diff options
Diffstat (limited to 'src/core/CL/cl_kernels')
| -rw-r--r-- | src/core/CL/cl_kernels/depthwise_convolution.cl | 420 | ||||
| -rw-r--r-- | src/core/CL/cl_kernels/depthwise_convolution_quantized.cl | 847 |
2 files changed, 170 insertions, 1097 deletions
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl index 8ce5617858..22a38e7094 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,110 +42,110 @@ inline __global uchar *ptr_offset(__global uchar *ptr, const int x, const int y, #if(DILATION_X == 1 && DILATION_Y == 1) -#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0, weights_row0) \ - ({ \ - acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ +#define CONVOLUTION1x3_2X1_STRIDE1(acc, src0, weights_row0) \ + ({ \ + acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ }) -#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0, weights_row0) \ - ({ \ - acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ - acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2); \ - acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2); \ - acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2); \ - acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3); \ - acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3); \ - acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3); \ +#define CONVOLUTION1x3_4X1_STRIDE1(acc, src0, weights_row0) \ + ({ \ + acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ + acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2); \ + acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2); \ + acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2); \ + acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3); \ + acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3); \ + acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3); \ }) -#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0, src1, weights_row0) \ - ({ \ - acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1); \ +#define CONVOLUTION1x3_2X1_STRIDE2(acc, src0, src1, weights_row0) \ + ({ \ + acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1); \ }) -#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0, src1, weights_row0) \ - ({ \ - acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1); \ - acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2); \ - acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2); \ - acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2); \ - acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3); \ - acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3); \ - acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3); \ +#define CONVOLUTION1x3_4X1_STRIDE2(acc, src0, src1, weights_row0) \ + ({ \ + acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1); \ + acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2); \ + acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2); \ + acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2); \ + acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3); \ + acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3); \ + acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3); \ }) #else /* DILATION_X==1 && DILATION_Y==1 */ -#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ - ({ \ - acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ +#define CONVOLUTION1x3_2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ + ({ \ + acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ }) -#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ - ({ \ - acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ +#define CONVOLUTION1x3_2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ + ({ \ + acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ }) -#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ - ({ \ - acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ - acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2); \ - acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2); \ - acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2); \ - acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3); \ - acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3); \ - acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3); \ +#define CONVOLUTION1x3_4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ + ({ \ + acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ + acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2); \ + acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2); \ + acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2); \ + acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3); \ + acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3); \ + acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3); \ }) -#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ - ({ \ - acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ - acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ - acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ - acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ - acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ - acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ - acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2); \ - acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2); \ - acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2); \ - acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3); \ - acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3); \ - acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3); \ +#define CONVOLUTION1x3_4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ + ({ \ + acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ + acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ + acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ + acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ + acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ + acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ + acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2); \ + acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2); \ + acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2); \ + acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3); \ + acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3); \ + acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3); \ }) #endif /* DILATION_X==1 && DILATION_Y==1 */ @@ -385,8 +385,8 @@ __kernel void depthwise_convolution_3x3( * @param[in] weights_addr Pointer from where to get weights * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension */ -inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, - const int y_offset, __global uchar *weights_addr, const int weights_stride_y) +inline float2 convolution_3x3_dilation_stridex1_stridey1_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, + const int y_offset, __global uchar *weights_addr, const int weights_stride_y) { // Load the weights float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); @@ -407,9 +407,9 @@ inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uc float2 src20_mid = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); float2 src20_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); return pixels0; } @@ -423,8 +423,8 @@ inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uc * @param[in] weights_addr Pointer from where to get weights * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension */ -inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, - const int y_offset, __global uchar *weights_addr, const int weights_stride_y) +inline float2 convolution_3x3_dilation_stridex2_stridey2_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, + const int y_offset, __global uchar *weights_addr, const int weights_stride_y) { // Load the weights float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); @@ -445,9 +445,9 @@ inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uc float3 src20_mid = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); float3 src20_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); return pixels0; } @@ -491,7 +491,7 @@ inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uc * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector */ -__kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32( +__kernel void depthwise_convolution_3x3_stridex1_stridey1_f32( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), TENSOR3D_DECLARATION(weights) @@ -531,29 +531,29 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32( float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 float4 src50 = vload4(0, (__global float *)(src_addr + 5 * src_stride_y)); // Row5 - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20, weights_row2); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src10, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src20, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src30, weights_row2); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src20, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src30, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src40, weights_row2); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src30, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src40, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src50, weights_row2); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src00, weights_row0); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src10, weights_row1); + CONVOLUTION1x3_2X1_STRIDE1(pixels0, src20, weights_row2); + CONVOLUTION1x3_2X1_STRIDE1(pixels1, src10, weights_row0); + CONVOLUTION1x3_2X1_STRIDE1(pixels1, src20, weights_row1); + CONVOLUTION1x3_2X1_STRIDE1(pixels1, src30, weights_row2); + CONVOLUTION1x3_2X1_STRIDE1(pixels2, src20, weights_row0); + CONVOLUTION1x3_2X1_STRIDE1(pixels2, src30, weights_row1); + CONVOLUTION1x3_2X1_STRIDE1(pixels2, src40, weights_row2); + CONVOLUTION1x3_2X1_STRIDE1(pixels3, src30, weights_row0); + CONVOLUTION1x3_2X1_STRIDE1(pixels3, src40, weights_row1); + CONVOLUTION1x3_2X1_STRIDE1(pixels3, src50, weights_row2); #else /* DILATION_X==1 && DILATION_Y==1 */ //3x3 Convolution of elements starting in 0th row - pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); + pixels0 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 1st row - pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y); + pixels1 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 2nd row - pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); + pixels2 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 3rd row - pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y); + pixels3 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y); #endif /* DILATION_X==1 && DILATION_Y==1 */ @@ -611,7 +611,7 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32( * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector */ -__kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32( +__kernel void depthwise_convolution_3x3_stridex2_stridey2_f32( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), TENSOR3D_DECLARATION(weights) @@ -654,19 +654,19 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32( float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 float2 src41 = vload2(2, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00, src01, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10, src11, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20, src21, weights_row2); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src20, src21, weights_row0); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src30, src31, weights_row1); - CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src40, src41, weights_row2); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src00, src01, weights_row0); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src10, src11, weights_row1); + CONVOLUTION1x3_2X1_STRIDE2(pixels0, src20, src21, weights_row2); + CONVOLUTION1x3_2X1_STRIDE2(pixels1, src20, src21, weights_row0); + CONVOLUTION1x3_2X1_STRIDE2(pixels1, src30, src31, weights_row1); + CONVOLUTION1x3_2X1_STRIDE2(pixels1, src40, src41, weights_row2); #else /* DILATION_X==1 && DILATION_Y==1 */ //3x3 Convolution of elements starting in 0th row - pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); + pixels0 = convolution_3x3_dilation_stridex2_stridey2_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 2nd row - pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); + pixels1 = convolution_3x3_dilation_stridex2_stridey2_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); #endif /* DILATION_X==1 && DILATION_Y==1 */ #ifdef HAS_BIAS @@ -684,104 +684,6 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32( #endif // defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F32) -#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH) -/** Reshape the weights for