aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/core/CL/CLKernelLibrary.cpp1
-rw-r--r--src/core/CL/cl_kernels/winograd.cl310
-rw-r--r--src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp9
-rw-r--r--tests/datasets/ShapeDatasets.h32
-rw-r--r--tests/validation/CL/Winograd.cpp20
-rw-r--r--tests/validation/reference/Winograd.cpp19
6 files changed, 355 insertions, 36 deletions
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 3ff3acd697..4760a1739e 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -356,6 +356,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
{ "warp_perspective_bilinear", "warp_perspective.cl" },
{ "winograd_filter_transform_2x2_3x3_nchw", "winograd.cl" },
{ "winograd_filter_transform_4x4_3x3_nchw", "winograd.cl" },
+ { "winograd_filter_transform_4x4_5x5_nchw", "winograd.cl" },
{ "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd.cl" },
{ "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd.cl" },
{ "winograd_output_transform_2x2_3x3_nchw", "winograd.cl" },
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
index bd51db6b03..03ff377d52 100644
--- a/src/core/CL/cl_kernels/winograd.cl
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -165,48 +165,48 @@ __kernel void winograd_filter_transform_4x4_3x3_nchw(
out0.s0 = (w0.s0) / 16.f;
out0.s1 = (-w0.s0 - w0.s1 - w0.s2) / 24.f;
out0.s2 = (-w0.s0 + w0.s1 - w0.s2) / 24.f;
- out0.s3 = (w0.s0 + 2 * w0.s1 + 4 * w0.s2) / 96.f;
- out0.s4 = (w0.s0 - 2 * w0.s1 + 4 * w0.s2) / 96.f;
+ out0.s3 = (w0.s0 + 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
+ out0.s4 = (w0.s0 - 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
out0.s5 = (w0.s2) / 4.f;
// Row 1
out1.s0 = (-w0.s0 - w1.s0 - w2.s0) / 24.f;
out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
out1.s2 = (w0.s0 + w1.s0 + w2.s0 - w0.s1 - w1.s1 - w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
- out1.s3 = (-w0.s0 - w1.s0 - w2.s0 + 2 * (-w0.s1 - w1.s1 - w2.s1) + 4 * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
- out1.s4 = (-w0.s0 - w1.s0 - w2.s0 + 2 * (w0.s1 + w1.s1 + w2.s1) + 4 * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
+ out1.s3 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (-w0.s1 - w1.s1 - w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
+ out1.s4 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (w0.s1 + w1.s1 + w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
out1.s5 = (-w0.s2 - w1.s2 - w2.s2) / 6.f;
// Row 2
out2.s0 = (-w0.s0 + w1.s0 - w2.s0) / 24.f;
out2.s1 = (w0.s0 - w1.s0 + w2.s0 + w0.s1 - w1.s1 + w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
out2.s2 = (w0.s0 - w1.s0 + w2.s0 - w0.s1 + w1.s1 - w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
- out2.s3 = (-w0.s0 + w1.s0 - w2.s0 + 2 * (-w0.s1 + w1.s1 - w2.s1) + 4 * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
- out2.s4 = (-w0.s0 + w1.s0 - w2.s0 + 2 * (w0.s1 - w1.s1 + w2.s1) + 4 * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
+ out2.s3 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (-w0.s1 + w1.s1 - w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
+ out2.s4 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (w0.s1 - w1.s1 + w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
out2.s5 = (-w0.s2 + w1.s2 - w2.s2) / 6.f;
// Row 3
- out3.s0 = (w0.s0 + 2 * w1.s0 + 4 * w2.s0) / 96.f;
- out3.s1 = (-w0.s0 - 2 * w1.s0 - 4 * w2.s0 - w0.s1 - 2 * w1.s1 - 4 * w2.s1 - w0.s2 - 2 * w1.s2 - 4 * w2.s2) / 144.f;
- out3.s2 = (-w0.s0 - 2 * w1.s0 - 4 * w2.s0 + w0.s1 + 2 * w1.s1 + 4 * w2.s1 - w0.s2 - 2 * w1.s2 - 4 * w2.s2) / 144.f;
- out3.s3 = ((w0.s0 + 2 * w1.s0 + 4 * w2.s0) + 2 * (w0.s1 + 2 * w1.s1 + 4 * w2.s1) + 4 * (w0.s2 + 2 * w1.s2 + 4 * w2.s2)) / 576.f;
- out3.s4 = ((w0.s0 + 2 * w1.s0 + 4 * w2.s0) + 2 * (-w0.s1 - 2 * w1.s1 - 4 * w2.s1) + 4 * (w0.s2 + 2 * w1.s2 + 4 * w2.s2)) / 576.f;
- out3.s5 = (w0.s2 + 2 * w1.s2 + 4 * w2.s2) / 24.f;
+ out3.s0 = (w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
+ out3.s1 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 - 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+ out3.s2 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 + 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+ out3.s3 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 + 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+ out3.s4 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 - 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+ out3.s5 = (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
