diff options
-rw-r--r-- | src/routines/levelx/xcol2im.cpp | 0 | ||||
-rw-r--r-- | src/routines/levelx/xim2col.cpp | 16 | ||||
-rw-r--r-- | src/routines/levelx/xim2col.hpp | 1 | ||||
-rw-r--r-- | test/routines/levelx/xim2col.hpp | 26 |
4 files changed, 22 insertions, 21 deletions
diff --git a/src/routines/levelx/xcol2im.cpp b/src/routines/levelx/xcol2im.cpp new file mode 100644 index 00000000..e69de29b --- /dev/null +++ b/src/routines/levelx/xcol2im.cpp diff --git a/src/routines/levelx/xim2col.cpp b/src/routines/levelx/xim2col.cpp index dfbb4bb5..09dcc42c 100644 --- a/src/routines/levelx/xim2col.cpp +++ b/src/routines/levelx/xim2col.cpp @@ -41,23 +41,23 @@ void Xim2col<T>::DoIm2col(const size_t channels, const size_t height, const size // Makes sure all dimensions are larger than zero if ((channels == 0) || (height == 0) || (width == 0)) { throw BLASError(StatusCode::kInvalidDimension); } - // Sets the output height and width + // Sets the height and width of the 'col' result 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 col_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; + const auto col_w = (size_w >= padding_w) ? (size_w - padding_w) / stride_w + 1 : 1; - // Retrieves the Xcopy kernel from the compiled binary + // Retrieves the kernel from the compiled binary auto kernel = Kernel(program_, "im2col"); // Sets the kernel arguments kernel.SetArgument(0, static_cast<int>(height)); kernel.SetArgument(1, static_cast<int>(width)); kernel.SetArgument(2, static_cast<int>(channels)); - kernel.SetArgument(3, static_cast<int>(output_h)); - kernel.SetArgument(4, static_cast<int>(output_w)); + kernel.SetArgument(3, static_cast<int>(col_h)); + kernel.SetArgument(4, static_cast<int>(col_w)); kernel.SetArgument(5, static_cast<int>(kernel_h)); kernel.SetArgument(6, static_cast<int>(kernel_w)); kernel.SetArgument(7, static_cast<int>(pad_h)); @@ -72,8 +72,8 @@ void Xim2col<T>::DoIm2col(const size_t channels, const size_t height, const size kernel.SetArgument(16, static_cast<int>(col_offset)); // Launches the kernel - const auto w_ceiled = Ceil(output_w, db_["COPY_DIMX"]); - const auto h_ceiled = Ceil(output_h, db_["COPY_DIMY"]); + const auto w_ceiled = Ceil(col_w, db_["COPY_DIMX"]); + const auto h_ceiled = Ceil(col_h, db_["COPY_DIMY"]); const auto global = std::vector<size_t>{w_ceiled, h_ceiled * channels}; const auto local = std::vector<size_t>{db_["COPY_DIMX"], db_["COPY_DIMY"]}; RunKernel(kernel, queue_, device_, global, local, event_); diff --git a/src/routines/levelx/xim2col.hpp b/src/routines/levelx/xim2col.hpp index 4448b54e..2c03b169 100644 --- a/src/routines/levelx/xim2col.hpp +++ b/src/routines/levelx/xim2col.hpp @@ -8,6 +8,7 @@ // Cedric Nugteren <www.cedricnugteren.nl> // // This file implements the Xim2col routine. The precision is implemented using a template argument. +// Uses the tuning parameters from the regular copy kernel. // // ================================================================================================= diff --git a/test/routines/levelx/xim2col.hpp b/test/routines/levelx/xim2col.hpp index 092e251d..9fd2af0c 100644 --- a/test/routines/levelx/xim2col.hpp +++ b/test/routines/levelx/xim2col.hpp @@ -39,20 +39,20 @@ public: static std::vector<std::string> BuffersOut() { return {kBufMatB}; } // Describes how to obtain the sizes of the buffers - static size_t OutputHeight(const Arguments<T> &args) { + static size_t ColHeight(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) { + static size_t ColWidth(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; + return ColHeight(args) * ColWidth(args) * args.channels; } static size_t GetSizeA(const Arguments<T> &args) { return args.height * args.width * args.channels + args.a_offset; @@ -156,13 +156,13 @@ public: template <typename T> StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host) { - const auto output_h = TestXim2col<T>::OutputHeight(args); - const auto output_w = TestXim2col<T>::OutputWidth(args); + const auto col_h = TestXim2col<T>::ColHeight(args); + const auto col_w = TestXim2col<T>::ColWidth(args); for (auto c_id = size_t{0}; c_id < args.channels; ++c_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 - for (auto h_id = size_t{0}; h_id < output_h; ++h_id) { // image height - for (auto w_id = size_t{0}; w_id < output_w; ++w_id) { // image width + for (auto h_id = size_t{0}; h_id < col_h; ++h_id) { // image height + for (auto w_id = size_t{0}; w_id < col_w; ++w_id) { // image width // Retrieves the input value const auto h_index = kh_id * args.dilation_h + args.stride_h * h_id - args.pad_h; @@ -170,16 +170,16 @@ StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host) auto val = ConstantZero<T>(); if (h_index >= 0 && h_index < args.height && w_index >= 0 && w_index < args.width) { - const auto input_index = w_index + args.width * (h_index + args.height * c_id); - val = buffers_host.a_mat[input_index + args.a_offset]; + const auto im_index = w_index + args.width * (h_index + args.height * c_id); + val = buffers_host.a_mat[im_index + args.a_offset]; } // Sets the output value const auto kernel_index = kw_id + args.kernel_w * kh_id; - const auto patch_index = w_id + output_w * h_id; - const auto output_index = patch_index + kernel_index * output_w * output_h + - c_id * output_w * output_h * args.kernel_h * args.kernel_w; - buffers_host.b_mat[output_index + args.b_offset] = val; + const auto patch_index = w_id + col_w * h_id; + const auto col_index = patch_index + kernel_index * col_w * col_h + + c_id * col_w * col_h * args.kernel_h * args.kernel_w; + buffers_host.b_mat[col_index + args.b_offset] = val; } } } |