quantized depthwise convolution - * - * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type, e.g. -DDATA_TYPE=uint8 - * @note Output width should be given as a preprocessor argument using -DDST_WIDTH=width, e.g. -DDST_WIDTH=128 - * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=vec_size, e.g., -DVEC_SIZE=4 - * @attention Input's height and width should be 3 - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: All - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void depthwise_convolution_reshape_weights( - TENSOR3D_DECLARATION(src), - IMAGE_DECLARATION(dst)) -{ - Vector src = CONVERT_TO_VECTOR_STRUCT(src); - const int x = get_global_id(0); - - // Load 3x3xVEC_SIZE weights - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w0 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 0 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w1 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 0 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w2 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 0 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w3 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 1 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w4 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 1 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w5 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 1 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w6 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 2 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w7 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 2 * src_stride_z); - VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - w8 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 2 * src_stride_z); - - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * DST_WIDTH * sizeof(DATA_TYPE); - -#if defined(TRANSPOSE) -#if VEC_SIZE != 4 -#error "VEC_SIZE not supported" -#else // VEC_SIZE != 4 - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w0.s0, w1.s0, w2.s0, w3.s0), 0, dst_addr + 0); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w4.s0, w5.s0, w6.s0, w7.s0), 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w8.s0, w0.s1, w1.s1, w2.s1), 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w3.s1, w4.s1, w5.s1, w6.s1), 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w7.s1, w8.s1, w0.s2, w1.s2), 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w2.s2, w3.s2, w4.s2, w5.s2), 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w6.s2, w7.s2, w8.s2, w0.s3), 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w1.s3, w2.s3, w3.s3, w4.s3), 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w5.s3, w6.s3, w7.s3, w8.s3), 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE); -#endif // VEC_SIZE != 4 -#else // !defined(TRANSPOSE) - VSTORE(VEC_SIZE) - (w0, 0, dst_addr + 0); - VSTORE(VEC_SIZE) - (w1, 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w2, 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w3, 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w4, 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w5, 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w6, 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w7, 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE); - VSTORE(VEC_SIZE) - (w8, 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE); -#endif // defined(TRANSPOSE) -} -#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH) - #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16) #if defined(CONV_STRIDE_X) #if CONV_STRIDE_X == 1 @@ -805,8 +707,8 @@ __kernel void depthwise_convolution_reshape_weights( * @param[in] weights_addr Pointer from where to get weights * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension */ -inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, - const int y_offset, __global uchar *weights_addr, const int weights_stride_y) +inline half4 convolution_3x3_dilation_stridex1_stridey1_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, + const int y_offset, __global uchar *weights_addr, const int weights_stride_y) { // Load the weights half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); @@ -827,9 +729,9 @@ inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uch half4 src20_mid = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); half4 src20_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); return pixels0; } @@ -843,8 +745,8 @@ inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uch * @param[in] weights_addr Pointer from where to get weights * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension */ -inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, - const int y_offset, __global uchar *weights_addr, const int weights_stride_y) +inline half4 convolution_3x3_dilation_stridex2_stridey2_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, + const int y_offset, __global uchar *weights_addr, const int weights_stride_y) { // Load the weights half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); @@ -865,9 +767,9 @@ inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uch half8 src20_mid = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); half8 src20_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); return pixels0; } @@ -1127,7 +1029,7 @@ __kernel void depthwise_convolution_3x3_f16( * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector */ -__kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16( +__kernel void depthwise_convolution_3x3_stridex1_stridey1_f16( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), TENSOR3D_DECLARATION(weights) @@ -1174,29 +1076,29 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16( half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 half8 src50 = vload8(0, (__global half *)(src_addr + 5 * src_stride_y)); // Row5 - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20, weights_row2); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src10, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src20, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src30, weights_row2); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src20, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src30, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src40, weights_row2); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src30, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src40, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src50, weights_row2); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src00, weights_row0); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src10, weights_row1); + CONVOLUTION1x3_4X1_STRIDE1(pixels0, src20, weights_row2); + CONVOLUTION1x3_4X1_STRIDE1(pixels1, src10, weights_row0); + CONVOLUTION1x3_4X1_STRIDE1(pixels1, src20, weights_row1); + CONVOLUTION1x3_4X1_STRIDE1(pixels1, src30, weights_row2); + CONVOLUTION1x3_4X1_STRIDE1(pixels2, src20, weights_row0); + CONVOLUTION1x3_4X1_STRIDE1(pixels2, src30, weights_row1); + CONVOLUTION1x3_4X1_STRIDE1(pixels2, src40, weights_row2); + CONVOLUTION1x3_4X1_STRIDE1(pixels3, src30, weights_row0); + CONVOLUTION1x3_4X1_STRIDE1(pixels3, src40, weights_row1); + CONVOLUTION1x3_4X1_STRIDE1(pixels3, src50, weights_row2); #else /* DILATION_X==1 && DILATION_Y==1 */ //3x3 Convolution of elements starting in 0th row - pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); + pixels0 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 1st row - pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y); + pixels1 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 2nd row - pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); + pixels2 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 3rd row - pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y); + pixels3 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y); #endif /* DILATION_X==1 && DILATION_Y==1 */ @@ -1250,7 +1152,7 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16( * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector */ -__kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16( +__kernel void depthwise_convolution_3x3_stridex2_stridey2_f16( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), TENSOR3D_DECLARATION(weights) @@ -1300,18 +1202,18 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16( half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 half2 src41 = vload2(4, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00, src01, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10, src11, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20, src21, weights_row2); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src20, src21, weights_row0); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src30, src31, weights_row1); - CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src40, src41, weights_row2); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src00, src01, weights_row0); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src10, src11, weights_row1); + CONVOLUTION1x3_4X1_STRIDE2(pixels0, src20, src21, weights_row2); + CONVOLUTION1x3_4X1_STRIDE2(pixels1, src20, src21, weights_row0); + CONVOLUTION1x3_4X1_STRIDE2(pixels1, src30, src31, weights_row1); + CONVOLUTION1x3_4X1_STRIDE2(pixels1, src40, src41, weights_row2); #else /* DILATION_X==1 && DILATION_Y==1 */ //3x3 Convolution of elements starting in 0th row - pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); + pixels0 = convolution_3x3_dilation_stridex2_stridey2_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 2nd row - pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); + pixels1 = convolution_3x3_dilation_stridex2_stridey2_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y); #endif /* DILATION_X==1 && DILATION_Y==1 */ #ifdef HAS_BIAS diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl index c7fe401f80..000dce1590 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl @@ -334,9 +334,9 @@ __kernel void dwc_3x3_native_quantized8_nchw( #else // defined(REAL_MULTIPLIER) #if defined(PER_CHANNEL_QUANTIZATION) - int8 res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8); - int8 res0_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8); - values0 = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0); + int8 res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8); + int8 res0_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8); + values0 = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0); #else // defined(PER_CHANNEL_QUANTIZATION) #if OUTPUT_SHIFT < 0 values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); @@ -360,9 +360,9 @@ __kernel void dwc_3x3_native_quantized8_nchw( #else // defined(REAL_MULTIPLIER) #if defined(PER_CHANNEL_QUANTIZATION) - int8 res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8); - int8 res1_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8); - values1 = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0); + int8 res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8); + int8 res1_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8); + values1 = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0); #else // defined(PER_CHANNEL_QUANTIZATION) #if OUTPUT_SHIFT < 0 values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); @@ -446,8 +446,8 @@ __kernel void dwc_3x3_native_quantized8_nchw( VEC_TYPE(16) \ temp0 = vload16(0, (__global DATA_TYPE *)(first_value)); \ VEC_TYPE(8) \ - temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE)))); \ - left = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \ + temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))); \ + left = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \ \ temp0 = vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))); \ temp1 = vload8(0, (__global DATA_TYPE *)(first_value + (16 + DILATION_X) * sizeof(DATA_TYPE))); \ @@ -776,835 +776,6 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw( #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) -#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && VEC_SIZE == 4 -/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width or height is not 1. - * - * @note This kernel assumes VEC_SIZE is 4. - * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel. - * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) - * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) - * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X) - * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1) - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor reshaped. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32 - * @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes) - * @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector - * @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32 - * @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes) - * @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector - * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32 - * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) - * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector - * @param[in] max_offset Max offset for the input tensor - */ -__kernel void dwc_3x3_reshaped_quantized8_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR4D_DECLARATION(dst), - IMAGE_DECLARATION(weights), - VECTOR_DECLARATION(output_multipliers), - VECTOR_DECLARATION(output_shifts), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - int max_offset) -{ - const int x = get_global_id(0); // channels - const int y = get_global_id(1); // spatial coordinate x -#if defined(DST_DEPTH) - int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y - int b = get_global_id(2) / (int)DST_DEPTH; // batch -#else // defined(DST_DEPTH) - int z = get_global_id(2); // spatial coordinate y -#endif // defined(DST_DEPTH) - - __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y; - -#if defined(DST_DEPTH) - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; -#endif /* defined(DST_DEPTH) */ - - int z_coord = 0; - int4 offset = 0; - int4 y_coord = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3)) - (int)CONV_PAD_LEFT; - - // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1 - y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1); - y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1); - y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1); - y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1); - - int4 y_offset = convert_int4(y_coord * (int)src_stride_y); - - // We compute VEC_SIZEx1x1 [C,W,H] elements - VEC_INT acc = 0, sum = 0; - - // Load weights - VEC_DATA_TYPE(WEIGHTS_TYPE, 16) - w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr)); - VEC_DATA_TYPE(WEIGHTS_TYPE, 16) - w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16)); - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16)); - - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w0 = w0_tmp.s0123; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w1 = w0_tmp.s4567; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w2 = w0_tmp.s89AB; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w3 = w0_tmp.sCDEF; - - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w4 = w1_tmp.s0123; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w5 = w1_tmp.s4567; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w6 = w1_tmp.s89AB; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w7 = w1_tmp.sCDEF; - -#if INPUT_OFFSET != 0 - VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT) - + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT) - + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT); -#endif /* INPUT_OFFSET != 0 */ - - // Load input values - // z == 0 - // Clamp z_coord as for z = 0, it can be negative - // z_coord is casted to unsigned int in order to use just a min() operation - // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 - z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - - // z == 1 - // z_coord can be only negative for z = 0 so we do not need to clamp it - // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset - z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - - // z == 2 - // Offset can be out-of-bound so we need to check if it is greater than max_offset - z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - - MULTIPLY_ADD_ACCUMULATE(values0, w0, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values1, w1, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values2, w2, acc, sum); - - MULTIPLY_ADD_ACCUMULATE(values3, w3, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values4, w4, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values5, w5, acc, sum); - - MULTIPLY_ADD_ACCUMULATE(values6, w6, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values7, w7, acc, sum); - MULTIPLY_ADD_ACCUMULATE(values8, w8, acc, sum); - -#if defined(HAS_BIAS) - Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); - VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); - acc += bias_values; -#endif // defined(HAS_BIAS) - -#if WEIGHTS_OFFSET != 0 - acc += WEIGHTS_OFFSET * sum; -#endif /* WEIGHTS_OFFSET != 0 */ - -#if INPUT_OFFSET != 0 - acc += INPUT_OFFSET * sum_we; -#endif /* INPUT_OFFSET != 0 */ - -#if K_OFFSET != 0 - acc += (VEC_INT)K_OFFSET; -#endif /* K_OFFSET != 0 */ - -#if defined(REAL_MULTIPLIER) - - acc = CONVERT(round(CONVERT(acc, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - -#else // defined(REAL_MULTIPLIER) - -#if defined(PER_CHANNEL_QUANTIZATION) - Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers); - Vector output_shifts = CONVERT_TO_VECTOR_STRUCT(output_shifts); - VEC_INT output_multiplier = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr); - VEC_INT output_shift = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr); - - VEC_INT res_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, output_multiplier, output_shift, VEC_SIZE); - VEC_INT res_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc, output_multiplier, output_shift); - acc = select(res_shift_lt0, res_shift_gt0, output_shift >= 0); -#else // defined(PER_CHANNEL_QUANTIZATION) -#if OUTPUT_SHIFT < 0 - acc = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); -#else // OUTPUT_SHIFT < 0 - acc = asymm_mult_by_quant_multiplier_less_than_one(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); -#endif // OUTPUT_SHIFT < 0 -#endif // defined(PER_CHANNEL_QUANTIZATION) - -#endif // defined(REAL_MULTIPLIER) - - acc += (VEC_INT)OUTPUT_OFFSET; - - VEC_TYPE(VEC_SIZE) - res = CONVERT_SAT(acc, VEC_TYPE(VEC_SIZE)); - -#if defined(DST_DEPTH) - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z; -#endif /* defined(DST_DEPTH) */ - - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res), 0, (__global DATA_TYPE *)(dst_addr)); -} -#endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) - -#if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) && VEC_SIZE == 4 -/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1. - * - * @note This kernel assumes VEC_SIZE is 4. - * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel. - * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) - * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) - * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2) - * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1). - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32 - * @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes) - * @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector - * @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32 - * @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes) - * @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector - * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32 - * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) - * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector - * @param[in] max_offset Max offset for the input tensor - */ - -__kernel void dwc_3x3_reshaped_quantized8_stride1_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR4D_DECLARATION(dst), - IMAGE_DECLARATION(weights), - VECTOR_DECLARATION(output_multipliers), - VECTOR_DECLARATION(output_shifts), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - int max_offset) -{ - int x = get_global_id(0); - int y = get_global_id(1); -#if defined(DST_DEPTH) - int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y - int b = get_global_id(2) / (int)DST_DEPTH; // batch -#else // defined(DST_DEPTH) - int z = get_global_id(2); // spatial coordinate y -#endif // defined(DST_DEPTH) - - __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y; - -#if defined(DST_DEPTH) - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; -#endif /* defined(DST_DEPTH) */ - - int z_coord = 0; - int4 offset = 0; - int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT; - - // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1 - y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1); - y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1); - y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1); - y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1); - - int4 y_offset = convert_int4(y_coord * (int)src_stride_y); - - // We compute 4x2x2 [C,W,H] elements - VEC_INT acc0 = 0, sum0 = 0; - VEC_INT acc1 = 0, sum1 = 0; - VEC_INT acc2 = 0, sum2 = 0; - VEC_INT acc3 = 0, sum3 = 0; - - // Load weights - VEC_DATA_TYPE(WEIGHTS_TYPE, 16) - w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr)); - VEC_DATA_TYPE(WEIGHTS_TYPE, 16) - w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16)); - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16)); - - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w0 = w0_tmp.s0123; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w1 = w0_tmp.s4567; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w2 = w0_tmp.s89AB; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w3 = w0_tmp.sCDEF; - - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w4 = w1_tmp.s0123; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w5 = w1_tmp.s4567; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w6 = w1_tmp.s89AB; - VEC_DATA_TYPE(WEIGHTS_TYPE, 4) - w7 = w1_tmp.sCDEF; - -#if INPUT_OFFSET != 0 - VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT) - + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT) - + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT); -#endif /* INPUT_OFFSET != 0 */ - - // Load input values - // z == 0 - // Clamp z_coord as for z = 0, it can be negative - // z_coord is casted to unsigned int in order to use just a min() operation - // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 - z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - // z == 1 - z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - // z == 2 - z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 2; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - // z == 3 - z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 3; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2); - VEC_TYPE(VEC_SIZE) - values12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - MULTIPLY_ADD_ACCUMULATE(values0, w0, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values1, w1, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values2, w2, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values1, w0, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values2, w1, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values3, w2, acc1, sum1); - - MULTIPLY_ADD_ACCUMULATE(values4, w3, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values5, w4, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values6, w5, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values5, w3, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values6, w4, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values7, w5, acc1, sum1); - - MULTIPLY_ADD_ACCUMULATE(values8, w6, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values9, w7, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values10, w8, acc0, sum0); - MULTIPLY_ADD_ACCUMULATE(values9, w6, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values10, w7, acc1, sum1); - MULTIPLY_ADD_ACCUMULATE(values11, w8, acc1, sum1); - - MULTIPLY_ADD_ACCUMULATE(values4, w0, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values5, w1, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values6, w2, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values5, w0, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values6, w1, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values7, w2, acc3, sum3); - - MULTIPLY_ADD_ACCUMULATE(values8, w3, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values9, w4, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values10, w5, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values9, w3, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values10, w4, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values11, w5, acc3, sum3); - - MULTIPLY_ADD_ACCUMULATE(values12, w6, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values13, w7, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values14, w8, acc2, sum2); - MULTIPLY_ADD_ACCUMULATE(values13, w6, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values14, w7, acc3, sum3); - MULTIPLY_ADD_ACCUMULATE(values15, w8, acc3, sum3); - -#if defined(HAS_BIAS) - Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); - - VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); - - acc0 += bias_values; - acc1 += bias_values; - acc2 += bias_values; - acc3 += bias_values; -#endif /* defined(HAS_BIAS) */ - -#if WEIGHTS_OFFSET != 0 - acc0 += WEIGHTS_OFFSET * sum0; - acc1 += WEIGHTS_OFFSET * sum1; - acc2 += WEIGHTS_OFFSET * sum2; - acc3 += WEIGHTS_OFFSET * sum3; -#endif /* WEIGHTS_OFFSET != 0 */ - -#if INPUT_OFFSET != 0 - VEC_INT offs = INPUT_OFFSET * sum_we; - - acc0 += offs; - acc1 += offs; - acc2 += offs; - acc3 += offs; -#endif /* INPUT_OFFSET != 0 */ - -#if K_OFFSET != 0 - acc0 += (VEC_INT)K_OFFSET; - acc1 += (VEC_INT)K_OFFSET; - acc2 += (VEC_INT)K_OFFSET; - acc3 += (VEC_INT)K_OFFSET; -#endif /* K_OFFSET != 0 */ - -#if defined(REAL_MULTIPLIER) - - acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - acc2 = CONVERT(round(CONVERT(acc2, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - acc3 = CONVERT(round(CONVERT(acc3, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - -#else // defined(REAL_MULTIPLIER) - -#if defined(PER_CHANNEL_QUANTIZATION) - Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers); - Vector output_shifts = CONVERT_TO_VECTOR_STRUCT(output_shifts); - VEC_INT output_multiplier = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr); - VEC_INT output_shift = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr); - - VEC_INT res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, output_multiplier, output_shift, VEC_SIZE); - VEC_INT res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, output_multiplier, output_shift, VEC_SIZE); - VEC_INT res2_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, output_multiplier, output_shift, VEC_SIZE); - VEC_INT res3_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, output_multiplier, output_shift, VEC_SIZE); - VEC_INT res0_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, output_multiplier, output_shift); - VEC_INT res1_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc1, output_multiplier, output_shift); - VEC_INT res2_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc2, output_multiplier, output_shift); - VEC_INT res3_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc3, output_multiplier, output_shift); - acc0 = select(res0_shift_lt0, res0_shift_gt0, output_shift >= 0); - acc1 = select(res1_shift_lt0, res1_shift_gt0, output_shift >= 0); - acc2 = select(res2_shift_lt0, res2_shift_gt0, output_shift >= 0); - acc3 = select(res3_shift_lt0, res3_shift_gt0, output_shift >= 0); -#else // defined(PER_CHANNEL_QUANTIZATION) -#if OUTPUT_SHIFT < 0 - acc0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); - acc1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); - acc2 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); - acc3 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); -#else // OUTPUT_SHIFT < 0 - acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); - acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); - acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); - acc3 = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); -#endif // OUTPUT_SHIFT < 0 -#endif // defined(PER_CHANNEL_QUANTIZATION) - -#endif // defined(REAL_MULTIPLIER) - - acc0 += (VEC_INT)OUTPUT_OFFSET; - acc1 += (VEC_INT)OUTPUT_OFFSET; - acc2 += (VEC_INT)OUTPUT_OFFSET; - acc3 += (VEC_INT)OUTPUT_OFFSET; - - VEC_TYPE(VEC_SIZE) - res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE)); - VEC_TYPE(VEC_SIZE) - res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE)); - VEC_TYPE(VEC_SIZE) - res2 = CONVERT_SAT(acc2, VEC_TYPE(VEC_SIZE)); - VEC_TYPE(VEC_SIZE) - res3 = CONVERT_SAT(acc3, VEC_TYPE(VEC_SIZE)); - -#if defined(DST_DEPTH) - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z + b * dst_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z; -#endif /* defined(DST_DEPTH) */ - - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res0), 0, dst_addr + 0 * dst_stride_y); - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res1), 0, dst_addr + 1 * dst_stride_y); - -#if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) - if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2) -#endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) - { - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res2), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z)); - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res3), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z)); - } -} - -#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE == 4 -/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1 using dot product. - * - * @note Per-channel quantization is not supported by this kernel. - * @note This kernel assumes VEC_SIZE is 4. - * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel. - * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) - * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) - * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2) - * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) - * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1). - * @note If REAL_MULTIPLIER is passed at compile time (i.e. -DREAL_MULTIPLIER=1.355f), the final quantization is performed using a floating point multiplication. - * If not, the quantization will be performed using a fixed point multiplication - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32 - * @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes) - * @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector - * @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32 - * @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes) - * @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector - * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32 - * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) - * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector - * @param[in] max_offset The maximum allowed offset for the input tensor - */ -__kernel void dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR4D_DECLARATION(dst), - IMAGE_DECLARATION(weights), - VECTOR_DECLARATION(output_multipliers), - VECTOR_DECLARATION(output_shifts), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(biases), -#endif // defined(HAS_BIAS) - int max_offset) -{ - int x = get_global_id(0); - int y = get_global_id(1); -#if defined(DST_DEPTH) - int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y - int b = get_global_id(2) / (int)DST_DEPTH; // batch -#else // defined(DST_DEPTH) - int z = get_global_id(2); // spatial coordinate y -#endif // defined(DST_DEPTH) - - __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y; - -#if defined(DST_DEPTH) - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; -#endif /* defined(DST_DEPTH) */ - - int z_coord = 0; - int4 offset = 