// Row 4
- out4.s0 = (w0.s0 - 2 * w1.s0 + 4 * w2.s0) / 96.f;
- out4.s1 = (-w0.s0 + 2 * w1.s0 - 4 * w2.s0 - w0.s1 + 2 * w1.s1 - 4 * w2.s1 - w0.s2 + 2 * w1.s2 - 4 * w2.s2) / 144.f;
- out4.s2 = (-w0.s0 + 2 * w1.s0 - 4 * w2.s0 + w0.s1 - 2 * w1.s1 + 4 * w2.s1 - w0.s2 + 2 * w1.s2 - 4 * w2.s2) / 144.f;
- out4.s3 = ((w0.s0 - 2 * w1.s0 + 4 * w2.s0) + 2 * (w0.s1 - 2 * w1.s1 + 4 * w2.s1) + 4 * (w0.s2 - 2 * w1.s2 + 4 * w2.s2)) / 576.f;
- out4.s4 = ((w0.s0 - 2 * w1.s0 + 4 * w2.s0) + 2 * (-w0.s1 + 2 * w1.s1 - 4 * w2.s1) + 4 * (w0.s2 - 2 * w1.s2 + 4 * w2.s2)) / 576.f;
- out4.s5 = (w0.s2 - 2 * w1.s2 + 4 * w2.s2) / 24.f;
+ out4.s0 = (w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
+ out4.s1 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 + 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+ out4.s2 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 - 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+ out4.s3 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 - 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+ out4.s4 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 + 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+ out4.s5 = (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
// Row 5
out5.s0 = (w2.s0) / 4.f;
out5.s1 = (-w2.s0 - w2.s1 - w2.s2) / 6.f;
out5.s2 = (-w2.s0 + w2.s1 - w2.s2) / 6.f;
- out5.s3 = (w2.s0 + 2 * w2.s1 + 4 * w2.s2) / 24.f;
- out5.s4 = (w2.s0 - 2 * w2.s1 + 4 * w2.s2) / 24.f;
+ out5.s3 = (w2.s0 + 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
+ out5.s4 = (w2.s0 - 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
out5.s5 = (w2.s2);
int z = get_global_id(2);
@@ -254,6 +254,276 @@ __kernel void winograd_filter_transform_4x4_3x3_nchw(
*(__global float *)(dst_addr + 34 * dst_stride_z) = out5.s4;
*(__global float *)(dst_addr + 35 * dst_stride_z) = out5.s5;
}
+
+/** This OpenCL kernel performs Winograd filter transform 5x5 when the data format is NCHW and the output tile is 4x4
+ *
+ * @note The number of channels must be passed at compile time using -DNUM_CHANNELS: e.g. -DNUM_CHANNELS=64
+ *
+ * @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_4x4_5x5_nchw(
+ TENSOR4D_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, NUM_CHANNELS);
+
+ const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0);
+
+ // Load the values from the input tensor
+ const char stride_x = 4 * sizeof(float); // Used for accessing the last value in each row
+ const uchar8 stride_y = (uchar8)(0, 1, 2, 3, 4, 0, 0, 0) * (uchar8)src_stride_y;
+
+ float4 w00 = vload4(0, (__global float *)(src_addr + stride_y.s0));
+ float w01 = *((__global float *)(src_addr + stride_y.s0 + stride_x));
+ float4 w10 = vload4(0, (__global float *)(src_addr + stride_y.s1));
+ float w11 = *((__global float *)(src_addr + stride_y.s1 + stride_x));
+ float4 w20 = vload4(0, (__global float *)(src_addr + stride_y.s2));
+ float w21 = *((__global float *)(src_addr + stride_y.s2 + stride_x));
+ float4 w30 = vload4(0, (__global float *)(src_addr + stride_y.s3));
+ float w31 = *((__global float *)(src_addr + stride_y.s3 + stride_x));
+ float4 w40 = vload4(0, (__global float *)(src_addr + stride_y.s4));
+ float w41 = *((__global float *)(src_addr + stride_y.s4 + stride_x));
+
+ // 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;
+
+ // Row 0
+ out0.s0 = w00.s0;
+ out0.s1 = -2.f * (w00.s0 + w00.s1 + w00.s2 + w00.s3 + w01) / 9.f;
+ out0.s2 = -2.f * (w00.s0 - w00.s1 + w00.s2 - w00.s3 + w01) / 9.f;
+ out0.s3 = (w00.s0 + 2.f * w00.s1 + 4.f * w00.s2 + 8.f * w00.s3 + 16.f * w01) / 90.f;
+ out0.s4 = (w00.s0 - 2.f * w00.s1 + 4.f * w00.s2 - 8.f * w00.s3 + 16.f * w01) / 90.f;
+ out0.s5 = (16.f * w00.s0 + 8.f * w00.s1 + 4.f * w00.s2 + 2.f * w00.s3 + w01) / 180.f;
+ out0.s6 = (16.f * w00.s0 - 8.f * w00.s1 + 4.f * w00.s2 - 2.f * w00.s3 + w01) / 180.f;
+ out0.s7 = w01;
+
+ // Row 1
+ out1.s0 = -2.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) / 9.f;
+ out1.s1 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) +
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
+ out1.s2 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) -
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
+ out1.s3 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 8.f *
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
