COMPMID-1318: Implementing Winograd 7x7 NHWC on OpenCL -- Part II

Change-Id: I036558d832c697da1fe9ea04ada0df38dc793914
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/923
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 528 insertions, 7 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 31e9a8acf5..cf6d4c9843 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -488,6 +488,9 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "winograd_filter_transform_4x4_5x5_nhwc", "winograd_filter_transform.cl" },
     { "winograd_filter_transform_4x1_5x1_nhwc", "winograd_filter_transform.cl" },
     { "winograd_filter_transform_1x4_1x5_nhwc", "winograd_filter_transform.cl" },
+    { "winograd_filter_transform_2x2_7x7_nhwc", "winograd_filter_transform.cl" },
+    { "winograd_filter_transform_2x1_7x1_nhwc", "winograd_filter_transform.cl" },
+    { "winograd_filter_transform_1x2_1x7_nhwc", "winograd_filter_transform.cl" },
     { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd_input_transform.cl" },
     { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd_input_transform.cl" },
     { "winograd_input_transform_2x1_3x1_stepz1_nchw", "winograd_input_transform.cl" },
diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl
index 3b9b1e918e..3f203b8a1e 100644
--- a/src/core/CL/cl_kernels/winograd_filter_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,6 +25,18 @@
 
 #if defined(SRC_DIM_Z)
 
