APPBROWSER-290: DC5x5 optimization for FP16
Change-Id: I0833437d8353515a29aad69c15c99c7b0fd65156
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111631
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Joel Liang <joel.liang@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs
index f3b843d..a36bd43 100644
--- a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs
+++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs
@@ -26,6 +26,48 @@
#include "helpers.h"
+#ifdef DATA_TYPE_FP32
+
+precision highp float;
+
+/** This kernel performs a direct convolution to convolve the low three dimensions
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" 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 Z 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
+ * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
+ * @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 the 4th dimension
+ * @param[in] weights_depth The third dimensions of the weights tensors
+ */
+
layout(std140) uniform shader_params
{
TENSOR3D_PARAM_DECLARATION(src);
@@ -38,10 +80,6 @@
uint weights_depth;
};
-#ifdef DATA_TYPE_FP32
-
-precision highp float;
-
BUFFER_DECLARATION(src, 1, float, readonly);
BUFFER_DECLARATION(dst, 2, float, writeonly);
BUFFER_DECLARATION(weights, 3, float, readonly);
@@ -145,6 +183,58 @@
precision mediump float;
+#if defined(PROCESS_4X_1Y_1Z)
+
+/** This kernel performs a direct convolution to convolve the low three dimensions
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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_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 Z 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
+ * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
+ * @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 the 4th dimension
+ * @param[in] weights_depth The third dimensions of the weights tensors
+ */
+
+layout(std140) uniform shader_params
+{
+ TENSOR3D_PARAM_DECLARATION(src);
+ TENSOR3D_PARAM_DECLARATION(dst);
+ TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+ VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+ uint weights_stride_w;
+ uint weights_depth;
+};
+
BUFFER_DECLARATION(src, 1, uvec2, readonly);
BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
BUFFER_DECLARATION(weights, 3, uint, readonly);
@@ -203,7 +293,6 @@
return ret;
}
-// output 4 element per thread
vec4 convolve1x5_stride1(vec4 tmp[2], vec2 w[3])
{
vec4 src0 = tmp[0];
@@ -308,6 +397,635 @@
GC_STORE1_3D_OFFSET(packed_d, dst, 0, 0, 0);
}
+#elif defined(PROCESS_4X_3Y_1Z)
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 3 elements @ Y at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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_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 Z 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
+ * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
+ * @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 the 4th dimension
+ * @param[in] weights_depth The third dimensions of the weights tensors
+ */
+
+layout(std140) uniform shader_params
+{
+ TENSOR3D_PARAM_DECLARATION(src);
+ TENSOR3D_PARAM_DECLARATION(dst);
+ TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+ VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+ uint weights_stride_w;
+ uint weights_depth;
+};
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(bias, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 1
+#define LOAD_SRC(src, row) load_src_stride1(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride1(src, weight)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define LOAD_SRC(src, row) load_src_stride2(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride2(src, weight)
+#else /* STRDIDE_X == 1 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 1 */
+
+vec4[2] load_src_stride1(Image src, int row)
+{
+ uvec2 packed[2];
+ vec4 ret[2];
+
+ GC_LOAD2_2D_OFFSET(packed, src, 0, row);
+
+ ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+ ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+
+ return ret;
+}
+
