commit 666635c68ebbb182d1db4a85f33ed5325d472a65
author zhenglin <zhenglin.li@arm.com> Mon Dec 04 14:38:09 2017 +0800
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:41:58 2018 +0000
tree bfd45697f2a2615681adc5c3510948f8350b4710
parent 83f826537a195378fb3cc5d266fa06faf301a894

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>

src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs

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;

src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp

diff --git a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
index 5c7320a..962f044 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
@@ -191,6 +191,10 @@
             case DataType::F16:
                 num_elems_read_per_iteration_x    = 8;
                 num_elems_written_per_iteration_x = 8;
+                if(weights->info()->dimension(2) % 2 == 0)
+                {
+                    options.emplace("#define WEIGHTS_OPTIMIZATION");
+                }
                 break;
 
             case DataType::F32:
@@ -255,7 +259,10 @@
     switch(weights->info()->data_type())
     {
         case DataType::F16:
-            weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size + 1, kernel_size);
+            if((weights->info()->dimension(2) % 2 != 0) || (kernel_size != 1))
+            {
+                weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size + 1, kernel_size);
+            }
             if(_bias != nullptr)
             {
                 bias_access = AccessWindowStatic(_bias->info(), 0, 0, _bias->info()->dimension(0) + 1, 1);