QU8 4x8 NEON MLA Lane microkernel AArch32 assembly language
- Template generated from QS8 source.
- Prefetch and non-prefetch version.
- RNDNU quantization.
PiperOrigin-RevId: 422755432
diff --git a/bench/qu8-gemm.cc b/bench/qu8-gemm.cc
index ba25e15..941eddc 100644
--- a/bench/qu8-gemm.cc
+++ b/bench/qu8-gemm.cc
@@ -308,6 +308,22 @@
#endif // BENCHMARK_RUY
+#if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
+ static void qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64(benchmark::State& state, const char* net) {
+ GEMMBenchmark(state,
+ xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64,
+ xnn_init_qu8_conv_minmax_rndnu_neon_params,
+ 4, 8, 1, 1, benchmark::utils::CheckNEON);
+ }
+ static void qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64(benchmark::State& state, const char* net) {
+ GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64,
+ xnn_init_qu8_conv_minmax_rndnu_neon_params,
+ 4, 8, 1, 1, benchmark::utils::CheckNEON);
+ }
+ BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_ld64)
+ BENCHMARK_GEMM(qu8_gemm_4x8__aarch32_neon_mlal_lane_prfm_ld64)
+#endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
+
#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
static void qu8_gemm_4x16c4__aarch64_neondot_cortex_a55(benchmark::State& state, const char* net) {
GEMMBenchmark(state,