IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[IE CLDNN] Improve performance of fc block fp16 implementation (#1993)

Browse Source 
Main purpose of this change is to fix weird behaviour of
fully_connected_gpu_fb_io_block_fp16 implementation where it shows
severe performance drop without bias.
Additionally assembly for case with bias is improved.
This commit is contained in:
Konrad Dobros
2020-08-31 08:52:47 +02:00
committed by GitHub
parent c7b3bd0195
commit 2b8249fc9f
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/fully_connected_gpu_fb_io_block_fp16.cl vendored
View File

@@ -214,8 +214,9 @@ KERNEL (fully_connected_gpu_xb_xb_block_fp16)(
CHUNK_TYPE acc[UNITS_PER_SG_READ] = {};
// Iterate over yxf linear plane (both filters/weights and input).
for (uint input_offset = input_base, filter_offset = filter_base; input_offset < input_byte_size; input_offset += input_sg_reads_distance)
{
uint input_offset = input_base;
uint filter_offset = filter_base;
do {
CHUNK_TYPE input_val = INPUT0_READ(input, input_offset + sg_elem_offset);
// Iterate over filters needed to process input read by sub-group.
@@ -233,7 +234,9 @@ KERNEL (fully_connected_gpu_xb_xb_block_fp16)(
acc[acc_pos] = AS_CHUNK(fma(AS_UNITS(input_val), SG_UNIT_SELECT(filter_val, acc_pos), AS_UNITS(acc[acc_pos])));
}
}
}
input_offset += input_sg_reads_distance;
} while (input_offset < input_byte_size);
// WRITE OUTPUT
// BATCH = 32x? (HF) / 16x? (F)