COMPMID-452 CL Generic Depthwise Convolution implementation.
Change-Id: I115e48fe6ce5e281f3791aa5d80fdc754cdd2b5e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85082
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index cbcdbf2..9c2c3a5 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -24,6 +24,8 @@
#include "helpers.h"
+#if defined(CONV_STRIDE_X)
+
#if CONV_STRIDE_X == 1
#define convolution1x3 convolution1x3_stride_1
#elif CONV_STRIDE_X == 2
@@ -186,4 +188,134 @@
weights_values2.s0, weights_values2.s1, weights_values2.s2);
vstore2(pixels, 0, (__global float *)dst.ptr);
-}
\ No newline at end of file
+}
+
+#endif //defined(CONV_STRIDE_X)
+
+#if defined(SRC_WIDTH) && defined(DATA_TYPE)
+/** This kernel reshapes each of the tensor's low three dimensions to single rows.
+ *
+ * @note Datatype and source width should be given as a preprocessor argument using -DDATA_TYPE=type and -DSRC_WIDTH=width. e.g. -DSRC_WIDTH=128
+ *
+ * @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)
+ * @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 Y 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. 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] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void depthwise_weights_reshape(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(dst))
+{
+ Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+
+ __global DATA_TYPE *input_ptr = (__global DATA_TYPE *)src.ptr;
+ __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
+
+ for(int i = 0; i < SRC_WIDTH; ++i, ++input_ptr)
+ {
+ *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *input_ptr;
+ }
+}
+#endif //defined(SRC_WIDTH) && defined(DATA_TYPE)
+
+#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE)
+/** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The convolution information must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y, -DPAD_X, -DPAD_Y, -DKERNEL_WIDHT, -DKERNEL_HEIGHT, -DSRC_WIDTH, -DSRC_HEIGHT
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/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)
+ * @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] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_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 depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst))
+{
+ Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+ const int src_pixel_linear = get_global_id(1) * STRIDE_X;
+ const int full_length = SRC_WIDTH + 2 * PAD_X;
+ const int max_initial_x = STRIDE_X * (((full_length - KERNEL_WIDTH) / STRIDE_X) + 1);
+
+ const int src_x = -PAD_X + src_pixel_linear % max_initial_x;
+ const int src_y = -PAD_Y + src_pixel_linear / max_initial_x * STRIDE_Y;
+ const int src_z = get_global_id(2);
+
+ __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + src_z * src_stride_z;
+ __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr));
+
+ for(int y = src_y; y < src_y + KERNEL_HEIGHT; ++y)
+ {
+ for(int x = src_x; x < src_x + KERNEL_WIDTH; ++x, ++output_ptr)
+ {
+ if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
+ {
+ *output_ptr = 0;
+ }
+ else
+ {
+ *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
+ }
+ }
+ }
+}
+
+#endif //defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(DATA_TYPE)
+
+#if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
+
+/** This kernel performs a reshaping of the output of the depthwise generic convolution.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The convolution information must be passed at compile time using -DCONV_WIDTH, -DCONV_HEIGHT, e.g -DCONV_WIDTH=32, -DCONV_HEIGHT=42
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/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_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] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_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 depthwise_vector_to_tensor(
+ VECTOR_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ Vector src = CONVERT_TO_VECTOR_STRUCT(src);
+
+ const int patch_size = CONV_WIDTH * CONV_HEIGHT;
+ const int id0 = get_global_id(0);
+ const int z = id0 / patch_size;
+ const int index2D = id0 - z * patch_size;
+
+ __global uchar *out_ptr = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * dst_stride_x + index2D / CONV_WIDTH * dst_stride_y + z * dst_stride_z;
+ *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr);
+}
+
+#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)