+ out1.s4 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 8.f *
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
+ out1.s5 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 2.f *
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
+ out1.s6 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 2.f *
+ (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
+ out1.s7 = -2.f * (w01 + w11 + w21 + w31 + w41) / 9.f;
+
+ // Row 2
+ out2.s0 = -2.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) / 9.f;
+ out2.s1 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) +
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
+ out2.s2 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) -
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
+ out2.s3 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 8.f *
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
+ out2.s4 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 8.f *
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
+ out2.s5 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 2.f *
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
+ out2.s6 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 2.f *
+ (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
+ out2.s7 = -2.f * (w01 - w11 + w21 - w31 + w41) / 9.f;
+
+ // Row 3
+ out3.s0 = (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
+ out3.s1 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
+ out3.s2 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
+ out3.s3 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
+ out3.s4 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
+ out3.s5 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
+ out3.s6 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
+ out3.s7 = (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) / 90.f;
+
+ // Row 4
+ out4.s0 = (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
+ out4.s1 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
+ out4.s2 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
+ out4.s3 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
+ out4.s4 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
+ out4.s5 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
+ out4.s6 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+ (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+ (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
+ out4.s7 = (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) / 90.f;
+
+ // Row 5
+ out5.s0 = (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) / 180.f;
+ out5.s1 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
+ out5.s2 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
+ out5.s3 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
+ out5.s4 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
+ out5.s5 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
+ out5.s6 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
+ out5.s7 = (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) / 180.f;
+
+ // Row 6
+ out6.s0 = (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) / 180.f;
+ out6.s1 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
+ out6.s2 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
+ out6.s3 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
+ out6.s4 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
+ out6.s5 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
+ out6.s6 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s0 + w40.s1) + 4.f *
+ (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+ (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
+ out6.s7 = (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) / 180.f;
+
+ // Row 7
+ out7.s0 = w40.s0;
+ out7.s1 = -2.f * (w40.s0 + w40.s1 + w40.s2 + w40.s3 + w41) / 9.f;
+ out7.s2 = -2.f * (w40.s0 - w40.s1 + w40.s2 - w40.s3 + w41) / 9.f;
+ out7.s3 = (w40.s0 + 2.f * w40.s1 + 4.f * w40.s2 + 8.f * w40.s3 + 16.f * w41) / 90.f;
+ out7.s4 = (w40.s0 - 2.f * w40.s1 + 4.f * w40.s2 - 8.f * w40.s3 + 16.f * w41) / 90.f;
+ out7.s5 = (16.f * w40.s0 + 8.f * w40.s1 + 4.f * w40.s2 + 2.f * w40.s3 + w41) / 180.f;
+ out7.s6 = (16.f * w40.s0 - 8.f * w40.s1 + 4.f * w40.s2 - 2.f * w40.s3 + w41) / 180.f;
+ out7.s7 = w41;
+
+ int z = get_global_id(2);
+ int x0 = z / NUM_CHANNELS; // idx filter
+ int y0 = z % NUM_CHANNELS; // idx channel
+
+ // Get output address
