const std::vector& source); -void FloatToHalfBuffer(std::vector& result, const std::vector& source); - -// As above, but now for OpenCL data-types instead of std::vectors -Buffer HalfToFloatBuffer(const Buffer& source, cl_command_queue queue_raw); -void FloatToHalfBuffer(Buffer& result, const Buffer& source, cl_command_queue queue_raw); - // Converts a 'real' value to a 'real argument' value to be passed to a kernel. Normally there is // no conversion, but half-precision is not supported as kernel argument so it is converted to float. template struct RealArg { using Type = T; }; diff --git a/test/correctness/tester.hpp b/test/correctness/tester.hpp index dfbefdfd..caf03787 100644 --- a/test/correctness/tester.hpp +++ b/test/correctness/tester.hpp @@ -22,7 +22,7 @@ #include #include -#include "utilities/utilities.hpp" +#include "test/test_utilities.hpp" // The libraries #ifdef CLBLAST_REF_CLBLAS diff --git a/test/performance/client.hpp b/test/performance/client.hpp index 47a13017..2ba09cb9 100644 --- a/test/performance/client.hpp +++ b/test/performance/client.hpp @@ -25,7 +25,7 @@ #include #include -#include "utilities/utilities.hpp" +#include "test/test_utilities.hpp" // The libraries to test #ifdef CLBLAST_REF_CLBLAS diff --git a/test/routines/common.hpp b/test/routines/common.hpp index 9708288a..47c8f8d7 100644 --- a/test/routines/common.hpp +++ b/test/routines/common.hpp @@ -18,6 +18,7 @@ #include #include "utilities/utilities.hpp" +#include "test/test_utilities.hpp" #ifdef CLBLAST_REF_CLBLAS #include "test/wrapper_clblas.hpp" diff --git a/test/test_utilities.cpp b/test/test_utilities.cpp new file mode 100644 index 00000000..b8fd94a9 --- /dev/null +++ b/test/test_utilities.cpp @@ -0,0 +1,114 @@ + +// ================================================================================================= +// 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 +// +// This file implements the test utility functions. +// +// ================================================================================================= + +#include "test/test_utilities.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Returns whether a scalar is close to zero +template bool IsCloseToZero(const T value) { return (value > -SmallConstant()) && (value < SmallConstant()); } +template bool IsCloseToZero(const float); +template bool IsCloseToZero(const double); +template <> bool IsCloseToZero(const half value) { return IsCloseToZero(HalfToFloat(value)); } +template <> bool IsCloseToZero(const float2 value) { return IsCloseToZero(value.real()) || IsCloseToZero(value.imag()); } +template <> bool IsCloseToZero(const double2 value) { return IsCloseToZero(value.real()) || IsCloseToZero(value.imag()); } + +// ================================================================================================= + +template +void DeviceToHost(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, + Queue &queue, const std::vector &names) { + for (auto &name: names) { + if (name == kBufVecX) {buffers_host.x_vec = std::vector(args.x_size, static_cast(0)); buffers.x_vec.Read(queue, args.x_size, buffers_host.x_vec); } + else if (name == kBufVecY) { buffers_host.y_vec = std::vector(args.y_size, static_cast(0)); buffers.y_vec.Read(queue, args.y_size, buffers_host.y_vec); } + else if (name == kBufMatA) { buffers_host.a_mat = std::vector(args.a_size, static_cast(0)); buffers.a_mat.Read(queue, args.a_size, buffers_host.a_mat); } + else if (name == kBufMatB) { buffers_host.b_mat = std::vector(args.b_size, static_cast(0)); buffers.b_mat.Read(queue, args.b_size, buffers_host.b_mat); } + else if (name == kBufMatC) { buffers_host.c_mat = std::vector(args.c_size, static_cast(0)); buffers.c_mat.Read(queue, args.c_size, buffers_host.c_mat); } + else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector(args.ap_size, static_cast(0)); buffers.ap_mat.Read(queue, args.ap_size, buffers_host.ap_mat); } + else if (name == kBufScalar) { buffers_host.scalar = std::vector(args.scalar_size, static_cast(0)); buffers.scalar.Read(queue, args.scalar_size, buffers_host.scalar); } + else { throw std::runtime_error("Invalid buffer name"); } + } +} + +template +void HostToDevice(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, + Queue &queue, const std::vector &names) { + for (auto &name: names) { + if (name == kBufVecX) { buffers.x_vec.Write(queue, args.x_size, buffers_host.x_vec); } + else if (name == kBufVecY) { buffers.y_vec.Write(queue, args.y_size, buffers_host.y_vec); } + else if (name == kBufMatA) { buffers.a_mat.Write(queue, args.a_size, buffers_host.a_mat); } + else if (name == kBufMatB) { buffers.b_mat.Write(queue, args.b_size, buffers_host.b_mat); } + else if (name == kBufMatC) { buffers.c_mat.Write(queue, args.c_size, buffers_host.c_mat); } + else if (name == kBufMatAP) { buffers.ap_mat.Write(queue, args.ap_size, buffers_host.ap_mat); } + else if (name == kBufScalar) { buffers.scalar.Write(queue, args.scalar_size, buffers_host.scalar); } + else { throw std::runtime_error("Invalid buffer name"); } + } +} + +// Compiles the above functions +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void DeviceToHost(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); +template void HostToDevice(const Arguments&, Buffers&, BuffersHost&, Queue&, const std::vector&); + +// ================================================================================================= + +// Conversion between half and single-precision +std::vector HalfToFloatBuffer(const std::vector& source) { + auto result = std::vector(source.size()); + for (auto i = size_t(0); i < source.size(); ++i) { result[i] = HalfToFloat(source[i]); } + return result; +} +void FloatToHalfBuffer(std::vector& result, const std::vector& source) { + for (auto i = size_t(0); i < source.size(); ++i) { result[i] = FloatToHalf(source[i]); } +} + +// As above, but now for OpenCL data-types instead of std::vectors +Buffer HalfToFloatBuffer(const Buffer& source, cl_command_queue queue_raw) { + const auto size = source.GetSize() / sizeof(half); + auto queue = Queue(queue_raw); + auto context = queue.GetContext(); + auto source_cpu = std::vector(size); + source.Read(queue, size, source_cpu); + auto result_cpu = HalfToFloatBuffer(source_cpu); + auto result = Buffer(context, size); + result.Write(queue, size, result_cpu); + return result; +} +void FloatToHalfBuffer(Buffer& result, const Buffer& source, cl_command_queue queue_raw) { + const auto size = source.GetSize() / sizeof(float); + auto queue = Queue(queue_raw); + auto context = queue.GetContext(); + auto source_cpu = std::vector(size); + source.Read(queue, size, source_cpu); + auto result_cpu = std::vector(size); + FloatToHalfBuffer(result_cpu, source_cpu); + result.Write(queue, size, result_cpu); +} + +// ================================================================================================= +} // namespace clblast diff --git a/test/test_utilities.hpp b/test/test_utilities.hpp new file mode 100644 index 00000000..fc50a754 --- /dev/null +++ b/test/test_utilities.hpp @@ -0,0 +1,99 @@ + +// ================================================================================================= +// 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 +// +// This file provides declarations for the common test utility functions (performance clients and +// correctness testers). +// +// ================================================================================================= + +#ifndef CLBLAST_TEST_UTILITIES_H_ +#define CLBLAST_TEST_UTILITIES_H_ + +#include + +#include "utilities/utilities.hpp" + +namespace clblast { +// ================================================================================================= + +// The client-specific arguments in string form +constexpr auto kArgCompareclblas = "clblas"; +constexpr auto kArgComparecblas = "cblas"; +constexpr auto kArgComparecublas = "cublas"; +constexpr auto kArgStepSize = "step"; +constexpr auto kArgNumSteps = "num_steps"; +constexpr auto kArgWarmUp = "warm_up"; + +// The test-specific arguments in string form +constexpr auto kArgFullTest = "full_test"; +constexpr auto kArgVerbose = "verbose"; + +// ================================================================================================= + +// Returns whether a scalar is close to zero +template bool IsCloseToZero(const T value); + +// ================================================================================================= + +// Structure containing all possible buffers for test clients +template +struct Buffers { + Buffer x_vec; + Buffer y_vec; + Buffer a_mat; + Buffer b_mat; + Buffer c_mat; + Buffer ap_mat; + Buffer scalar; +}; +template +struct BuffersHost { + std::vector x_vec; + std::vector y_vec; + std::vector a_mat; + std::vector b_mat; + std::vector c_mat; + std::vector ap_mat; + std::vector scalar; +}; + +// ================================================================================================= + +// Converts a value (e.g. an integer) to a string. This also covers special cases for CLBlast +// data-types such as the Layout and Transpose data-types. +template +std::string ToString(T value); + +// ================================================================================================= + +// Copies buffers from the OpenCL device to the host +template +void DeviceToHost(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, + Queue &queue, const std::vector &names); + +// Copies buffers from the host to the OpenCL device +template +void HostToDevice(const Arguments &args, Buffers &buffers, BuffersHost &buffers_host, + Queue &queue, const std::vector &names); + +// ================================================================================================= + +// Conversion between half and single-precision +std::vector HalfToFloatBuffer(const std::vector& source); +void FloatToHalfBuffer(std::vector& result, const std::vector& source); + +// As above, but now for OpenCL data-types instead of std::vectors +Buffer HalfToFloatBuffer(const Buffer& source, cl_command_queue queue_raw); +void FloatToHalfBuffer(Buffer& result, const Buffer& source, cl_command_queue queue_raw); + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_UTILITIES_H_ +#endif -- cgit v1.2.3