diff options
-rw-r--r-- | CMakeLists.txt | 2 | ||||
-rw-r--r-- | doc/api.md | 70 | ||||
-rwxr-xr-x | scripts/generator/generator.py | 2 | ||||
-rw-r--r-- | src/clblast.cpp | 20 | ||||
-rw-r--r-- | src/clblast_cuda.cpp | 22 | ||||
-rw-r--r-- | src/routines/levelx/xconvgemm.cpp | 68 | ||||
-rw-r--r-- | src/routines/levelx/xconvgemm.hpp | 48 | ||||
-rw-r--r-- | src/routines/routines.hpp | 1 | ||||
-rw-r--r-- | src/utilities/utilities.hpp | 4 | ||||
-rw-r--r-- | test/correctness/routines/levelx/xconvgemm.cpp | 26 | ||||
-rw-r--r-- | test/correctness/testblas.hpp | 10 | ||||
-rw-r--r-- | test/performance/routines/levelx/xconvgemm.cpp | 33 | ||||
-rw-r--r-- | test/routines/levelx/xconvgemm.hpp | 219 |
13 files changed, 508 insertions, 17 deletions
diff --git a/CMakeLists.txt b/CMakeLists.txt index 0715b866..4974545e 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -204,7 +204,7 @@ 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 xhad xomatcopy xim2col xaxpybatched xgemmbatched xgemmstridedbatched) +set(LEVELX_ROUTINES xhad xomatcopy xim2col xconvgemm xaxpybatched xgemmbatched xgemmstridedbatched) set(ROUTINES ${LEVEL1_ROUTINES} ${LEVEL2_ROUTINES} ${LEVEL3_ROUTINES} ${LEVELX_ROUTINES}) set(PRECISIONS 32 64 3232 6464 16) @@ -3072,6 +3072,76 @@ Arguments to IM2COL: +xCONVGEMM: Batched convolution as GEMM (non-BLAS function) +------------- + +Integrates im2col and GEMM for batched 3D convolution, in which _im_ is the 4D input tensor, _kernel_ the 4D kernel weights tensor, and _result_ the 4D output tensor. + +C++ API: +``` +template <typename T> +StatusCode Convgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +``` + +C API: +``` +CLBlastStatusCode CLBlastSconvgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastDconvgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastCconvgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastZconvgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +CLBlastStatusCode CLBlastHconvgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) +``` + +Arguments to CONVGEMM: + +* `const size_t channels`: Integer size argument. This value must be positive. +* `const size_t height`: Integer size argument. This value must be positive. +* `const size_t width`: Integer size argument. This value must be positive. +* `const size_t kernel_h`: Integer size argument. This value must be positive. +* `const size_t kernel_w`: Integer size argument. This value must be positive. +* `const size_t pad_h`: Integer size argument. This value must be positive. +* `const size_t pad_w`: Integer size argument. This value must be positive. +* `const size_t stride_h`: Integer size argument. This value must be positive. +* `const size_t stride_w`: Integer size argument. This value must be positive. +* `const size_t dilation_h`: Integer size argument. This value must be positive. +* `const size_t dilation_w`: Integer size argument. This value must be positive. +* `const size_t num_kernels`: Integer size argument. This value must be positive. +* `const size_t batch_count`: Integer size argument. This value must be positive. +* `const cl_mem im_buffer`: OpenCL buffer to store the input im tensor. +* `const size_t im_offset`: The offset in elements from the start of the input im tensor. +* `const cl_mem kernel_buffer`: OpenCL buffer to store the input kernel tensor. +* `const size_t kernel_offset`: The offset in elements from the start of the input kernel tensor. +* `cl_mem result_buffer`: OpenCL buffer to store the output result tensor. +* `const size_t result_offset`: The offset in elements from the start of the output result tensor. +* `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. + + + xAXPYBATCHED: Batched version of AXPY ------------- diff --git a/scripts/generator/generator.py b/scripts/generator/generator.py index e2837dd5..f04d9f3d 100755 --- a/scripts/generator/generator.py +++ b/scripts/generator/generator.py @@ -181,7 +181,7 @@ ROUTINES = [ Routine(True, True, 0, False, "x", "had", T, [S,D,C,Z,H], ["n"], [], ["x","y"], ["z"], [xn,yn,zn], ["alpha","beta"], "", "Element-wise