+ __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + 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;
+ *(__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;
+ *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s3;
+ *(__global float *)(dst_addr + 12 * dst_stride_z) = out1.s4;
+ *(__global float *)(dst_addr + 13 * dst_stride_z) = out1.s5;
+ *(__global float *)(dst_addr + 14 * dst_stride_z) = out1.s6;
+ *(__global float *)(dst_addr + 15 * dst_stride_z) = out1.s7;
+ *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s0;
+ *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s1;
+ *(__global float *)(dst_addr + 18 * dst_stride_z) = out2.s2;
+ *(__global float *)(dst_addr + 19 * dst_stride_z) = out2.s3;
+ *(__global float *)(dst_addr + 20 * dst_stride_z) = out2.s4;
+ *(__global float *)(dst_addr + 21 * dst_stride_z) = out2.s5;
+ *(__global float *)(dst_addr + 22 * dst_stride_z) = out2.s6;
+ *(__global float *)(dst_addr + 23 * dst_stride_z) = out2.s7;
+ *(__global float *)(dst_addr + 24 * dst_stride_z) = out3.s0;
+ *(__global float *)(dst_addr + 25 * dst_stride_z) = out3.s1;
+ *(__global float *)(dst_addr + 26 * dst_stride_z) = out3.s2;
+ *(__global float *)(dst_addr + 27 * dst_stride_z) = out3.s3;
+ *(__global float *)(dst_addr + 28 * dst_stride_z) = out3.s4;
+ *(__global float *)(dst_addr + 29 * dst_stride_z) = out3.s5;
+ *(__global float *)(dst_addr + 30 * dst_stride_z) = out3.s6;
+ *(__global float *)(dst_addr + 31 * dst_stride_z) = out3.s7;
+ *(__global float *)(dst_addr + 32 * dst_stride_z) = out4.s0;
+ *(__global float *)(dst_addr + 33 * dst_stride_z) = out4.s1;
+ *(__global float *)(dst_addr + 34 * dst_stride_z) = out4.s2;
+ *(__global float *)(dst_addr + 35 * dst_stride_z) = out4.s3;
+ *(__global float *)(dst_addr + 36 * dst_stride_z) = out4.s4;
+ *(__global float *)(dst_addr + 37 * dst_stride_z) = out4.s5;
+ *(__global float *)(dst_addr + 38 * dst_stride_z) = out4.s6;
+ *(__global float *)(dst_addr + 39 * dst_stride_z) = out4.s7;
+ *(__global float *)(dst_addr + 40 * dst_stride_z) = out5.s0;
+ *(__global float *)(dst_addr + 41 * dst_stride_z) = out5.s1;
+ *(__global float *)(dst_addr + 42 * dst_stride_z) = out5.s2;
+ *(__global float *)(dst_addr + 43 * dst_stride_z) = out5.s3;
+ *(__global float *)(dst_addr + 44 * dst_stride_z) = out5.s4;
+ *(__global float *)(dst_addr + 45 * dst_stride_z) = out5.s5;
+ *(__global float *)(dst_addr + 46 * dst_stride_z) = out5.s6;
+ *(__global float *)(dst_addr + 47 * dst_stride_z) = out5.s7;
+ *(__global float *)(dst_addr + 48 * dst_stride_z) = out6.s0;
+ *(__global float *)(dst_addr + 49 * dst_stride_z) = out6.s1;
+ *(__global float *)(dst_addr + 50 * dst_stride_z) = out6.s2;
+ *(__global float *)(dst_addr + 51 * dst_stride_z) = out6.s3;
+ *(__global float *)(dst_addr + 52 * dst_stride_z) = out6.s4;
+ *(__global float *)(dst_addr + 53 * dst_stride_z) = out6.s5;
+ *(__global float *)(dst_addr + 54 * dst_stride_z) = out6.s6;
+ *(__global float *)(dst_addr + 55 * dst_stride_z) = out6.s7;
+ *(__global float *)(dst_addr + 56 * dst_stride_z) = out7.s0;
+ *(__global float *)(dst_addr + 57 * dst_stride_z) = out7.s1;
+ *(__global float *)(dst_addr + 58 * dst_stride_z) = out7.s2;
+ *(__global float *)(dst_addr + 59 * dst_stride_z) = out7.s3;
+ *(__global float *)(dst_addr + 60 * dst_stride_z) = out7.s4;
+ *(__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(NUM_CHANNELS)
#if defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP)
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
index 5b8921b8e4..d3a33c01f9 100644
--- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -55,10 +55,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
- ARM_COMPUTE_RETURN_ERROR_ON(kernel_size != Size2D(3U, 3U));
+ ARM_COMPUTE_RETURN_ERROR_ON(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U));
+ ARM_COMPUTE_RETURN_ERROR_ON(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U) && output_tile_size != Size2D(4U, 4U));
+ ARM_COMPUTE_RETURN_ERROR_ON(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U));
ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != kernel_size.width || input->dimension(idx_h) != kernel_size.height);
ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
- ARM_COMPUTE_RETURN_ERROR_ON(output_tile_size != Size2D(2U, 2U) && output_tile_size != Size2D(4U, 4U));
// Checks performed when output is configured
if(output->total_size() != 0)
