COMPMID-1337 Implementing Winograd Convolution Layer 1x3 and 3x1 kernels on OpenCL NHWC
Change-Id: Ia07e0dfcbcd07366c4bcb956e298369fb12a0369
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138759
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl
index 5f528d4..e53da9b 100644
--- a/src/core/CL/cl_kernels/winograd_filter_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl
@@ -285,9 +285,11 @@
#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !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
+/** This OpenCL kernel performs Winograd filter transform 3x3/3x1/1x3 when the data layout is NHWC and the output tile is 4x4/4x1/1x4
*
* @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 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note If this kernel is used to perform Winograd filter transform 1x3, -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)
@@ -317,32 +319,26 @@
const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w;
// Load the values from the input tensor
- float w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y));
- float w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y));
- float w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y));
- float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
- float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
- float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
- float w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
- float w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
- float w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
-
- // Transform the 3x3 tile in a 6x6 tile
- float out00, out01, out02, out03, out04, out05;
- float out10, out11, out12, out13, out14, out15;
- float out20, out21, out22, out23, out24, out25;
- float out30, out31, out32, out33, out34, out35;
- float out40, out41, out42, out43, out44, out45;
- float out50, out51, out52, out53, out54, out55;
-
- out00 = out01 = out02 = out03 = out04 = out05 = 0.f;
- out10 = out11 = out12 = out13 = out14 = out15 = 0.f;
- out20 = out21 = out22 = out23 = out24 = out25 = 0.f;
- out30 = out31 = out32 = out33 = out34 = out35 = 0.f;
- out40 = out41 = out42 = out43 = out44 = out45 = 0.f;
- out50 = out51 = out52 = out53 = out54 = out55 = 0.f;
+#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+ float w00 = *((__global float *)(src_addr + 0 * src_stride_z));
+ float w01 = *((__global float *)(src_addr + 1 * src_stride_z));
+ float w02 = *((__global float *)(src_addr + 2 * src_stride_z));
+#else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+ float w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y));
+ float w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y));
+ float w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y));
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+ float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
+ float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
+ float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
+ float w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
+ float w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
+ float w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
// Row 0
+ float out00, out01, out02, out03, out04, out05;
out00 = (w00) / 16.f;
out01 = (-w00 - w01 - w02) / 24.f;
out02 = (-w00 + w01 - w02) / 24.f;
@@ -350,7 +346,9 @@
out04 = (w00 - 2.f * w01 + 4.f * w02) / 96.f;
out05 = (w02) / 4.f;
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
// Row 1
+ float out10, out11, out12, out13, out14, out15;
out10 = (-w00 - w10 - w20) / 24.f;
out11 = (w00 + w10 + w20 + w01 + w11 + w21 + w02 + w12 + w22) / 36.f;
out12 = (w00 + w10 + w20 - w01 - w11 - w21 + w02 + w12 + w22) / 36.f;
@@ -359,6 +357,7 @@
out15 = (-w02 - w12 - w22) / 6.f;
// Row 2
+ float out20, out21, out22, out23, out24, out25;
out20 = (-w00 + w10 - w20) / 24.f;
out21 = (w00 - w10 + w20 + w01 - w11 + w21 + w02 - w12 + w22) / 36.f;
out22 = (w00 - w10 + w20 - w01 + w11 - w21 + w02 - w12 + w22) / 36.f;
@@ -367,6 +366,7 @@
out25 = (-w02 + w12 - w22) / 6.f;
// Row 3
+ float out30, out31, out32, out33, out34, out35;
out30 = (w00 + 2.f * w10 + 4.f * w20) / 96.f;