+vec4[3] load_src_stride2(Image src, int row)
+{
+ uvec2 packed[3];
+ vec4 ret[3];
+
+ GC_LOAD3_2D_OFFSET(packed, src, 0, row);
+
+ ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+ ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+ ret[2] = vec4(unpackHalf2x16(packed[2].x), unpackHalf2x16(packed[2].y));
+
+ return ret;
+}
+
+vec2[3] load_weight(Tensor3D weights, int row)
+{
+ uvec3 packed_w;
+ vec2 ret[3];
+
+ GC_LOAD3_3D_OFFSET(packed_w, weights, 0, row, 0);
+
+ ret[0] = vec2(unpackHalf2x16(packed_w[0]));
+ ret[1] = vec2(unpackHalf2x16(packed_w[1]));
+ ret[2] = vec2(unpackHalf2x16(packed_w[2]));
+
+ return ret;
+}
+
+vec4 convolve1x5_stride1(vec4 tmp[2], vec2 w[3])
+{
+ vec4 src0 = tmp[0];
+ vec4 src1 = vec4(tmp[0].yzw, tmp[1].x);
+ vec4 src2 = vec4(tmp[0].zw, tmp[1].xy);
+ vec4 src3 = vec4(tmp[0].w, tmp[1].xyz);
+ vec4 src4 = tmp[1];
+ vec4 ret = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ return ret;
+}
+
+vec4 convolve1x5_stride2(vec4 tmp[3], vec2 w[3])
+{
+ vec4 src0 = vec4(tmp[0].xz, tmp[1].xz);
+ vec4 src1 = vec4(tmp[0].yw, tmp[1].yw);
+ vec4 src2 = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+ vec4 src3 = vec4(tmp[0].w, tmp[1].yw, tmp[2].y);
+ vec4 src4 = vec4(tmp[1].x, tmp[1].z, tmp[2].xz);
+ vec4 ret = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ return ret;
+}
+
+void main()
+{
+ Image src = GC_CONVERT_TO_IMAGE_STRUCT(src);
+ Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+ Tensor3D dst = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+ Vector bias = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
+#endif /* BIAS */
+
+ vec4 res[3];
+ vec2 w[5][3];
+ vec4 s[STRIDE_X + 1];
+ uvec2 packed_d;
+ uint z_index = gl_GlobalInvocationID.z;
+ int i;
+
+ for(i = 0; i < 3; i++)
+ {
+ res[i] = vec4(0);
+ }
+
+ weights.current_offset += z_index * weights_stride_w;
+
+ for(int d = 0; d < int(weights_depth); ++d)
+ {
+ // load weights once
+ for(int row = 0; row < 5; row++)
+ {
+ w[row] = load_weight(weights, row);
+ }
+
+ // 1st line
+ s = LOAD_SRC(src, 0);
+ res[0] += CONVOLVE1x5(s, w[0]);
+
+ // 2nd line
+ s = LOAD_SRC(src, 1);
+ res[0] += CONVOLVE1x5(s, w[1]);
+ res[1] += CONVOLVE1x5(s, w[0]);
+
+ // 3rd line
+ s = LOAD_SRC(src, 2);
+ res[0] += CONVOLVE1x5(s, w[2]);
+ res[1] += CONVOLVE1x5(s, w[1]);
+ res[2] += CONVOLVE1x5(s, w[0]);
+
+ // 4th line
+ s = LOAD_SRC(src, 3);
+ res[0] += CONVOLVE1x5(s, w[3]);
+ res[1] += CONVOLVE1x5(s, w[2]);
+ res[2] += CONVOLVE1x5(s, w[1]);
+
+ // 5th line
+ s = LOAD_SRC(src, 4);
+ res[0] += CONVOLVE1x5(s, w[4]);
+ res[1] += CONVOLVE1x5(s, w[3]);
+ res[2] += CONVOLVE1x5(s, w[2]);
+
+ // 6th line
+ s = LOAD_SRC(src, 5);
+ res[1] += CONVOLVE1x5(s, w[4]);
+ res[2] += CONVOLVE1x5(s, w[3]);
+
+ // 7th line
+ s = LOAD_SRC(src, 6);
+ res[2] += CONVOLVE1x5(s, w[4]);
+
+ src.current_offset += src_stride_z;
+ weights.current_offset += weights_stride_z;
+ }
+
+#ifdef BIAS
+ uint packed_b;
+ float b;
+
+ GC_LOAD1_1D_OFFSET(packed_b, bias, z_index);
+ b = (z_index % uint(2) == uint(0)) ? unpackHalf2x16(packed_b).x : unpackHalf2x16(packed_b).y;
+ for(i = 0; i < 3; i++)
+ {
+ res[i] += vec4(b);
+ }
+#endif /* BIAS */
+
+ for(i = 0; i < 3; i++)
+ {
+ packed_d = uvec2(packHalf2x16(res[i].xy), packHalf2x16(res[i].zw));
+ GC_STORE1_3D_OFFSET(packed_d, dst, 0, i, 0);
+ }
+}
+
+#elif defined(PROCESS_4X_3Y_2Z)
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 3 elements @ Y and 2 elements @ Z at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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_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 Z 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
+ * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
+ * @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 the 4th dimension
+ * @param[in] weights_depth The third dimensions of the weights tensors
+ */
+
+layout(std140) uniform shader_params
+{
+ TENSOR3D_PARAM_DECLARATION(src);
+ TENSOR3D_PARAM_DECLARATION(dst);
+ TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+ VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+ uint weights_stride_w;