0; - int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT; - - // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1 - y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1); - y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1); - y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1); - y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1); - - int4 y_offset = convert_int4(y_coord * (int)src_stride_y); - - // We compute 4x2x1 [C,W,H] elements - VEC_INT acc0 = 0; - VEC_INT acc1 = 0; - VEC_INT sum0 = 0; - VEC_INT sum1 = 0; - - // Load weights - VEC_TYPE(16) - w0 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr)); - VEC_TYPE(16) - w1 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16)); - VEC_TYPE(4) - w2 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 32)); - -#if INPUT_OFFSET != 0 - // Initilize the final result with the weights reduction multiplied by INPUT_OFFSET - DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s0, w0.s01234567, w0.s8); - DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1); - DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s2, w1.s23456789, w1.sA); - DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3); - - // Multiply the weights reduction with INPUT_OFFSET - acc0 = INPUT_OFFSET * acc0; - - acc1 = acc0; -#endif // INPUT_OFFSET != 0 - - // Load input values - // z == 0 - // Clamp z_coord as for z = 0, it can be negative - // z_coord is casted to unsigned int in order to use just a min() operation - // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 - z_coord = z - (int)CONV_PAD_TOP; - z_coord = min((uint)z_coord, (uint)SRC_DIM_2); - offset = y_offset + (int4)(z_coord * src_stride_z); - offset = min(offset, (int4)max_offset); - - VEC_TYPE(VEC_SIZE) - values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - // z == 1 - // z_coord can be only negative for z = 0 so we do not need to clamp it - // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset - z_coord = z - (int)CONV_PAD_TOP + 1; - offset = y_offset + (int4)(z_coord * src_stride_z); - VEC_TYPE(VEC_SIZE) - values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - // z == 2 - // After z = 1 we can simply add src_stride_z to offset without updating z_coord - // However offset can be out-of-bound so we need to check if it is greater than max_offset - offset += (int4)src_stride_z; - offset = min(offset, (int4)max_offset); - VEC_TYPE(VEC_SIZE) - values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); - VEC_TYPE(VEC_SIZE) - values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); - VEC_TYPE(VEC_SIZE) - values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); - VEC_TYPE(VEC_SIZE) - values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); - - DOT_PRODUCT_REDUCTION(sum0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0); - DOT_PRODUCT_REDUCTION(sum1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0); - DOT_PRODUCT(acc0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0, w0.s01234567, w0.s8); - DOT_PRODUCT(acc1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0, w0.s01234567, w0.s8); - - DOT_PRODUCT_REDUCTION(sum0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1); - DOT_PRODUCT_REDUCTION(sum1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1); - DOT_PRODUCT(acc0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1); - DOT_PRODUCT(acc1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1); - - DOT_PRODUCT_REDUCTION(sum0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2); - DOT_PRODUCT_REDUCTION(sum1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2); - DOT_PRODUCT(acc0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2, w1.s23456789, w1.sA); - DOT_PRODUCT(acc1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2, w1.s23456789, w1.sA); - - DOT_PRODUCT_REDUCTION(sum0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3); - DOT_PRODUCT_REDUCTION(sum1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3); - DOT_PRODUCT(acc0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3); - DOT_PRODUCT(acc1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3); - -#if defined(HAS_BIAS) - Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); - - VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); - - acc0 += bias_values; - acc1 += bias_values; - -#endif // defined(HAS_BIAS) - -#if WEIGHTS_OFFSET != 0 - acc0 += WEIGHTS_OFFSET * sum0; - acc1 += WEIGHTS_OFFSET * sum1; -#endif // WEIGHTS_OFFSET != 0 - -#if K_OFFSET != 0 - acc0 += (VEC_INT)K_OFFSET; - acc1 += (VEC_INT)K_OFFSET; - -#endif // K_OFFSET != 0 - -#if defined(REAL_MULTIPLIER) - - acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT); - -#else // defined(REAL_MULTIPLIER) - -#if OUTPUT_SHIFT < 0 - acc0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); - acc1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE); -#else // OUTPUT_SHIFT < 0 - acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); - acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); -#endif // OUTPUT_SHIFT < 0 - -#endif // defined(REAL_MULTIPLIER) - acc0 += (VEC_INT)OUTPUT_OFFSET; - acc1 += (VEC_INT)OUTPUT_OFFSET; - - VEC_TYPE(VEC_SIZE) - res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE)); - VEC_TYPE(VEC_SIZE) - res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE)); - -#if defined(DST_DEPTH) - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w; -#else /* defined(DST_DEPTH) */ - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z; -#endif /* defined(DST_DEPTH) */ - - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res0), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)); - VSTORE(VEC_SIZE) - (ACTIVATION_FUNC(res1), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)); -} -#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE==4 - -#endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) - #endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) #endif // defined(WEIGHTS_PROMOTED_TYPE) @@ -1612,7 +783,7 @@ __kernel void dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc( #endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER)) #if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_SHIFT) && defined(OUTPUT_MULTIPLIER) && defined(VEC_SIZE_LEFTOVER) -/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped +/** This function computes the depthwise convolution for NHWC data layout. * * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2) * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1) |