APPBROWSER-324 FP16 Weights optimization in DirectConvolution1x1
Change-Id: I94f09caa3d9852b7ecd1d3716ba7853de9dc1c43
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111636
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_convolution1x1.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs
index 071c185..1338299 100644
--- a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs
+++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs
@@ -160,6 +160,29 @@
uint z_index = gl_GlobalInvocationID.z;
TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w);
+#ifdef WEIGHTS_OPTIMIZATION
+ float w1, w2;
+ int nums = (int(weights_depth)) / 2;
+ for(int d = 0; d < nums; ++d)
+ {
+ vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter);
+
+ w1 = vec2_w.x;
+ vec4 r1[2] = CONVOLVE(src_iter, w1);
+ pixels[0] += r1[0];
+ pixels[1] += r1[1];
+
+ TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z);
+
+ w2 = vec2_w.y;
+ vec4 r2[2] = CONVOLVE(src_iter, w2);
+ pixels[0] += r2[0];
+ pixels[1] += r2[1];
+
+ TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z);
+ TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2));
+ }
+#else /* WEIGHTS_OPTIMIZATION */
float w;
for(int d = 0; d < int(weights_depth); ++d)
{
@@ -172,6 +195,7 @@
TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z);
TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z);
}
+#endif /* WEIGHTS_OPTIMIZATION */
#ifdef BIAS
vec2 vec2_b;