+#define OUTPUT_ROW_2x2_7x7(out, tmp)                                                                                               \
+    ({                                                                                                                             \
+        out.s0 = -tmp.s0 / 36.f;                                                                                                   \
+        out.s1 = (tmp.s0 - tmp.s1 + tmp.s2 - tmp.s3 + tmp.s4 - tmp.s5 + tmp.s6) / 48.f;                                            \
+        out.s2 = (tmp.s0 + tmp.s1 + tmp.s2 + tmp.s3 + tmp.s4 + tmp.s5 + tmp.s6) / 48.f;                                            \
+        out.s3 = (-tmp.s0 + 2.f * tmp.s1 - 4.f * tmp.s2 + 8.f * tmp.s3 - 16.f * tmp.s4 + 32.f * tmp.s5 - 64.f * tmp.s6) / 120.f;   \
+        out.s4 = (-tmp.s0 - 2.f * tmp.s1 - 4.f * tmp.s2 - 8.f * tmp.s3 - 16.f * tmp.s4 - 32.f * tmp.s5 - 64.f * tmp.s6) / 120.f;   \
+        out.s5 = (tmp.s0 - 3.f * tmp.s1 + 9.f * tmp.s2 - 27.f * tmp.s3 + 81.f * tmp.s4 - 243.f * tmp.s5 + 729.f * tmp.s6) / 720.f; \
+        out.s6 = (tmp.s0 + 3.f * tmp.s1 + 9.f * tmp.s2 + 27.f * tmp.s3 + 81.f * tmp.s4 + 243.f * tmp.s5 + 729.f * tmp.s6) / 720.f; \
+        out.s7 = tmp.s6;                                                                                                           \
+    })
+
 /** This OpenCL kernel performs Winograd filter transform 3x3/3x1/1x3 when the data layout is NCHW and the output tile is 2x2/2x1/1x2
  *
  * @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
@@ -1045,6 +1057,306 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc(
     *(__global DATA_TYPE *)(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 7x7/7x1 or 1x7 when the data layout is NHWC and the output tile is 2x2/2x1 or 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 If this kernel is used to perform Winograd filter transform 7x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note If this kernel is used to perform Winograd filter transform 1x7, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
+ * @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_2x2_7x7_nhwc(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z);
+
+    const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(DATA_TYPE) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w;
+
+#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    // Load the values from the input tensor
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE w03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE w04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE w05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE w06 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+#else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    // Load the values from the input tensor
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    DATA_TYPE w03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
+    DATA_TYPE w04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
+    DATA_TYPE w05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_y));
+    DATA_TYPE w06 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_y));
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    DATA_TYPE w10 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w11 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w12 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w13 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w14 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w15 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w16 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 6 * src_stride_y));
+
+    DATA_TYPE w20 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w21 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w22 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w23 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w24 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w25 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w26 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 6 * src_stride_y));
+
+    DATA_TYPE w30 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w31 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w32 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w33 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w34 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w35 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w36 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 6 * src_stride_y));
+
+    DATA_TYPE w40 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w41 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w42 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w43 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w44 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w45 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w46 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 6 * src_stride_y));
+
+    DATA_TYPE w50 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w51 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w52 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w53 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w54 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w55 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w56 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z + 6 * src_stride_y));
+
+    DATA_TYPE w60 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w61 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w62 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w63 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w64 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w65 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 5 * src_stride_y));
+    DATA_TYPE w66 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z + 6 * src_stride_y));
+
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    tmp = 0.0f;
+
+    // Row 0
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0 = 0.0f;
+
+    out0.s0 = -w00 / 36.0f;
+    out0.s1 = (w00 - w01 + w02 - w03 + w04 - w05 + w06) / 48.f;
+    out0.s2 = (w00 + w01 + w02 + w03 + w04 + w05 + w06) / 48.f;
+    out0.s3 = (-w00 + 2.f * w01 - 4.f * w02 + 8.f * w03 - 16.f * w04 + 32.f * w05 - 64.f * w06) / 120.f;
+    out0.s4 = (-w00 - 2.f * w01 - 4.f * w02 - 8.f * w03 - 16.f * w04 - 32.f * w05 - 64.f * w06) / 120.f;
+    out0.s5 = (w00 - 3.f * w01 + 9.f * w02 - 27.f * w03 + 81.f * w04 - 243.f * w05 + 729.f * w06) / 720.f;
+    out0.s6 = (w00 + 3.f * w01 + 9.f * w02 + 27.f * w03 + 81.f * w04 + 243.f * w05 + 729.f * w06) / 720.f;
+    out0.s7 = w06;
+
+    out0 /= (VEC_DATA_TYPE(DATA_TYPE, 8)) - 36.f;
+
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+
+    // Row 1
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out1 = 0.0f;
+
+    tmp.s0 = (w00 - w10 + w20 - w30 + w40 - w50 + w60) / 48.f;
+    tmp.s1 = (w01 - w11 + w21 - w31 + w41 - w51 + w61) / 48.f;
+    tmp.s2 = (w02 - w12 + w22 - w32 + w42 - w52 + w62) / 48.f;
+    tmp.s3 = (w03 - w13 + w23 - w33 + w43 - w53 + w63) / 48.f;
+    tmp.s4 = (w04 - w14 + w24 - w34 + w44 - w54 + w64) / 48.f;
+    tmp.s5 = (w05 - w15 + w25 - w35 + w45 - w55 + w65) / 48.f;
+    tmp.s6 = (w06 - w16 + w26 - w36 + w46 - w56 + w66) / 48.f;
+
+    OUTPUT_ROW_2x2_7x7(out1, tmp);
+
+    // Row 2
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out2 = 0.0f;
+
+    tmp.s0 = (w00 + w10 + w20 + w30 + w40 + w50 + w60) / 48.f;
+    tmp.s1 = (w01 + w11 + w21 + w31 + w41 + w51 + w61) / 48.f;
+    tmp.s2 = (w02 + w12 + w22 + w32 + w42 + w52 + w62) / 48.f;
+    tmp.s3 = (w03 + w13 + w23 + w33 + w43 + w53 + w63) / 48.f;
+    tmp.s4 = (w04 + w14 + w24 + w34 + w44 + w54 + w64) / 48.f;
+    tmp.s5 = (w05 + w15 + w25 + w35 + w45 + w55 + w65) / 48.f;
+    tmp.s6 = (w06 + w16 + w26 + w36 + w46 + w56 + w66) / 48.f;
+
+    OUTPUT_ROW_2x2_7x7(out2, tmp);
+
+    // Row 3
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out3 = 0.0f;
+
+    tmp.s0 = (-w00 + 2.f * w10 - 4.f * w20 + 8.f * w30 - 16.f * w40 + 32.f * w50 - 64.f * w60) / 120.f;
+    tmp.s1 = (-w01 + 2.f * w11 - 4.f * w21 + 8.f * w31 - 16.f * w41 + 32.f * w51 - 64.f * w61) / 120.f;
