COMPMID-1339 - Implementing Winograd Convolution Layer 1x5 and 5x1 kernels on OpenCL NCHW

Change-Id: Ia293cd89651146a0e27e5f7c74ca9c924807e83c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138707
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 472 insertions, 360 deletions

diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl
index 1ee6981a07..5f528d4b0e 100644
--- a/src/core/CL/cl_kernels/winograd_filter_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl
@@ -66,9 +66,9 @@ __kernel void winograd_filter_transform_2x2_3x3_nchw(
                          *((__global float *)(src_addr + 1 * src_stride_y)),
                          *((__global float *)(src_addr + 2 * src_stride_y)));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float3 w0  = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float3 w1  = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float3 w2  = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
     // Row 0
@@ -173,9 +173,9 @@ __kernel void winograd_filter_transform_4x4_3x3_nchw(
                          *((__global float *)(src_addr + 1 * src_stride_y)),
                          *((__global float *)(src_addr + 2 * src_stride_y)));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float3 w0  = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float3 w1  = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float3 w2  = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
     // Row 0
@@ -285,202 +285,6 @@ __kernel void winograd_filter_transform_4x4_3x3_nchw(
 #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 NCHW and the output tile is 2x1
- *
- * @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_2x1_3x1_nchw(
-    TENSOR4D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
-{
-    winograd_filter_transform_2x2_3x3_nchw(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 3x1 when the data layout is NCHW 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_nchw(
-    TENSOR4D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
-{
-    winograd_filter_transform_4x4_3x3_nchw(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 NCHW and the output tile is 1x2
- *
- * @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_1x2_1x3_nchw(
-    TENSOR4D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
-{
-    winograd_filter_transform_2x2_3x3_nchw(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 1x3 when the data layout is NCHW 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_nchw(
-    TENSOR4D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
-{
-    winograd_filter_transform_4x4_3x3_nchw(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 3x3 when the data layout is NHWC and the output tile is 4x4
  *
  * @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
@@ -630,10 +434,13 @@ __kernel void winograd_filter_transform_4x4_3x3_nhwc(
     *(__global float *)(dst_addr + 34 * dst_stride_z) = out54;
     *(__global float *)(dst_addr + 35 * dst_stride_z) = out55;
 }
-/** This OpenCL kernel performs Winograd filter transform 5x5 when the data layout is NCHW and the output tile is 4x4
+/** 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
  *
+ * @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)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
@@ -662,186 +469,198 @@ __kernel void winograd_filter_transform_4x4_5x5_nchw(
     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));
+#if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+    float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float  w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4);
+#elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    float4 w00 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
+                          *((__global float *)(src_addr + 1 * src_stride_y)),
+                          *((__global float *)(src_addr + 2 * src_stride_y)),
+                          *((__global float *)(src_addr + 3 * src_stride_y)));
+    float w01 = *((__global float *)(src_addr + 4 * src_stride_y));
+#else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float  w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4);
+    float4 w10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float  w11 = *((__global float *)(src_addr + 1 * src_stride_y) + 4);
+    float4 w20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float  w21 = *((__global float *)(src_addr + 2 * src_stride_y) + 4);
+    float4 w30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+    float  w31 = *((__global float *)(src_addr + 3 * src_stride_y) + 4);
+    float4 w40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y));
+    float  w41 = *((__global float *)(src_addr + 4 * src_stride_y) + 4);
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
-    // 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;
+    // Transform the input tile
 
     // 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;
+    float8 out0 = 0.0f;
+    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;
 
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // 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;
+    float8 out1 = 0.0f;
+    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;
+    float8 out2 = 0.0f;
+    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;
+    float8 out3 = 0.0f;
+    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;
+    float8 out4 = 0.0f;
+    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.s1 + 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.s1 + 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.s1 + 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.s1 + 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;
+    float8 out5 = 0.0f;
+    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.s1 + 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.s1 + 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.s1 + 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.s1 + 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.s1 + 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.s1 + 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.s1 + 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.s1 + 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;
+    float8 out6 = 0.0f;
+    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.s1 + 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.s1 + 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.s1 + 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.s1 + 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;
+    float8 out7 = 0.0f;
+    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;
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     int z  = get_global_id(2);
     int x0 = z / SRC_DIM_Z; // idx filter
     int y0 = z % SRC_DIM_Z; // 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;
+    __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;
@@ -898,6 +717,7 @@ __kernel void winograd_filter_transform_4x4_5x5_nchw(
     *(__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)
 }
 
 /** This OpenCL kernel performs Winograd filter transform 5x5 when the data layout is NHWC and the output tile is 4x4
@@ -1152,4 +972,296 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc(
     *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6;
     *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7;
 }
-#endif // defined(SRC_DIM_Z)
\ No newline at end of file
+#endif // defined(SRC_DIM_Z)
+
+#if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NCHW and the output tile is 2x1
+ *
+ * @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_2x1_3x1_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_2x2_3x3_nchw(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 3x1 when the data layout is NCHW 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_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_4x4_3x3_nchw(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 NCHW 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_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_4x4_5x5_nchw(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 NCHW and the output tile is 1x2
+ *
+ * @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_1x2_1x3_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_2x2_3x3_nchw(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 1x3 when the data layout is NCHW 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_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_4x4_3x3_nchw(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 NCHW 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_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_4x4_5x5_nchw(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)
\ No newline at end of file