vector product (Hadamard)", "Performs the Hadamard element-wise product _z = alpha * x * y + beta * z_, in which _x_, _y_, and _z_ are vectors and _alpha_ and _beta_ are scalar constants.", []), 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.", []), - Routine(False, True, 0, False, "x", "convgemm", T, [S,D,C,Z,H], convgemm_constants, [], ["im","kernel"], ["result"], [imb,kernel,result],[""], "", "Batched convolution as GEMM (non-BLAS function)", "Integrates im2col and GEMM for batched convolution, in which _im_ is the 4D input tensor, _kernel_ the 3D kernel weights tensor, and _result_ the 4D output tensor.", []), + Routine(True, True, 0, False, "x", "convgemm", T, [S,D,C,Z,H], convgemm_constants, [], ["im","kernel"], ["result"], [imb,kernel,result],[""], "", "Batched convolution as GEMM (non-BLAS function)", "Integrates im2col and GEMM for batched 3D convolution, in which _im_ is the 4D input tensor, _kernel_ the 4D kernel weights tensor, and _result_ the 4D output tensor.", []), # Batched routines: 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]), diff --git a/src/clblast.cpp b/src/clblast.cpp index 026285bb..3a96136a 100644 --- a/src/clblast.cpp +++ b/src/clblast.cpp @@ -2254,12 +2254,20 @@ template StatusCode PUBLIC_API Im2col<half>(const size_t, const size_t, const si // Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/CCONVGEMM/ZCONVGEMM/HCONVGEMM template <typename T> -StatusCode Convgemm(const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, - const cl_mem, const size_t, - const cl_mem, const size_t, - cl_mem, const size_t, - cl_command_queue*, cl_event*) { - return StatusCode::kNotImplemented; +StatusCode Convgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event) { + try { + auto queue_cpp = Queue(*queue); + auto routine = Xconvgemm<T>(queue_cpp, event); + routine.DoConvgemm(channels, height, width, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, num_kernels, batch_count, + Buffer<T>(im_buffer), im_offset, + Buffer<T>(kernel_buffer), kernel_offset, + Buffer<T>(result_buffer), result_offset); + return StatusCode::kSuccess; + } catch (...) { return DispatchException(); } } template StatusCode PUBLIC_API Convgemm<float>(const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const cl_mem, const size_t, diff --git a/src/clblast_cuda.cpp b/src/clblast_cuda.cpp index f89fb77d..5aab1626 100644 --- a/src/clblast_cuda.cpp +++ b/src/clblast_cuda.cpp @@ -2352,12 +2352,22 @@ template StatusCode PUBLIC_API Im2col<half>(const size_t, const size_t, const si // Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/CCONVGEMM/ZCONVGEMM/HCONVGEMM template <typename T> -StatusCode Convgemm(const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, - const CUdeviceptr, const size_t, - const CUdeviceptr, const size_t, - CUdeviceptr, const size_t, - const CUcontext, const CUdevice) { - return StatusCode::kNotImplemented; +StatusCode Convgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const CUdeviceptr im_buffer, const size_t im_offset, + const CUdeviceptr kernel_buffer, const size_t kernel_offset, + CUdeviceptr result_buffer, const size_t result_offset, + 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 = Xconvgemm<T>(queue_cpp, nullptr); + routine.DoConvgemm(channels, height, width, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, num_kernels, batch_count, + Buffer<T>(im_buffer), im_offset, + Buffer<T>(kernel_buffer), kernel_offset, + Buffer<T>(result_buffer), result_offset); + return StatusCode::kSuccess; + } catch (...) { return DispatchException(); } } template StatusCode PUBLIC_API Convgemm<float>(const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const size_t, const CUdeviceptr, const size_t, diff --git a/src/routines/levelx/xconvgemm.cpp b/src/routines/levelx/xconvgemm.cpp new file mode 100644 index 00000000..2676dbda --- /dev/null +++ b/src/routines/levelx/xconvgemm.cpp @@ -0,0 +1,68 @@ + +// ================================================================================================= +// 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 