+ uint weights_depth;
+};
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(bias, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 1
+#define LOAD_SRC(src, row) load_src_stride1(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride1(src, weight)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define LOAD_SRC(src, row) load_src_stride2(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride2(src, weight)
+#else /* STRDIDE_X == 1 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 1 */
+
+vec4[2] load_src_stride1(Image src, int row)
+{
+ uvec2 packed[2];
+ vec4 ret[2];
+
+ GC_LOAD2_2D_OFFSET(packed, src, 0, row);
+
+ ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+ ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+
+ return ret;
+}
+
+vec4[3] load_src_stride2(Image src, int row)
+{
+ uvec2 packed[3];
+ vec4 ret[3];
+
+ GC_LOAD3_2D_OFFSET(packed, src, 0, row);
+
+ ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+ ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+ ret[2] = vec4(unpackHalf2x16(packed[2].x), unpackHalf2x16(packed[2].y));
+
+ return ret;
+}
+
+vec2[3] load_weight(Tensor3D weights, int row)
+{
+ uvec3 packed_w;
+ vec2 ret[3];
+
+ GC_LOAD3_3D_OFFSET(packed_w, weights, 0, row, 0);
+
+ ret[0] = vec2(unpackHalf2x16(packed_w[0]));
+ ret[1] = vec2(unpackHalf2x16(packed_w[1]));
+ ret[2] = vec2(unpackHalf2x16(packed_w[2]));
+
+ return ret;
+}
+
+vec4 convolve1x5_stride1(vec4 tmp[2], vec2 w[3])
+{
+ vec4 src0 = tmp[0];
+ vec4 src1 = vec4(tmp[0].yzw, tmp[1].x);
+ vec4 src2 = vec4(tmp[0].zw, tmp[1].xy);
+ vec4 src3 = vec4(tmp[0].w, tmp[1].xyz);
+ vec4 src4 = tmp[1];
+ vec4 ret = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ return ret;
+}
+
+vec4 convolve1x5_stride2(vec4 tmp[3], vec2 w[3])
+{
+ vec4 src0 = vec4(tmp[0].xz, tmp[1].xz);
+ vec4 src1 = vec4(tmp[0].yw, tmp[1].yw);
+ vec4 src2 = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+ vec4 src3 = vec4(tmp[0].w, tmp[1].yw, tmp[2].y);
+ vec4 src4 = vec4(tmp[1].x, tmp[1].z, tmp[2].xz);
+ vec4 ret = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ return ret;
+}
+
+void main()
+{
+ Image src = GC_CONVERT_TO_IMAGE_STRUCT(src);
+ Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+ Tensor3D dst = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+ Vector bias = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
+#endif /* BIAS */
+
+ vec4 res[3];
+ vec2 w[5][3];
+ vec4 s[STRIDE_X + 1];
+ uvec2 packed_d;
+ uint z_index = (gl_GlobalInvocationID.z);
+ uint s_offset = src.current_offset;
+ int i, z;
+
+ weights.current_offset += z_index * weights_stride_w;
+
+ for(z = 0; z < 2; z++)
+ {
+ z_index += uint(z);
+ src.current_offset = s_offset;
+
+ for(i = 0; i < 3; i++)
+ {
+ res[i] = vec4(0);
+ }
+
+ for(int d = 0; d < int(weights_depth); ++d)
+ {
+ // load weights once
+ for(int row = 0; row < 5; row++)
+ {
+ w[row] = load_weight(weights, row);
+ }
+
+ // 1st line
+ s = LOAD_SRC(src, 0);
+ res[0] += CONVOLVE1x5(s, w[0]);
+
+ // 2nd line
+ s = LOAD_SRC(src, 1);
+ res[0] += CONVOLVE1x5(s, w[1]);
+ res[1] += CONVOLVE1x5(s, w[0]);
+
+ // 3rd line
+ s = LOAD_SRC(src, 2);
+ res[0] += CONVOLVE1x5(s, w[2]);
+ res[1] += CONVOLVE1x5(s, w[1]);
+ res[2] += CONVOLVE1x5(s, w[0]);
+
+ // 4th line
+ s = LOAD_SRC(src, 3);
+ res[0] += CONVOLVE1x5(s, w[3]);
+ res[1] += CONVOLVE1x5(s, w[2]);
+ res[2] += CONVOLVE1x5(s, w[1]);
+
+ // 5th line
+ s = LOAD_SRC(src, 4);
+ res[0] += CONVOLVE1x5(s, w[4]);
+ res[1] += CONVOLVE1x5(s, w[3]);
+ res[2] += CONVOLVE1x5(s, w[2]);
+
+ // 6th line
+ s = LOAD_SRC(src, 5);
+ res[1] += CONVOLVE1x5(s, w[4]);
+ res[2] += CONVOLVE1x5(s, w[3]);
+
+ // 7th line
+ s = LOAD_SRC(src, 6);
+ res[2] += CONVOLVE1x5(s, w[4]);
+
+ src.current_offset += src_stride_z;
+ weights.current_offset += weights_stride_z;
+ }
+
+#ifdef BIAS
+ uint packed_b;
+ float b;
+
+ GC_LOAD1_1D_OFFSET(packed_b, bias, z_index);
+ b = (z_index % uint(2) == uint(0)) ? unpackHalf2x16(packed_b).x : unpackHalf2x16(packed_b).y;
+ for(i = 0; i < 3; i++)
+ {
+ res[i] += vec4(b);
+ }
+#endif /* BIAS */
+
+ for(i = 0; i < 3; i++)