+    tmp.s2 = (-w02 + 2.f * w12 - 4.f * w22 + 8.f * w32 - 16.f * w42 + 32.f * w52 - 64.f * w62) / 120.f;
+    tmp.s3 = (-w03 + 2.f * w13 - 4.f * w23 + 8.f * w33 - 16.f * w43 + 32.f * w53 - 64.f * w63) / 120.f;
+    tmp.s4 = (-w04 + 2.f * w14 - 4.f * w24 + 8.f * w34 - 16.f * w44 + 32.f * w54 - 64.f * w64) / 120.f;
+    tmp.s5 = (-w05 + 2.f * w15 - 4.f * w25 + 8.f * w35 - 16.f * w45 + 32.f * w55 - 64.f * w65) / 120.f;
+    tmp.s6 = (-w06 + 2.f * w16 - 4.f * w26 + 8.f * w36 - 16.f * w46 + 32.f * w56 - 64.f * w66) / 120.f;
+
+    OUTPUT_ROW_2x2_7x7(out3, tmp);
+
+    // Row 4
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out4 = 0.0f;
+
+    tmp.s0 = (-w00 - 2.f * w10 - 4.f * w20 - 8.f * w30 - 16.f * w40 - 32.f * w50 - 64.f * w60) / 120.f;
+    tmp.s1 = (-w01 - 2.f * w11 - 4.f * w21 - 8.f * w31 - 16.f * w41 - 32.f * w51 - 64.f * w61) / 120.f;
+    tmp.s2 = (-w02 - 2.f * w12 - 4.f * w22 - 8.f * w32 - 16.f * w42 - 32.f * w52 - 64.f * w62) / 120.f;
+    tmp.s3 = (-w03 - 2.f * w13 - 4.f * w23 - 8.f * w33 - 16.f * w43 - 32.f * w53 - 64.f * w63) / 120.f;
+    tmp.s4 = (-w04 - 2.f * w14 - 4.f * w24 - 8.f * w34 - 16.f * w44 - 32.f * w54 - 64.f * w64) / 120.f;
+    tmp.s5 = (-w05 - 2.f * w15 - 4.f * w25 - 8.f * w35 - 16.f * w45 - 32.f * w55 - 64.f * w65) / 120.f;
+    tmp.s6 = (-w06 - 2.f * w16 - 4.f * w26 - 8.f * w36 - 16.f * w46 - 32.f * w56 - 64.f * w66) / 120.f;
+
+    OUTPUT_ROW_2x2_7x7(out4, tmp);
+
+    // Row 5
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out5 = 0.0f;
+
+    tmp.s0 = (w00 - 3.f * w10 + 9.f * w20 - 27.f * w30 + 81.f * w40 - 243.f * w50 + 729.f * w60) / 720.f;
+    tmp.s1 = (w01 - 3.f * w11 + 9.f * w21 - 27.f * w31 + 81.f * w41 - 243.f * w51 + 729.f * w61) / 720.f;
+    tmp.s2 = (w02 - 3.f * w12 + 9.f * w22 - 27.f * w32 + 81.f * w42 - 243.f * w52 + 729.f * w62) / 720.f;
+    tmp.s3 = (w03 - 3.f * w13 + 9.f * w23 - 27.f * w33 + 81.f * w43 - 243.f * w53 + 729.f * w63) / 720.f;
+    tmp.s4 = (w04 - 3.f * w14 + 9.f * w24 - 27.f * w34 + 81.f * w44 - 243.f * w54 + 729.f * w64) / 720.f;
+    tmp.s5 = (w05 - 3.f * w15 + 9.f * w25 - 27.f * w35 + 81.f * w45 - 243.f * w55 + 729.f * w65) / 720.f;
+    tmp.s6 = (w06 - 3.f * w16 + 9.f * w26 - 27.f * w36 + 81.f * w46 - 243.f * w56 + 729.f * w66) / 720.f;
+
+    OUTPUT_ROW_2x2_7x7(out5, tmp);
+
+    // Row 6
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out6 = 0.0f;
+
+    tmp.s0 = (w00 + 3.f * w10 + 9.f * w20 + 27.f * w30 + 81.f * w40 + 243.f * w50 + 729.f * w60) / 720.f;
+    tmp.s1 = (w01 + 3.f * w11 + 9.f * w21 + 27.f * w31 + 81.f * w41 + 243.f * w51 + 729.f * w61) / 720.f;
+    tmp.s2 = (w02 + 3.f * w12 + 9.f * w22 + 27.f * w32 + 81.f * w42 + 243.f * w52 + 729.f * w62) / 720.f;
+    tmp.s3 = (w03 + 3.f * w13 + 9.f * w23 + 27.f * w33 + 81.f * w43 + 243.f * w53 + 729.f * w63) / 720.f;
+    tmp.s4 = (w04 + 3.f * w14 + 9.f * w24 + 27.f * w34 + 81.f * w44 + 243.f * w54 + 729.f * w64) / 720.f;
+    tmp.s5 = (w05 + 3.f * w15 + 9.f * w25 + 27.f * w35 + 81.f * w45 + 243.f * w55 + 729.f * w65) / 720.f;
+    tmp.s6 = (w06 + 3.f * w16 + 9.f * w26 + 27.f * w36 + 81.f * w46 + 243.f * w56 + 729.f * w66) / 720.f;
+
+    OUTPUT_ROW_2x2_7x7(out6, tmp);
+
+    // Row 7
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out7 = 0.0f;
+
+    tmp.s0 = w60;
+    tmp.s1 = w61;
+    tmp.s2 = w62;
+    tmp.s3 = w63;
+    tmp.s4 = w64;
+    tmp.s5 = w65;
+    tmp.s6 = w66;
+
+    OUTPUT_ROW_2x2_7x7(out7, tmp);
+
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+
+    int x0 = get_global_id(2); // idx filter
+    int y0 = get_global_id(0); // idx channel
+
+    // Get output address
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(DATA_TYPE) + y0 * dst_stride_y;
+
+    // Store the values across the channels
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z) = out0.s4;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z) = out0.s5;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z) = out0.s6;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z) = out0.s7;
+
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out1.s0;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out1.s1;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out1.s2;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out1.s3;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out1.s4;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out1.s5;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out1.s6;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out1.s7;
+    *(__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z) = out2.s0;
+    *(__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z) = out2.s1;
+    *(__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z) = out2.s2;
+    *(__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z) = out2.s3;
+    *(__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z) = out2.s4;
+    *(__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z) = out2.s5;
+    *(__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z) = out2.s6;
+    *(__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z) = out2.s7;
+    *(__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z) = out3.s0;
+    *(__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z) = out3.s1;
+    *(__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z) = out3.s2;
+    *(__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z) = out3.s3;
+    *(__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z) = out3.s4;
+    *(__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z) = out3.s5;
+    *(__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z) = out3.s6;
+    *(__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z) = out3.s7;
+    *(__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z) = out4.s0;
+    *(__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z) = out4.s1;
+    *(__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z) = out4.s2;
+    *(__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z) = out4.s3;
+    *(__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z) = out4.s4;
+    *(__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z) = out4.s5;
+    *(__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z) = out4.s6;
+    *(__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z) = out4.s7;
+    *(__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z) = out5.s0;
+    *(__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z) = out5.s1;
+    *(__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z) = out5.s2;
+    *(__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z) = out5.s3;
+    *(__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z) = out5.s4;
+    *(__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z) = out5.s5;
+    *(__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z) = out5.s6;
+    *(__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z) = out5.s7;
+    *(__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z) = out6.s0;
+    *(__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z) = out6.s1;
+    *(__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z) = out6.s2;
+    *(__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z) = out6.s3;
+    *(__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z) = out6.s4;
+    *(__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z) = out6.s5;
+    *(__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z) = out6.s6;
+    *(__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z) = out6.s7;
+    *(__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z) = out7.s0;
+    *(__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z) = out7.s1;
+    *(__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z) = out7.s2;
+    *(__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z) = out7.s3;
+    *(__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z) = out7.s4;
+    *(__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z) = out7.s5;
+    *(__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z) = out7.s6;
+    *(__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z) = out7.s7;
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+}
 #endif // defined(SRC_DIM_Z)
 