out31 = (-w00 - 2.f * w10 - 4.f * w20 - w01 - 2.f * w11 - 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f;
out32 = (-w00 - 2.f * w10 - 4.f * w20 + w01 + 2.f * w11 + 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f;
@@ -375,6 +375,7 @@
out35 = (w02 + 2.f * w12 + 4.f * w22) / 24.f;
// Row 4
+ float out40, out41, out42, out43, out44, out45;
out40 = (w00 - 2.f * w10 + 4.f * w20) / 96.f;
out41 = (-w00 + 2.f * w10 - 4.f * w20 - w01 + 2.f * w11 - 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f;
out42 = (-w00 + 2.f * w10 - 4.f * w20 + w01 - 2.f * w11 + 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f;
@@ -383,26 +384,31 @@
out45 = (w02 - 2.f * w12 + 4.f * w22) / 24.f;
// Row 5
+ float out50, out51, out52, out53, out54, out55;
out50 = (w20) / 4.f;
out51 = (-w20 - w21 - w22) / 6.f;
out52 = (-w20 + w21 - w22) / 6.f;
out53 = (w20 + 2.f * w21 + 4.f * w22) / 24.f;
out54 = (w20 - 2.f * w21 + 4.f * w22) / 24.f;
out55 = (w22);
+#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 * 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 values across the channels
- *(__global float *)(dst_addr + 0 * dst_stride_z) = out00;
- *(__global float *)(dst_addr + 1 * dst_stride_z) = out01;
- *(__global float *)(dst_addr + 2 * dst_stride_z) = out02;
- *(__global float *)(dst_addr + 3 * dst_stride_z) = out03;
- *(__global float *)(dst_addr + 4 * dst_stride_z) = out04;
- *(__global float *)(dst_addr + 5 * dst_stride_z) = out05;
+ // 36 channels for 3x3 kernels
+ // 6 channels for 3x1 or 1x3 kernels
+ *(__global float *)(dst_addr + 0 * dst_stride_z) = out00;
+ *(__global float *)(dst_addr + 1 * dst_stride_z) = out01;
+ *(__global float *)(dst_addr + 2 * dst_stride_z) = out02;
+ *(__global float *)(dst_addr + 3 * dst_stride_z) = out03;
+ *(__global float *)(dst_addr + 4 * dst_stride_z) = out04;
+ *(__global float *)(dst_addr + 5 * dst_stride_z) = out05;
+#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
*(__global float *)(dst_addr + 6 * dst_stride_z) = out10;
*(__global float *)(dst_addr + 7 * dst_stride_z) = out11;
*(__global float *)(dst_addr + 8 * dst_stride_z) = out12;
@@ -433,7 +439,108 @@
*(__global float *)(dst_addr + 33 * dst_stride_z) = out53;
*(__global float *)(dst_addr + 34 * dst_stride_z) = out54;
*(__global float *)(dst_addr + 35 * dst_stride_z) = out55;
+#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
}
+
+#if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1
+ *
+ * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
+ * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void winograd_filter_transform_4x1_3x1_nhwc(
+ TENSOR4D_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ winograd_filter_transform_4x4_3x3_nhwc(src_ptr,
+ src_stride_x,
+ src_step_x,
+ src_stride_y,
+ src_step_y,
+ src_stride_z,
+ src_step_z,
+ src_stride_w,
+ src_step_w,
+ src_offset_first_element_in_bytes,
+ dst_ptr,
+ dst_stride_x,
+ dst_step_x,
+ dst_stride_y,
+ dst_step_y,
+ dst_stride_z,
+ dst_step_z,
+ dst_offset_first_element_in_bytes);
+}
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
+
+#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
+/** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4
+ *
+ * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
+ * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void winograd_filter_transform_1x4_1x3_nhwc(
+ TENSOR4D_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ winograd_filter_transform_4x4_3x3_nhwc(src_ptr,
+ src_stride_x,
+ src_step_x,
+ src_stride_y,
+ src_step_y,
+ src_stride_z,
+ src_step_z,
+ src_stride_w,
+ src_step_w,
+ src_offset_first_element_in_bytes,
+ dst_ptr,
+ dst_stride_x,
+ dst_step_x,
+ dst_stride_y,
+ dst_step_y,
+ dst_stride_z,
+ dst_step_z,
+ dst_offset_first_element_in_bytes);
+}
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
/** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NCHW and the output tile is 4x4/4x1 or 1x4
*
* @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
@@ -1264,4 +1371,4 @@
dst_step_z,
dst_offset_first_element_in_bytes);
}
-#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
\ No newline at end of file
+#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)