COMPMID-477 - Optimizing CLDirectConvolution 3x3 on OpenCL and added the auto configuration

Change-Id: I3c8384dcbc9d7786943134bb658dafb35356d90d
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/83253
Reviewed-by: Steven Niu <steven.niu@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 51 insertions, 140 deletions

diff --git a/src/core/CL/cl_kernels/direct_convolution3x3.cl b/src/core/CL/cl_kernels/direct_convolution3x3.cl
index b5524e1d4b..51886efe64 100644
--- a/src/core/CL/cl_kernels/direct_convolution3x3.cl
+++ b/src/core/CL/cl_kernels/direct_convolution3x3.cl
@@ -23,124 +23,48 @@
  */
 #include "helpers.h"
 
-#if STRIDE_X == 2
-#define CONVOLVE1x3(left_pixel_position, left_coeff, middle_coeff, right_coeff) convolution1x3_stride2(left_pixel_position, left_coeff, middle_coeff, right_coeff)
-#elif STRIDE_X == 1 /* STRIDE_X == 1 */
-#define CONVOLVE1x3(left_pixel_position, left_coeff, middle_coeff, right_coeff) convolution1x3_stride1(left_pixel_position, left_coeff, middle_coeff, right_coeff)
+#if STRIDE_X == 1
+#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr)
 #else /* STRIDE_X not equals 1 or 2 */
 #error "STRIDE_X larger than 2 is not supported"
 #endif /* STRIDE_X == 2 */
 
-/** Compute a 1D horizontal convolution of size 3 with stride as 1.
- *
- * @param[in] left_pixel   Pointer to the left pixel.
- * @param[in] left_coeff   Weight of the left pixel
- * @param[in] middle_coeff Weight of the middle pixel
- * @param[in] right_coeff  Weight of the right pixel
- *
- * @return a convoluted values.
- */
-inline VEC_DATA_TYPE(DATA_TYPE, 8) convolution1x3_stride1(__global const DATA_TYPE *left_pixel,
-                                                          const DATA_TYPE left_coeff,
-                                                          const DATA_TYPE middle_coeff,
-                                                          const DATA_TYPE right_coeff)
-{
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    temp = vload16(0, left_pixel);
-
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    left = temp.s01234567;
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    middle = temp.s12345678;
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    right = temp.s23456789;
-
-    return left * (VEC_DATA_TYPE(DATA_TYPE, 8))left_coeff + middle * (VEC_DATA_TYPE(DATA_TYPE, 8))middle_coeff + right * (VEC_DATA_TYPE(DATA_TYPE, 8))right_coeff;
-}
-
-/** Compute a 1D horizontal convolution of size 3 with stride as 2.
- *
- * @param[in] left_pixel   Pointer to the left pixel.
- * @param[in] left_coeff   Weight of the left pixel
- * @param[in] middle_coeff Weight of the middle pixel
- * @param[in] right_coeff  Weight of the right pixel
- *
- * @return a convoluted values.
- */
-inline VEC_DATA_TYPE(DATA_TYPE, 8) convolution1x3_stride2(__global const DATA_TYPE *left_pixel,
-                                                          const DATA_TYPE left_coeff,
-                                                          const DATA_TYPE middle_coeff,
-                                                          const DATA_TYPE right_coeff)
-{
-    const int stride_size = 2;
-
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    temp1 = vload16(0, left_pixel);
-
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    temp2 = vload16(0, left_pixel + 8);
-
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    left = (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0246, temp2.s0246);
-
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    middle = (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s1357, temp2.s1357);
-
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    right = (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s2468, temp2.s2468);
-
-    return left * (VEC_DATA_TYPE(DATA_TYPE, 8))left_coeff + middle * (VEC_DATA_TYPE(DATA_TYPE, 8))middle_coeff + right * (VEC_DATA_TYPE(DATA_TYPE, 8))right_coeff;
-}
-
-/** Apply a 3x3 2D convolution matrix on the input and return the result.
- *
- * Convolution matrix layout:
- *
- * [ mat0, mat1, mat2 ]\n
- * [ mat3, mat4, mat5 ]\n
- * [ mat6, mat7, mat8 ]\n
- *
- * @param[in] src  A pointer to source Image structure
- * @param[in] mat0 Coefficient from the convolution matrix
- * @param[in] mat1 Coefficient from the convolution matrix
- * @param[in] mat2 Coefficient from the convolution matrix
- * @param[in] mat3 Coefficient from the convolution matrix
- * @param[in] mat4 Coefficient from the convolution matrix
- * @param[in] mat5 Coefficient from the convolution matrix
- * @param[in] mat6 Coefficient from the convolution matrix
- * @param[in] mat0 Coefficient from the convolution matrix
- * @param[in] mat7 Coefficient from the convolution matrix
- * @param[in] mat8 Coefficient from the convolution matrix
- *
- * @return convoluted values.
- */
-inline VEC_DATA_TYPE(DATA_TYPE, 8) convolution3x3(
-    Image          *src,
-    const DATA_TYPE mat0, const DATA_TYPE mat1, const DATA_TYPE mat2,
-    const DATA_TYPE mat3, const DATA_TYPE mat4, const DATA_TYPE mat5,
-    const DATA_TYPE mat6, const DATA_TYPE mat7, const DATA_TYPE mat8)
-{
-    // Output pixels
-    VEC_DATA_TYPE(DATA_TYPE, 8)
-    pixels;
-
-    // Row 0
-    pixels = CONVOLVE1x3((__global DATA_TYPE *)offset(src, 0, 0), mat0, mat1, mat2);
-    // Row
-    pixels += CONVOLVE1x3((__global DATA_TYPE *)offset(src, 0, 1), mat3, mat4, mat5);
-    // Row 2
-    pixels += CONVOLVE1x3((__global DATA_TYPE *)offset(src, 0, 2), mat6, mat7, mat8);
-
-    return pixels;
-}
+#define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr)                                                               \
+    ({                                                                                                                          \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                             \
+        weights_values0 = vload4(0, weights_row_ptr);                                                                           \
+        VEC_DATA_TYPE(DATA_TYPE, 8)                                                                                             \
+        src0 = vload8(0, src_row_ptr);                                                                                          \
+        VEC_DATA_TYPE(DATA_TYPE, 2)                                                                                             \
+        src1 = vload2(0, src_row_ptr + 8);                                                                                      \
+        \
+        acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0;                                                          \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \
+    })
+
+#define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr)                                                            \
+    ({                                                                                                                       \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                          \
+        weights_values0 = vload4(0, weights_row_ptr);                                                                        \
+        VEC_DATA_TYPE(DATA_TYPE, 16)                                                                                         \
+        src0           = vload16(0, src_row_ptr);                                                                            \
+        DATA_TYPE src1 = *(src_row_ptr + 16);                                                                                \
+        \
+        acc += src0.even * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0;                                                  \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1;      \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \
+    })
 
