diff options
Diffstat (limited to 'src/core/CL/cl_kernels/winograd_filter_transform.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/winograd_filter_transform.cl | 550 |
1 files changed, 330 insertions, 220 deletions
diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl index e53da9b278..73da005996 100644 --- a/src/core/CL/cl_kernels/winograd_filter_transform.cl +++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl @@ -442,105 +442,6 @@ __kernel void winograd_filter_transform_4x4_3x3_nhwc( #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } -#if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) -/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1 - * - * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 - * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @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 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[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] 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 Z 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 winograd_filter_transform_4x1_3x1_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_filter_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_stride_w, - src_step_w, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) - -#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) -/** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4 - * - * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 - * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @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 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[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] 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 Z 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 winograd_filter_transform_1x4_1x3_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_filter_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_stride_w, - src_step_w, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) /** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NCHW and the output tile is 4x4/4x1 or 1x4 * * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 @@ -827,9 +728,11 @@ __kernel void winograd_filter_transform_4x4_5x5_nchw( #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } -/** This OpenCL kernel performs Winograd filter transform 5x5 when the data layout is NHWC and the output tile is 4x4 +/** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NHWC and the output tile is 4x4/4x1 or 1x4 * * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time + * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) @@ -858,12 +761,23 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(float) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w; +#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Load the values from the input tensor - float w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y)); - float w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y)); - float w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y)); - float w03 = *((__global float *)(src_addr + 0 * src_stride_z + 3 * src_stride_y)); - float w04 = *((__global float *)(src_addr + 0 * src_stride_z + 4 * src_stride_y)); + float w00 = *((__global float *)(src_addr + 0 * src_stride_z)); + float w01 = *((__global float *)(src_addr + 1 * src_stride_z)); + float w02 = *((__global float *)(src_addr + 2 * src_stride_z)); + float w03 = *((__global float *)(src_addr + 3 * src_stride_z)); + float w04 = *((__global float *)(src_addr + 4 * src_stride_z)); +#else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) + // Load the values from the input tensor + float w00 = *((__global float *)(src_addr + 0 * src_stride_y)); + float w01 = *((__global float *)(src_addr + 1 * src_stride_y)); + float w02 = *((__global float *)(src_addr + 2 * src_stride_y)); + float w03 = *((__global float *)(src_addr + 3 * src_stride_y)); + float w04 = *((__global float *)(src_addr + 4 * src_stride_y)); +#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) + +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y)); float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y)); float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y)); @@ -884,128 +798,129 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( float w42 = *((__global float *)(src_addr + 4 * src_stride_z + 2 * src_stride_y)); float w43 = *((__global float *)(src_addr + 4 * src_stride_z + 3 * src_stride_y)); float w44 = *((__global float *)(src_addr + 4 * src_stride_z + 4 * src_stride_y)); - - // Transform the 3x3 tile in a 8x8 tile - float8 out0 = 0.0f; - float8 out1 = 0.0f; - float8 out2 = 0.0f; - float8 out3 = 0.0f; - float8 out4 = 0.0f; - float8 out5 = 0.0f; - float8 out6 = 0.0f; - float8 out7 = 0.0f; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Row 0 - out0.s0 = w00; - out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f; - out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f; - out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f; - out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f; - out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f; - out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f; - out0.s7 = w04; + float8 out0 = 0.0f; + out0.s0 = w00; + out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f; + out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f; + out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f; + out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f; + out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f; + out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f; + out0.s7 = w04; +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Row 1 - out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f; - out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; - out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; - out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * - (w04 + w14 + w24 + w34 + w44)) / 405.f; - out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * - (w04 + w14 + w24 + w34 + w44)) / 405.f; - out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) + - (w04 + w14 + w24 + w34 + w44)) / 810.f; - out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) + - (w04 + w14 + w24 + w34 + w44)) / 810.f; - out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f; + float8 out1 = 0.0f; + out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f; + out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; + out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; + out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * + (w04 + w14 + w24 + w34 + w44)) / 405.f; + out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * + (w04 + w14 + w24 + w34 + w44)) / 405.f; + out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) + + (w04 + w14 + w24 + w34 + w44)) / 810.f; + out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) + + (w04 + w14 + w24 + w34 + w44)) / 810.f; + out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f; // Row 2 - out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f; - out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; - out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; - out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * - (w04 - w14 + w24 - w34 + w44)) / 405.f; - out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * - (w04 - w14 + w24 - w34 + w44)) / 405.f; - out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) + - (w04 - w14 + w24 - w34 + w44)) / 810.f; - out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) + - (w04 - w14 + w24 - w34 + w44)) / 810.f; - out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f; + float8 out2 = 0.0f; + out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f; + out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; + out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; + out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * + (w04 - w14 + w24 - w34 + w44)) / 405.f; + out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * + (w04 - w14 + w24 - w34 + w44)) / 405.f; + out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) + + (w04 - w14 + w24 - w34 + w44)) / 810.f; + out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) + + (w04 - w14 + w24 - w34 + w44)) / 810.f; + out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f; // Row 3 - out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f; - out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + - (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; - out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - - (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; - out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f - * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; - out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f - * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; - out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * - (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; - out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * - (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; - out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f; + float8 out3 = 0.0f; + out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f; + out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + + (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; + out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - + (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; + out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; + out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; + out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; + out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; + out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f; // Row 4 - out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f; - out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + - (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; - out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - - (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; - out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f - * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; - out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f - * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; - out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * - (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; - out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * - (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; - out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f; + float8 out4 = 0.0f; + out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f; + out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + + (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; + out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - + (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; + out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; + out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; + out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; + out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; + out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f; // Row 5 - out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f; - out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + - (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; - out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - - (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; - out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f - * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; - out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f - * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; - out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * - (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; - out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * - (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; - out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f; + float8 out5 = 0.0f; + out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f; + out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + + (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; + out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - + (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; + out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; + out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; + out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; + out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; + out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f; // Row 6 - out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f; - out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + - (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; - out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - - (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; - out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f - * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; - out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f - * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; - out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * - (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; - out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * - (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; - out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f; + float8 out6 = 0.0f; + out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f; + out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + + (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; + out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - + (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; + out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; + out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; + out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; + out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; + out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f; // Row 7 - out7.s0 = w40; - out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f; - out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f; - out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f; - out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f; - out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f; - out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f; - out7.s7 = w44; + float8 out7 = 0.0f; + out7.s0 = w40; + out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f; + out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f; + out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f; + out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f; + out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f; + out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f; + out7.s7 = w44; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) int x0 = get_global_id(2); // idx filter int y0 = get_global_id(0); // idx channel @@ -1013,15 +928,17 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( // Get output address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y; - // Store the 64 values across the 64 channels - *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; - *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; - *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; - *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; - *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; - *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; - *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; - *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; + // Store the values across the channels + *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; + *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; + *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; + *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; + *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; + *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; + *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; + *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; + +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) *(__global float *)(dst_addr + 8 * dst_stride_z) = out1.s0; *(__global float *)(dst_addr + 9 * dst_stride_z) = out1.s1; *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2; @@ -1078,6 +995,7 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5; *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6; *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } #endif // defined(SRC_DIM_Z) @@ -1225,6 +1143,102 @@ __kernel void winograd_filter_transform_4x1_5x1_nchw( dst_step_z, dst_offset_first_element_in_bytes); } + +/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @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 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[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] 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 Z 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 winograd_filter_transform_4x1_3x1_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel performs Winograd filter transform 5x1 when the data layout is NHWC and the output tile is 4x1 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @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 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[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] 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 Z 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 winograd_filter_transform_4x1_5x1_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) @@ -1371,4 +1385,100 @@ __kernel void winograd_filter_transform_1x4_1x5_nchw( dst_step_z, dst_offset_first_element_in_bytes); } + +/** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @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 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[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] 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 Z 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 winograd_filter_transform_1x4_1x3_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel performs Winograd filter transform 1x5 when the data layout is NHWC and the output tile is 1x4 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @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 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[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] 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 Z 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 winograd_filter_transform_1x4_1x5_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |