Enable premultiplication for depthwise convolution

with fp16 and quantized types

Resolves: COMPMID-6337
Change-Id: I81542e51c9c0329f202ac8452f173b138e51a0f6
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9883
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 15 insertions, 1 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp b/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
index ad4c821cfb..8a49c775d3 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
@@ -45,7 +45,9 @@ void do_premultiply_float_6(const float       *in_ptr,
         {
             const float *ip = ip2;
             float       *op = op2;
-            for(unsigned int c = 0; c < input_channels; c += BLOCK_SIZE)
+
+            unsigned int num_blocks = input_channels / BLOCK_SIZE;
+            for(unsigned int c = 0; c < num_blocks; c++)
             {
                 float vals[BLOCK_SIZE];
                 for(unsigned int v = 0; v < BLOCK_SIZE; v++)
@@ -63,6 +65,18 @@ void do_premultiply_float_6(const float       *in_ptr,
                     op += CHANNEL_MULTIPLIER;
                 }
             }
+
+            unsigned int rem = input_channels - num_blocks * BLOCK_SIZE;
+            for(unsigned int c = 0; c < rem; c++)
+            {
+                float val = ip[c];
+                for(unsigned int r = 0; r < CHANNEL_MULTIPLIER; r++)
+                {
+                    op[r] = val;
+                }
+                op += CHANNEL_MULTIPLIER;
+            }
+
             ip2 += ld_col;
             op2 += out_ld_col;
         }