 #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
@@ -1292,6 +1604,55 @@ __kernel void winograd_filter_transform_4x1_5x1_nhwc(
                                            dst_step_z,
                                            dst_offset_first_element_in_bytes);
 }
+
+/** This OpenCL kernel performs Winograd filter transform 7x1 when the data layout is NHWC 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
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
+ * @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_7x1_nhwc(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_2x2_7x7_nhwc(src_ptr,
+                                           src_stride_x,
+                                           src_step_x,
+                                           src_stride_y,
+                                           src_step_y,
+                                           src_stride_z,
+                                           src_step_z,
+                                           src_stride_w,
+                                           src_step_w,
+                                           src_offset_first_element_in_bytes,
+                                           dst_ptr,
+                                           dst_stride_x,
+                                           dst_step_x,
+                                           dst_stride_y,
+                                           dst_step_y,
+                                           dst_stride_z,
+                                           dst_step_z,
+                                           dst_offset_first_element_in_bytes);
+}
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
 #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
@@ -1539,4 +1900,53 @@ __kernel void winograd_filter_transform_1x4_1x5_nhwc(
                                            dst_step_z,
                                            dst_offset_first_element_in_bytes);
 }
+
+/** This OpenCL kernel performs Winograd filter transform 1x7 when the data layout is NHWC 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
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
+ * @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_1x7_nhwc(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    winograd_filter_transform_2x2_7x7_nhwc(src_ptr,
+                                           src_stride_x,
+                                           src_step_x,
+                                           src_stride_y,
+                                           src_step_y,
+                                           src_stride_z,
+                                           src_step_z,
+                                           src_stride_w,
+                                           src_step_w,
+                                           src_offset_first_element_in_bytes,
+                                           dst_ptr,
+                                           dst_stride_x,
+                                           dst_step_x,
+                                           dst_stride_y,
+                                           dst_step_y,
+                                           dst_stride_z,
+                                           dst_step_z,
+                                           dst_offset_first_element_in_bytes);
+}
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index b6923c15c9..f3e3220fd6 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -580,6 +580,102 @@ public:
     }
 };
 