 /** This kernel performs a direct convolution to convolve the low three dimensions.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The convolution stride x and stride y must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1, _DSTRIDE_Y=1
+ * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
  * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
  *
- * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: QS8/F16/F32
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16/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)
@@ -168,9 +92,9 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) convolution3x3(
  * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
  * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
- * @param[in]  weights_stride_w                      Stride of the weights tensor in W dimension
- * @param[in]  filter_depth                          The depth size of the filter
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
  */
+#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 __kernel void direct_convolution3x3(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
@@ -178,50 +102,37 @@ __kernel void direct_convolution3x3(
 #ifdef HAS_BIAS
     VECTOR_DECLARATION(biases),
 #endif /* defined(HAS_BIAS) */
-    unsigned int weights_stride_w,
-    unsigned int filter_depth)
+    unsigned int weights_stride_w)
 {
     Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
     Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
     Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
-#ifdef HAS_BIAS
-    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
-#endif /* defined(HAS_BIAS) */
-
     VEC_DATA_TYPE(DATA_TYPE, 8)
-    pixels = 0;
+    pixels0 = 0;
 
-    const uint z_index = get_global_id(2);
+    __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
+    __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
 
-    weights.ptr += z_index * weights_stride_w;
+    const int kernel_index = get_global_id(2);
+    weights_addr += kernel_index * weights_stride_w;
 
-    for(int d = 0; d < filter_depth; ++d)
+    for(int d = 0; d < WEIGHTS_DEPTH; ++d)
     {
-        VEC_DATA_TYPE(DATA_TYPE, 4)
-        weights_row1 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0));
-        VEC_DATA_TYPE(DATA_TYPE, 4)
-        weights_row2 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 1, 0));
-        VEC_DATA_TYPE(DATA_TYPE, 4)
-        weights_row3 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 2, 0));
-
-        pixels += convolution3x3(&src, weights_row1.s0,
-                                 weights_row1.s1,
-                                 weights_row1.s2,
-                                 weights_row2.s0,
-                                 weights_row2.s1,
-                                 weights_row2.s2,
-                                 weights_row3.s0,
-                                 weights_row3.s1,
-                                 weights_row3.s2);
+        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y));
+        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
+        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
 
-        src.ptr += src_stride_z;
-        weights.ptr += weights_stride_z;
+        src_addr += src_stride_z;
+        weights_addr += weights_stride_z;
     }
 
 #ifdef HAS_BIAS
-    pixels += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index)));
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+
+    pixels0 += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index)));
 #endif /* defined(HAS_BIAS) */
 
-    vstore8(pixels, 0, (__global DATA_TYPE *)dst.ptr);
+    vstore8(pixels0, 0, (__global DATA_TYPE *)dst.ptr);
 }
+#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
\ No newline at end of file