conv2d_winograd_?

2D convolution interface for Winograd algorithm, supporting only non-dilated convolution computation with a stride of 1 and a dilation coefficient of 1.

Interface Definition

C interface:

void conv2d_winograd_fp32(const float *input, const int batch, const int inputChannels, const int inputHeight, const int inputWidth, const float *kernel, const int kernelHeight, const int kernelWidth,const int padHeight, const int padWidth, const float *bias, float *output, const int outputChannels);

void conv2d_winograd_fp16(const __fp16 *input, const int batch, const int inputChannels, const int inputHeight, const int inputWidth, const __fp16 *kernel, const int kernelHeight, const int kernelWidth, const int padHeight, const int padWidth, const __fp16 *bias, __fp16 *output, const int outputChannels);

Parameters

Parameter	Type	Description	Input/Output
input	conv2d_winograd_fp32: float conv2d_winograd_fp16: __fp16	Input data	Input
batch	Int	Number of batches of input data	Input
inputChannels	Int	Number of input channels	Input
inputHeight	Int	Height of the input data.	Input
inputWidth	Int	Width of the input data.	Input
kernel	conv2d_winograd_fp32: float conv2d_winograd_fp16: __fp16	Convolution kernel.	Input
kernelHeight	Int	Height of the convolution kernel.	Input
kernelWidth	Int	Width of the convolution kernel.	Input
padHeight	Int	Zero-padding length at both ends of the original input data's height.	Input
padWidth	Int	Zero-padding length at both ends of the original input data's width.	Input
bias	conv2d_winograd_fp32: float conv2d_winograd_fp16: __fp16	Bias. The value NULL indicates that there is no bias.	Input
output	conv2d_winograd_fp32: float conv2d_winograd_fp16: __fp16	Output result data.	Output
outputChannels	Int	Number of output channels.	Input

Dependencies

#include "conv.h"

Examples

C interface:

    int batch = 1;
    int inputChannels = 1;
    int inputHeight = 6;
    int inputWidth = 6;
    int kernelHeight = 3;
    int kernelWidth = 3;
    int strideY = 1;
    int strideX = 1;
    int padHeight = 0;
    int padWidth = 0;
    int dilationY = 1;
    int dilationX = 1;
    int outputChannels = 1;
    float input[36] = {1.0, 2.0, 3.0, 4.0, 5.0, 6.0,
                       7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
                       13.0, 14.0, 15.0, 16.0, 17.0, 18.0,
                       19.0, 20.0, 21.0, 22.0, 23.0, 24.0,
                       25.0, 26.0, 27.0, 28.0, 29.0, 30.0,
                       31.0, 32.0, 33.0, 34.0, 35.0, 36.0};
    float kernel[9] = {1.0, 2.0, 3.0,
                       4.0, 5.0, 6.0,
                       7.0, 8.0, 9.0};

    float *bias = NULL;
    int outputHeight = (inputHeight + 2 * padHeight - (kernelHeight - 1) - 1) + 1;
    int outputWidth = (inputWidth + 2 * padWidth - (kernelWidth - 1) - 1) + 1;
    /*
     *   outputHeight x outputWidth = 4 x 4
     */

    float output[16] = {0.0};
    conv2d_winograd_fp32(input, batch, inputChannels, inputHeight, inputWidth, kernel, kernelHeight, kernelWidth, padHeight, padWidth, bias, output, outputChannels);
    /*
     * output = [474.0 519.0 564.0 609.0,
     *           744.0 789.0 834.0 879.0, 
     *           1014.0 1059.0 1104.0 1149.0, 
     *           1284.0 1329.0 1374.0 1419.0]
     */

Parent topic: General Convolution Interfaces