+/** Data set containing small 1x7 tensor shapes. */
+class Small1x7Shapes final : public ShapeDataset
+{
+public:
+    Small1x7Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 1U, 7U, 7U, 4U },
+                     TensorShape{ 1U, 7U, 4U, 13U },
+                     TensorShape{ 1U, 7U, 9U, 2U },
+                     TensorShape{ 1U, 7U, 3U, 5U },
+    })
+    {
+    }
+};
+
+/** Data set containing large 1x7 tensor shapes. */
+class Large1x7Shapes final : public ShapeDataset
+{
+public:
+    Large1x7Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 1U, 7U, 32U, 64U },
+                     TensorShape{ 1U, 7U, 51U, 13U },
+                     TensorShape{ 1U, 7U, 53U, 47U },
+                     TensorShape{ 1U, 7U, 128U, 384U },
+    })
+    {
+    }
+};
+
+/** Data set containing small 7x7 tensor shapes. */
+class Small7x7Shapes final : public ShapeDataset
+{
+public:
+    Small7x7Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 7U, 7U, 4U },
+                     TensorShape{ 7U, 7U, 4U, 13U },
+                     TensorShape{ 7U, 7U, 9U, 2U },
+                     TensorShape{ 7U, 7U, 3U, 5U },
+    })
+    {
+    }
+};
+
+/** Data set containing large 7x7 tensor shapes. */
+class Large7x7Shapes final : public ShapeDataset
+{
+public:
+    Large7x7Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 7U, 32U, 64U },
+                     TensorShape{ 7U, 7U, 51U, 13U },
+                     TensorShape{ 7U, 7U, 53U, 47U },
+                     TensorShape{ 7U, 7U, 128U, 384U },
+    })
+    {
+    }
+};
+
+/** Data set containing small 7x1 tensor shapes. */
+class Small7x1Shapes final : public ShapeDataset
+{
+public:
+    Small7x1Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 1U, 7U, 4U },
+                     TensorShape{ 7U, 1U, 4U, 13U },
+                     TensorShape{ 7U, 1U, 9U, 2U },
+                     TensorShape{ 7U, 1U, 3U, 5U },
+    })
+    {
+    }
+};
+
+/** Data set containing large 7x1 tensor shapes. */
+class Large7x1Shapes final : public ShapeDataset
+{
+public:
+    Large7x1Shapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 1U, 32U, 64U },
+                     TensorShape{ 7U, 1U, 51U, 13U },
+                     TensorShape{ 7U, 1U, 53U, 47U },
+                     TensorShape{ 7U, 1U, 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 f933a287cc..1042dd7e08 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -118,7 +118,7 @@ const auto SmallWinogradFilterTransformDatasetNCHW =
            framework::dataset::concat(combine(datasets::Small5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
                                       combine(datasets::Small1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })))))));
 