@@ -76,8 +77,8 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
- constexpr unsigned int num_elems_processed_per_iteration_x = 3;
- constexpr unsigned int num_elems_processed_per_iteration_y = 3;
+ const unsigned int num_elems_processed_per_iteration_x = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH));
+ const unsigned int num_elems_processed_per_iteration_y = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT));
Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
bool window_changed = false;
diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 8255672430..76cf9a1835 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -405,6 +405,38 @@ public:
}
};
+/** Data set containing small 5x5 tensor shapes. */
+class Small5x5Shapes final : public ShapeDataset
+{
+public:
+ Small5x5Shapes()
+ : ShapeDataset("Shape",
+ {
+ TensorShape{ 5U, 5U, 7U, 4U },
+ TensorShape{ 5U, 5U, 4U, 13U },
+ TensorShape{ 5U, 5U, 9U, 2U },
+ TensorShape{ 5U, 5U, 3U, 5U },
+ })
+ {
+ }
+};
+
+/** Data set containing large 5x5 tensor shapes. */
+class Large5x5Shapes final : public ShapeDataset
+{
+public:
+ Large5x5Shapes()
+ : ShapeDataset("Shape",
+ {
+ TensorShape{ 5U, 5U, 32U, 64U },
+ TensorShape{ 5U, 5U, 51U, 13U },
+ TensorShape{ 5U, 5U, 53U, 47U },
+ TensorShape{ 5U, 5U, 128U, 384U },
+ })
+ {
+ }
+};
+
/** Data set containing small tensor shapes for deconvolution. */
class SmallDeconvolutionShapes final : public ShapeDataset
{
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
index 8fa5826470..08810eaf58 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -201,19 +201,23 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
winograd_filter_transform.configure(&a, &b, winograd_info);
}
-FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small3x3Shapes(),
- framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::ALL,
+ combine(combine(framework::dataset::concat(framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", Size2D(2U, 2U))), combine(datasets::Small3x3Shapes(),
+ framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
+ combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large3x3Shapes(),
- framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY,
+ combine(combine(framework::dataset::concat(framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", Size2D(2U, 2U))), combine(datasets::Large3x3Shapes(),
+ framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
+ combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index 604e25214b..757a06d847 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -43,9 +43,6 @@ namespace
template <typename T>
void initialize_matrix_transform(SimpleTensor<T> &src, const Size2D &output_tile_size, const Size2D &kernel_size, WinogradTransformType winograd_transform_type)
{
- ARM_COMPUTE_ERROR_ON((output_tile_size != Size2D(2U, 2U)) && (output_tile_size != Size2D(4U, 4U)));
- ARM_COMPUTE_ERROR_ON(kernel_size != Size2D(3U, 3U));
-
// Winograd input transform matrices
static const float imatrix2x2_3x3[] =
{
@@ -86,6 +83,19 @@ void initialize_matrix_transform(SimpleTensor<T> &src, const Size2D &output_tile
0.0f, 0.0f, 1.0f
};
+ static const float fmatrix4x4_5x5[] =
+ {
+ 1.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+ -2.0f / 9.0f, -2.0f / 9.0f, -2.0f / 9.0f, -2.0f / 9.0f, -2.0f / 9.0f,
+ -2.0f / 9.0f, 2.0f / 9.0f, -2.0f / 9.0f, 2.0f / 9.0f, -2.0f / 9.0f,
+ 1.0f / 90.0f, 1.0f / 45.0f, 2.0f / 45.0f, 4.0f / 45.0f, 8.0f / 45.0f,
+ 1.0f / 90.0f, -1.0f / 45.0f, 2.0f / 45.0f, -4.0f / 45.0f, 8.0f / 45.0f,
+ 4.0f / 45.0f, 2.0f / 45.0f, 1.0f / 45.0f, 1.0f / 90.0f, 1.0f / 180.0f,
+ 4.0f / 45.0f, -2.0f / 45.0f, 1.0f / 45.0f, -1.0f / 90.0f, 1.0f / 180.0f,
+ 0.0f, 0.0f, 0.0f, 0.0f, 1.0f
+
+ };
+
// ------------------------------------------
// Winograd output transform matrices
@@ -114,11 +124,12 @@ void initialize_matrix_transform(SimpleTensor<T> &src, const Size2D &output_tile
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::INPUT), imatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::FILTER), fmatrix4x4_5x5 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
};
- // Find input matrix transform
+ // Find transformation matrix
std::map<WinogradKey, const float *>::iterator it;
it = matrix_map.find(WinogradKey(std::pair<int, int>(output_tile_size.width, output_tile_size.height),
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-25 06:44:28 +0000