Xconvgemm class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/levelx/xconvgemm.hpp" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xconvgemm<T>::Xconvgemm(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy"}, PrecisionValue<T>(), {}, { +#include "../../kernels/levelx/im2col.opencl" + }) { +} + +// ================================================================================================= + +template <typename T> +void Xconvgemm<T>::DoConvgemm(const size_t channels, const size_t height, const size_t width, + const size_t kernel_h, const size_t kernel_w, const size_t pad_h, + const size_t pad_w, const size_t stride_h, const size_t stride_w, + const size_t dilation_h, const size_t dilation_w, + const size_t num_kernels, const size_t batch_count, + const Buffer<T> &im_buffer, const size_t im_offset, + const Buffer<T> &kernel_buffer, const size_t kernel_offset, + const Buffer<T> &result_buffer, const size_t result_offset) { + + // Makes sure all dimensions are larger than zero + if ((channels == 0) || (height == 0) || (width == 0) || (num_kernels == 0) || (batch_count == 0)) { + throw BLASError(StatusCode::kInvalidDimension); + } + + // Sets the output height and width + const auto size_h = height + 2 * pad_h; + const auto padding_h = dilation_h * (kernel_h - 1) + 1; + const auto output_h = (size_h >= padding_h) ? (size_h - padding_h) / stride_h + 1 : 1; + const auto size_w = width + 2 * pad_w; + const auto padding_w = dilation_w * (kernel_w - 1) + 1; + const auto output_w = (size_w >= padding_w) ? (size_w - padding_w) / stride_w + 1 : 1; + + throw BLASError(StatusCode::kNotImplemented); +} + +// ================================================================================================= + +// Compiles the templated class +template class Xconvgemm<half>; +template class Xconvgemm<float>; +template class Xconvgemm<double>; +template class Xconvgemm<float2>; +template class Xconvgemm<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/levelx/xconvgemm.hpp b/src/routines/levelx/xconvgemm.hpp new file mode 100644 index 00000000..01795ea8 --- /dev/null +++ b/src/routines/levelx/xconvgemm.hpp @@ -0,0 +1,48 @@ + +// ================================================================================================= +// 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 Xconvgemm routine. The precision is implemented as a template argument. +// This implements batched convolution of a 4D input 'image' tensor, a 3D input 'kernel' matrix, +// resulting in a 4D output 'result' tensor. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XCONVGEMM_H_ +#define CLBLAST_ROUTINES_XCONVGEMM_H_ + +#include "routine.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class Xconvgemm: public Routine { + public: + + // Constructor + Xconvgemm(Queue &queue, EventPointer event, const std::string &name = "CONVGEMM"); + + // Templated-precision implementation of the routine + void DoConvgemm(const size_t channels, const size_t height, const size_t width, + const size_t kernel_h, const size_t kernel_w, + const size_t pad_h, const size_t pad_w, + const size_t stride_h, const size_t stride_w, + const size_t dilation_h, const size_t dilation_w, + const size_t num_kernels, const size_t batch_count, + const Buffer<T> &im_buffer, const size_t im_offset, + const Buffer<T> &kernel_buffer, const size_t kernel_offset, + const Buffer<T> &result_buffer, const size_t result_offset); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XCONVGEMM_H_ +#endif diff --git a/src/routines/routines.hpp b/src/routines/routines.hpp index 2ab16a75..e080ed47 100644 --- a/src/routines/routines.hpp +++ b/src/routines/routines.hpp @@ -70,6 +70,7 @@ #include "routines/levelx/xhad.hpp" #include "routines/levelx/xomatcopy.hpp" #include "routines/levelx/xim2col.hpp" +#include "routines/levelx/xconvgemm.hpp" #include "routines/levelx/xaxpybatched.hpp" #include "routines/levelx/xgemmbatched.hpp" #include "routines/levelx/xgemmstridedbatched.hpp" diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp index 0edf77fe..2d2cd62e 100644 --- a/src/utilities/utilities.hpp +++ b/src/utilities/utilities.hpp @@ -84,6 +84,7 @@ constexpr auto kArgImaxOffset = "offimax"; constexpr auto kArgAlpha = "alpha"; constexpr auto kArgBeta = "beta"; constexpr auto kArgBatchCount = "batch_num"; +constexpr auto kArgNumKernels = "num_kernels"; // Constants for im2col constexpr auto kArgChannels = "channels"; @@ -195,7 +196,7 @@ struct Arguments { size_t imax_offset = 0; T alpha = ConstantOne<T>(); T beta = ConstantOne<T>(); - // Arguments for im2col + // Arguments for im2col and convgemm size_t channels = 1; size_t height = 1; size_t width = 1; @@ -207,6 +208,7 @@ struct Arguments { size_t stride_w = 1; size_t dilation_h = 1; size_t dilation_w = 1; + size_t num_kernels = 1; // Batch-specific arguments size_t batch_count = 1; std::vector<size_t> x_offsets; // = {0}; diff --git a/test/correctness/routines/levelx/xconvgemm.cpp b/test/correctness/routines/levelx/xconvgemm.cpp new file mode 100644 index 00000000..77a0f543 --- /dev/null +++ b/test/correctness/routines/levelx/xconvgemm.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/xconvgemm.hpp" + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + auto errors = size_t{0}; + errors += clblast::RunTests<clblast::TestXconvgemm<float>, float, float>(argc, argv, false, "SCONVGEMM"); + errors += clblast::RunTests<clblast::TestXconvgemm<double>, double, double>(argc, argv, true, "DCONVGEMM"); + errors += clblast::RunTests<clblast::TestXconvgemm<clblast::float2>, clblast::float2, clblast::float2>(argc, argv, true, "CCONVGEMM"); + errors += clblast::RunTests<clblast::TestXconvgemm<clblast::double2>, clblast::double2, clblast::double2>(argc, argv, true, "ZCONVGEMM"); + errors += clblast::RunTests<clblast::TestXconvgemm<clblast::half>, clblast::half, clblast::half>(argc, argv, true, "HCONVGEMM"); + if (errors > 0) { return 1; } else { return 0; } +} + +// ================================================================================================= diff --git a/test/correctness/testblas.hpp b/test/correctness/testblas.hpp index 54b2d6f8..1d1d2ca9 100644 --- a/test/correctness/testblas.hpp +++ b/test/correctness/testblas.hpp @@ -60,6 +60,7 @@ class TestBlas: public Tester<T,U> { static const std::vector<size_t> kDilationSizes; static const std::vector<size_t> kKernelSizes; static const std::vector<size_t> kBatchCounts; + static const std::vector<size_t> kNumKernels; const std::vector<size_t> kOffsets; const std::vector<U> kAlphaValues; const std::vector<U> kBetaValues; @@ -136,6 +137,7 @@ template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kBatc template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kPadSizes = { 0, 1 }; template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kDilationSizes = { 1, 2 }; template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kKernelSizes = { 1, 3 }; +template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kNumKernels = { 1, 2 }; // Test settings for the invalid tests template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kInvalidIncrements = { 0, 1 }; @@ -241,6 +243,7 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na auto dilation_hs = std::vector<size_t>{args.dilation_h}; auto dilation_ws = std::vector<size_t>{args.dilation_w}; auto batch_counts = std::vector<size_t>{args.batch_count}; + auto num_kernelss = std::vector<size_t>{args.num_kernels}; auto x_sizes = std::vector<size_t>{args.x_size}; auto y_sizes = std::vector<size_t>{args.y_size}; auto a_sizes = std::vector<size_t>{args.a_size}; @@ -296,6 +299,7 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na if (option == kArgDilationH) { dilation_hs = tester.kDilationSizes; } if (option == kArgDilationW) { dilation_ws = tester.kDilationSizes; } if (option == kArgBatchCount) { batch_counts = tester.kBatchCounts; } + if (option == kArgNumKernels) { num_kernelss = tester.kNumKernels; } if (option == kArgXOffset) { x_sizes = tester.kVecSizes; } if (option == kArgYOffset) { y_sizes = tester.kVecSizes; } @@ -350,8 +354,10 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na for (auto &dilation_h: dilation_hs) { r_args.dilation_h = dilation_h; for (auto &dilation_w: dilation_ws) { r_args.dilation_w = dilation_w; for (auto &batch_count: batch_counts) { r_args.batch_count = batch_count; - C::SetSizes(r_args, tester.queue_); - regular_test_vector.push_back(r_args); + for (auto &num_kernels: num_kernelss) { r_args.num_kernels = num_kernels; + C::SetSizes(r_args, tester.queue_); + regular_test_vector.push_back(r_args); + } } } } diff --git a/test/performance/routines/levelx/xconvgemm.cpp b/test/performance/routines/levelx/xconvgemm.cpp new file mode 100644 index 00000000..e6bbe8e8 --- /dev/null +++ b/test/performance/routines/levelx/xconvgemm.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/xconvgemm.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::TestXconvgemm<clblast::half>, clblast::half, clblast::half>(argc, argv); break; + case clblast::Precision::kSingle: + clblast::RunClient<clblast::TestXconvgemm<float>, float, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient<clblast::TestXconvgemm<double>, double, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + clblast::RunClient<clblast::TestXconvgemm<clblast::float2>, clblast::float2, clblast::float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient<clblast::TestXconvgemm<clblast::double2>, clblast::double2, clblast::double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/levelx/xconvgemm.hpp b/test/routines/levelx/xconvgemm.hpp new file mode 100644 index 00000000..6ca5965b --- /dev/null +++ b/test/routines/levelx/xconvgemm.hpp @@ -0,0 +1,219 @@ + +// ================================================================================================= +// 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 Xconvgemm 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_XCONVGEMM_H_ +#define CLBLAST_TEST_ROUTINES_XCONVGEMM_H_ + +#include "test/routines/common.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXconvgemm { +public: + + // The BLAS level: 4 for the extra routines + static size_t BLASLevel() { return 4; } + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW, + kArgStrideH, kArgStrideW, kArgDilationH, kArgDilationW, kArgNumKernels, kArgBatchCount, + kArgAOffset, kArgBOffset, kArgCOffset}; + } + static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; } + static std::vector<std::string> BuffersOut() { return {kBufMatC}; } + + // Describes how to obtain the sizes of the buffers + static size_t OutputHeight(const Arguments<T> &args) { + const auto size = args.height + 2 * args.pad_h; + const auto padding = args.dilation_h * (args.kernel_h - 1) + 1; + if (size >= padding) { return (size - padding) / args.stride_h + 1; } + return 1; + } + static size_t OutputWidth(const Arguments<T> &args) { + const auto size = args.width + 2 * args.pad_w; + const auto padding = args.dilation_w * (args.kernel_w - 1) + 1; + if (size >= padding) { return (size - padding) / args.stride_w + 1; } + return 1; + } + static size_t NumPatches(const Arguments<T> &args) { + return OutputHeight(args) * OutputWidth(args) * args.channels; + } + static size_t GetSizeA(const Arguments<T> &args) { // 4D: NCHW == batch-channel-height-width + return args.batch_count * args.channels * args.height * args.width + args.a_offset; + } + static size_t GetSizeB(const Arguments<T> &args) { // 4D: KCHW == kernel-channel-height-width + return args.num_kernels * args.channels * args.kernel_h * args.kernel_w + args.b_offset; + } + static size_t GetSizeC(const Arguments<T> &args) { // 4D: NCHW == batch-channel-height-width + return args.batch_count * args.num_kernels * OutputHeight(args) * OutputWidth(args) + 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> &) { return 1; } // N/A for this routine + 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 &) { return {}; } // N/A for this routine + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // 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 = Convgemm<T>(args.channels, args.height, args.width, + args.kernel_h, args.kernel_w, + args.pad_h, args.pad_w, + args.stride_h, args.stride_w, + args.dilation_h, args.dilation_w, + args.num_kernels, args.batch_count, + buffers.a_mat(), args.a_offset, + buffers.b_mat(), args.b_offset, + buffers.c_mat(), args.c_offset, + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } +#elif CUDA_API + auto status = Convgemm<T>(args.channels, args.height, args.width, + args.kernel_h, args.kernel_w, + args.pad_h, args.pad_w, + args.stride_h, args.stride_w, + args.dilation_h, args.dilation_w, + args.num_kernels, args.batch_count, + buffers.a_mat(), args.a_offset, + buffers.b_mat(), args.b_offset, + buffers.c_mat(), args.c_offset, + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); +#endif + return status; + } + + // Describes how to run a naive version of the routine (for correctness/performance comparison). + // Note that a proper clBLAS or CPU BLAS comparison is not available for non-BLAS routines. + static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + auto buffers_host = BuffersHost<T>(); + DeviceToHost(args, buffers, buffers_host, queue, BuffersIn()); + const auto status = RunReference(args, buffers_host); + HostToDevice(args, buffers, buffers_host, queue, BuffersOut()); + return status; + } + + static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue&) { + return RunReference(args, buffers_host); + } + static StatusCode RunReference3(const Arguments<T> &, BuffersCUDA<T> &, Queue &) { + return StatusCode::kUnknownError; + } + + // 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 OutputHeight(args) * OutputWidth(args); } + static size_t ResultID2(const Arguments<T> &args) { return args.num_kernels * args.batch_count; } + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t id2) { + return id1 + OutputHeight(args) * OutputWidth(args) * id2 + args.c_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return args.batch_count; // TODO + } + static size_t GetBytes(const Arguments<T> &args) { + return (GetSizeA(args) + GetSizeB(args) + GetSizeC(args)) * sizeof(T); + } +}; + +// ================================================================================================= + +template <typename T> +StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host) { + const auto output_h = TestXconvgemm<T>::OutputHeight(args); + const auto output_w = TestXconvgemm<T>::OutputWidth(args); + for (auto batch_id = size_t{0}; batch_id < args.batch_count; ++batch_id) { + for (auto co_id = size_t{0}; co_id < args.num_kernels; ++co_id) { // output channels == num-kernels + for (auto ho_id = size_t{0}; ho_id < output_h; ++ho_id) { // image height + for (auto wo_id = size_t{0}; wo_id < output_w; ++wo_id) { // image width + auto result = ConstantZero<T>(); + + // 3D convolution + for (auto ci_id = size_t{0}; ci_id < args.channels; ++ci_id) { // input channels + for (auto kh_id = size_t{0}; kh_id < args.kernel_h; ++kh_id) { // kernel height + for (auto kw_id = size_t{0}; kw_id < args.kernel_w; ++kw_id) { // kernel width + + // Retrieves the value from the input image + const auto hi_id = kh_id * args.dilation_h + args.stride_h * ho_id - args.pad_h; + const auto wi_id = kw_id * args.dilation_w + args.stride_w * wo_id - args.pad_w; + if (hi_id >= 0 && hi_id < args.height && + wi_id >= 0 && wi_id < args.width) { + const auto input_index = wi_id + args.width * ( + hi_id + args.height * ( + ci_id + args.channels * ( + batch_id))); + const auto input_value = buffers_host.a_mat[input_index + args.a_offset]; + + // Multiplies with the kernel tensor + const auto kernel_index = kw_id + args.kernel_w * ( + kh_id + args.kernel_h * ( + ci_id + args.channels * ( + co_id))); + const auto kernel_value = buffers_host.b_mat[kernel_index + args.b_offset]; + result += input_value * kernel_value; + + } + } + } + } + + // Sets the output value (NCHW == batch-channel-height-width) + const auto output_index = wo_id + output_w * ( + ho_id + output_h * ( + co_id + args.num_kernels * ( + batch_id))); + buffers_host.c_mat[output_index + args.c_offset] = result; + } + } + } + } + return StatusCode::kSuccess; +} + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XCONVGEMM_H_ +#endif |