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>

1 files changed, 168 insertions, 15 deletions

diff --git a/src/core/CL/cl_kernels/winograd_output_transform.cl b/src/core/CL/cl_kernels/winograd_output_transform.cl
index c63b206080..61f0f61db7 100644
--- a/src/core/CL/cl_kernels/winograd_output_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_output_transform.cl
@@ -351,9 +351,13 @@ __kernel void winograd_output_transform_4x4_3x3_nchw(
 #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 }
 
-/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 3x3 and the data layout is NHWC
+/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, 4x1 or 1x4, the filter size 3x3, 3x1 or 1x3 and the data layout is NHWC
  *
  * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
+ * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
+ * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
+ * @note If this kernel is used to perform Winograd output transform 3x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note If this kernel is used to perform Winograd output transform 1x3, -DWINOGRAD_OUTPUT_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)
@@ -381,12 +385,12 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
 #endif // defined(HAS_BIAS)
     int dst_size)
 {
-    // Each thread stores a 4x4 tile
+    // Each thread stores a 4x4/4x1 or 1x4 tile
     Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
 
     const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0);
 
-    // Load the values across the 36 channels to compose the 6x6 tile
+    // Load the values across the 36 channels to compose the 6x6 or 6x1 tile
     float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
     float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
     float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
@@ -394,6 +398,14 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
     float d04 = *((__global float *)(src_addr + 4 * src_stride_z));
     float d05 = *((__global float *)(src_addr + 5 * src_stride_z));
 
+#if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
+    // Compute out00, out01, out02 and out03
+    float out00 = d00 + d01 + d02 + d03 + d04;
+    float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04;
+    float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04;
+    float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05;
+#else  // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
+
     float d10 = *((__global float *)(src_addr + 6 * src_stride_z));
     float d11 = *((__global float *)(src_addr + 7 * src_stride_z));
     float d12 = *((__global float *)(src_addr + 8 * src_stride_z));
@@ -484,11 +496,12 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
     out31 += k1 - d12 + d22 - 8.0f * d32 + 8.0f * d42 - d52;
     out32 += 4.0f * k0 + d12 - d22 + 8.0f * d32 - 8.0f * d42 + d52;
     out33 += 4.0f * k1 - d12 + d15 + d22 - d25 - 8.0f * d32 + 8.0f * d35 + 8.0f * d42 - 8.0f * d45 - d52 + d55;
+#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
     int y_in  = get_global_id(1);
     int x_out = get_global_id(0);
-    int y_out = (y_in % NUM_TILES_X) * 4;
-    int z_out = (y_in / NUM_TILES_X) * 4;
+    int y_out = (y_in % NUM_TILES_X) * OUTPUT_TILE_W;
+    int z_out = (y_in / NUM_TILES_X) * OUTPUT_TILE_H;
 
 #if defined(HAS_BIAS)
     // Add bias
@@ -500,7 +513,7 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
     out01 += (float)b;
     out02 += (float)b;
     out03 += (float)b;
-
+#if !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) & !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     out10 += (float)b;
     out11 += (float)b;
     out12 += (float)b;
@@ -515,9 +528,29 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
     out31 += (float)b;
     out32 += (float)b;
     out33 += (float)b;
+#endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) & !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #endif // defined(HAS_BIAS)
 
+#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
+    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
+    offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
+
+    // Store the 1x4 output tile
+    *((__global float *)(dst_ptr + offset.s0)) = out00;
+    *((__global float *)(dst_ptr + offset.s1)) = out01;
+    *((__global float *)(dst_ptr + offset.s2)) = out02;
+    *((__global float *)(dst_ptr + offset.s3)) = out03;
+#elif defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
+    // Store the 4x1 output tile
+    int  offset = dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    int4 mult_y = min(dst_size - offset, 1);
+
+    *((__global float *)(dst_ptr + mult_y.s0 * 0 * dst_stride_y + offset)) = out00;
+    *((__global float *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset)) = out01;
+    *((__global float *)(dst_ptr + mult_y.s0 * 2 * dst_stride_y + offset)) = out02;
+    *((__global float *)(dst_ptr + mult_y.s0 * 3 * dst_stride_y + offset)) = out03;
+#else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
     // Get output address
     int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
     offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
@@ -540,7 +573,127 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
     *((__global float *)(dst_ptr + mult_y.s3 * 1 * dst_stride_y + offset.s3)) = out31;
     *((__global float *)(dst_ptr + mult_y.s3 * 2 * dst_stride_y + offset.s3)) = out32;
     *((__global float *)(dst_ptr + mult_y.s3 * 3 * dst_stride_y + offset.s3)) = out33;
+
+#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
+}
+
+#if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
+/** This OpenCL kernel performs Winograd output transform when the output tile is 4x1, the filter size 3x1 and the data layout is NHWC
+ *
+ * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
+ * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
+ * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
+ * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL 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)
+ * @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_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_output_transform_4x1_3x1_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(bias),
+#endif // defined(HAS_BIAS)
+    int dst_size)
+{
+    winograd_output_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_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,
+#if defined(HAS_BIAS)
+                                           bias_ptr,
+                                           bias_stride_x,
+                                           bias_step_x,
+                                           bias_offset_first_element_in_bytes,
+#endif // defined(HAS_BIAS)
+                                           dst_size);
 }
+#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
+
+#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
+/** This OpenCL kernel performs Winograd output transform when the output tile is 1x4, the filter size 1x3 and the data layout is NHWC
+ *
+ * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
+ * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
+ * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
+ * @note -DWINOGRAD_OUTPUT_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)
+ * @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_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_output_transform_1x4_1x3_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(bias),
+#endif // defined(HAS_BIAS)
+    int dst_size)
+{
+    winograd_output_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_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,
+#if defined(HAS_BIAS)
+                                           bias_ptr,
+                                           bias_stride_x,
+                                           bias_step_x,
+                                           bias_offset_first_element_in_bytes,
+#endif // defined(HAS_BIAS)
+                                           dst_size);
+}
+#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #define COMPUTE_TMP_COL(col, d0, d1, d2, d3, d4, d5, d6, d7, comm_fact)  \
     ({                                                                   \
@@ -646,14 +799,14 @@ __kernel void winograd_output_transform_4x4_5x5_nchw(
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    float d10 = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d11 = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d12 = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d13 = *((__global float *)(src_addr + 11 * src_stride_z));
-    float d14 = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d15 = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d16 = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d17 = *((__global float *)(src_addr + 15 * src_stride_z));
+    float d10                                                              = *((__global float *)(src_addr + 8 * src_stride_z));
+    float d11                                                              = *((__global float *)(src_addr + 9 * src_stride_z));
+    float d12                                                              = *((__global float *)(src_addr + 10 * src_stride_z));
+    float d13                                                              = *((__global float *)(src_addr + 11 * src_stride_z));
+    float d14                                                              = *((__global float *)(src_addr + 12 * src_stride_z));
+    float d15                                                              = *((__global float *)(src_addr + 13 * src_stride_z));
+    float d16                                                              = *((__global float *)(src_addr + 14 * src_stride_z));
+    float d17                                                              = *((__global float *)(src_addr + 15 * src_stride_z));
 
     float d20 = *((__global float *)(src_addr + 16 * src_stride_z));
     float d21 = *((__global float *)(src_addr + 17 * src_stride_z));
@@ -1290,4 +1443,4 @@ __kernel void winograd_output_transform_1x4_1x5_nchw(
                                           );
 }
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-#endif // defined(NUM_TILES_X) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H)
\ No newline at end of file
+#endif // defined(NUM_TILES_X) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H)