+ {
+ packed_d = uvec2(packHalf2x16(res[i].xy), packHalf2x16(res[i].zw));
+ GC_STORE1_3D_OFFSET(packed_d, dst, 0, i, 0);
+ }
+
+ dst.current_offset += dst_stride_z;
+ }
+}
+
+#elif defined(PROCESS_8X_1Y_1Z)
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 8 elements @ X at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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_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 Z 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
+ * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
+ * @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 the 4th dimension
+ * @param[in] weights_depth The third dimensions of the weights tensors
+ */
+
+layout(std140) uniform shader_params
+{
+ TENSOR3D_PARAM_DECLARATION(src);
+ TENSOR3D_PARAM_DECLARATION(dst);
+ TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+ VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+ uint weights_stride_w;
+ uint weights_depth;
+};
+
+BUFFER_DECLARATION(src, 1, uvec4, readonly);
+BUFFER_DECLARATION(dst, 2, uvec4, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(bias, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 1
+#define LOAD_SRC(src, row) load_src_stride1(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride1(src, weight)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#error stride == 2 for PROCESS_8X_1Y not implemented
+#else /* STRDIDE_X == 1 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 1 */
+
+vec4[3] load_src_stride1(Image src, int row)
+{
+ uvec4 packed[2];
+ vec4 ret[3];
+
+ GC_LOAD2_2D_OFFSET(packed, src, 0, row);
+
+ ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+ ret[1] = vec4(unpackHalf2x16(packed[0].z), unpackHalf2x16(packed[0].w));
+ ret[2] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+
+ return ret;
+}
+
+vec2[3] load_weight(Tensor3D weights, int row)
+{
+ uvec3 packed_w;
+ vec2 ret[3];
+
+ GC_LOAD3_3D_OFFSET(packed_w, weights, 0, row, 0);
+
+ ret[0] = vec2(unpackHalf2x16(packed_w[0]));
+ ret[1] = vec2(unpackHalf2x16(packed_w[1]));
+ ret[2] = vec2(unpackHalf2x16(packed_w[2]));
+
+ return ret;
+}
+
+vec4[2] convolve1x5_stride1(vec4 tmp[3], vec2 w[3])
+{
+ vec4 src0 = tmp[0];
+ vec4 src1 = vec4(tmp[0].yzw, tmp[1].x);
+ vec4 src2 = vec4(tmp[0].zw, tmp[1].xy);
+ vec4 src3 = vec4(tmp[0].w, tmp[1].xyz);
+ vec4 src4 = tmp[1];
+ vec4 ret[2];
+
+ ret[0] = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ src0 = tmp[1];
+ src1 = vec4(tmp[1].yzw, tmp[2].x);
+ src2 = vec4(tmp[1].zw, tmp[2].xy);
+ src3 = vec4(tmp[1].w, tmp[2].xyz);
+ src4 = tmp[2];
+ ret[1] = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+ return ret;
+}
+
+void main()
+{
+ Image src = GC_CONVERT_TO_IMAGE_STRUCT(src);
+ Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+ Tensor3D dst = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+ Vector bias = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
+#endif /* BIAS */
+
+ vec4 res[2];
+ vec2 w[3];
+ vec4 s[STRIDE_X + 2];
+ uvec4 packed_d;
+ uint z_index = gl_GlobalInvocationID.z;
+
+ res[0] = vec4(0);
+ res[1] = vec4(0);
+ weights.current_offset += z_index * weights_stride_w;
+
+ for(int d = 0; d < int(weights_depth); ++d)
+ {
+ for(int row = 0; row < 5; row++)
+ {
+ w = load_weight(weights, row);
+ s = LOAD_SRC(src, row);
+ res[0] += CONVOLVE1x5(s, w)[0];
+ res[1] += CONVOLVE1x5(s, w)[1];
+ }
+
+ src.current_offset += src_stride_z;
+ weights.current_offset += weights_stride_z;
+ }
+
+#ifdef BIAS
+ uint packed_b;
+ float b;
+
+ GC_LOAD1_1D_OFFSET(packed_b, bias, z_index);
+ b = (z_index % uint(2) == uint(0)) ? unpackHalf2x16(packed_b).x : unpackHalf2x16(packed_b).y;
+ res[0] += vec4(b);
+ res[1] += vec4(b);
+#endif /* BIAS */
+
+ packed_d.xy = uvec2(packHalf2x16(res[0].xy), packHalf2x16(res[0].zw));
+ packed_d.zw = uvec2(packHalf2x16(res[1].xy), packHalf2x16(res[1].zw));
+ GC_STORE1_3D_OFFSET(packed_d, dst, 0, 0, 0);
+}
+
+#else /* defined(PROCESS_4X_1Y_1Z) */
+
+#endif /* defined(PROCESS_4X_1Y_1Z) */
+
#else /* DATA_TYPE_FP16 */
#error Data type not supported
#endif /* DATA_TYPE_FP16 */