-const auto SmallWinogradFilterTransformDatasetNHWC =
+const auto SmallWinogradFilterTransformDatasetNHWC_F16 =
            framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })),
            framework::dataset::concat(combine(datasets::Small3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
            framework::dataset::concat(combine(datasets::Small1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })),
@@ -126,6 +126,11 @@ const auto SmallWinogradFilterTransformDatasetNHWC =
            framework::dataset::concat(combine(datasets::Small5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
                                      (combine(datasets::Small1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) }))))))));
 
+const auto SmallWinogradFilterTransformDatasetNHWC_F32 =
+           framework::dataset::concat(SmallWinogradFilterTransformDatasetNHWC_F16,
+           framework::dataset::concat(combine(datasets::Small7x7Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U) })),
+           framework::dataset::concat(combine(datasets::Small7x1Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 1U) })),
+                                      combine(datasets::Small1x7Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 2U) })))));
 
 const auto LargeWinogradFilterTransformDatasetNCHW =
            framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
@@ -135,7 +140,7 @@ const auto LargeWinogradFilterTransformDatasetNCHW =
            framework::dataset::concat(combine(datasets::Large5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
                                       combine(datasets::Large1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })))))));
 
-const auto LargeWinogradFilterTransformDatasetNHWC =
+const auto LargeWinogradFilterTransformDatasetNHWC_F16 =
            framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })),
            framework::dataset::concat(combine(datasets::Large3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
            framework::dataset::concat(combine(datasets::Large1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })),
@@ -143,6 +148,12 @@ const auto LargeWinogradFilterTransformDatasetNHWC =
            framework::dataset::concat(combine(datasets::Large5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })),
                                       combine(datasets::Large1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })))))));
 
+const auto LargeWinogradFilterTransformDatasetNHWC_F32 =
+           framework::dataset::concat(LargeWinogradFilterTransformDatasetNHWC_F16,
+           framework::dataset::concat(combine(datasets::Large7x7Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U) })),
+           framework::dataset::concat(combine(datasets::Large7x1Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 1U) })),
+                                      combine(datasets::Large1x7Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 2U) })))));
+
 // Output transform
 const auto SmallWinogradOutputTransformDatasetNCHW = datasets::SmallWinogradOutputTransformDatasetNCHW();
 
@@ -364,7 +375,7 @@ TEST_SUITE_END() // NCHW
 TEST_SUITE(NHWC)
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixtureFP16, framework::DatasetMode::PRECOMMIT,
-                       combine(combine(SmallWinogradFilterTransformDatasetNHWC,
+                       combine(combine(SmallWinogradFilterTransformDatasetNHWC_F16,
                                        framework::dataset::make("DataLayout", { DataLayout::NHWC })),
                                        framework::dataset::make("DataType", { DataType::F16 })))
 {
@@ -373,7 +384,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixtureFP16, framework
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixtureFP16, framework::DatasetMode::NIGHTLY,
-                       combine(combine(LargeWinogradFilterTransformDatasetNHWC,
+                       combine(combine(LargeWinogradFilterTransformDatasetNHWC_F16,
                                        framework::dataset::make("DataLayout", { DataLayout::NHWC })),
                                        framework::dataset::make("DataType", { DataType::F16 })))
 {
@@ -383,7 +394,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixtureFP16, framework
 TEST_SUITE_END() // FP16
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixtureFP32, framework::DatasetMode::PRECOMMIT,
-                       combine(combine(SmallWinogradFilterTransformDatasetNHWC,
+                       combine(combine(SmallWinogradFilterTransformDatasetNHWC_F32,
                                        framework::dataset::make("DataLayout", { DataLayout::NHWC })),
                                        framework::dataset::make("DataType", { DataType::F32 })))
 {
@@ -392,7 +403,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixtureFP32, framework
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixtureFP32, framework::DatasetMode::NIGHTLY,
-                       combine(combine(LargeWinogradFilterTransformDatasetNHWC,
+                       combine(combine(LargeWinogradFilterTransformDatasetNHWC_F32,
                                        framework::dataset::make("DataLayout", { DataLayout::NHWC })),
                                        framework::dataset::make("DataType", { DataType::F32 })))
 {
diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index f09b2205d9..5525bc4535 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -193,6 +193,7 @@ void initialize_matrix_transform(SimpleTensor<T> &src, const Size2D &output_tile
         { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(5, 1), WinogradTransformType::FILTER), fmatrix4x4_5x5 },
         { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(7, 1), WinogradTransformType::FILTER), fmatrix2x1_7x7 },
         { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 7), WinogradTransformType::FILTER), fmatrix2x1_7x7 },
+        { WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7), WinogradTransformType::FILTER), fmatrix2x1_7x7 },
         { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 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 },