XNNPACK icon indicating copy to clipboard operation
XNNPACK copied to clipboard

enable AVX_VNNI_INT8 instruction for qd8-f32-qc8w-gemm/igemm

Open xujuntwt95329 opened this issue 5 months ago • 0 comments

$ qd8_f32_qc8w_gemm_bench.exe

qd8_f32_qc8w_gemm_minmax_ukernel_1x8c8:
..._1x8c8__avxvnni_prfm/mobilenet_v1/M:12544/N:32/K:27/real_time         142256 ns       143173 ns         4911 OPS=152.373G/s    <-- base
..._1x8c8__avxvnniint8_prfm/mobilenet_v1/M:12544/N:32/K:27/real_time     139960 ns       126325 ns         5566 OPS=154.873G/s    <-- opt

qd8_f32_qc8w_gemm_minmax_ukernel_5x8c8:
..._5x8c8__avxvnni_prfm/mobilenet_v1/M:12544/N:32/K:27/real_time         109927 ns       109497 ns         6136 OPS=197.186G/s    <-- base
..._5x8c8__avxvnniint8_prfm/mobilenet_v1/M:12544/N:32/K:27/real_time      99856 ns       103637 ns         7086 OPS=217.073G/s    <-- opt

xujuntwt95329 avatar Sep 04 '24 12:09 xujuntwt95329