Skip to content

Commit

Permalink
Enable prefetch for FP32 1x8 GEMM Cortex A53 microkernel
Browse files Browse the repository at this point in the history 
PiperOrigin-RevId: 518076824
  • Loading branch information
@fbarchard @xnnpack-bot
fbarchard authored and xnnpack-bot committed Mar 20, 2023
1 parent 92cebdc commit 14d5fcc
Showing 1 changed file with 6 additions and 6 deletions.
12 changes: 6 additions & 6 deletions src/init.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -924,8 +924,8 @@ static void init(void) {
case cpuinfo_uarch_kryo:
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_4x8__asm_aarch32_neon_cortex_a7);
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_4x8__asm_aarch32_neon_cortex_a7);
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch32_neon_cortex_a53);
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_1x8__asm_aarch32_neon_cortex_a53);
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch32_neon_prfm_cortex_a53);
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_1x8__asm_aarch32_neon_prfm_cortex_a53);
xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params;
xnn_params.f32.gemm.mr = 4;
xnn_params.f32.gemm.nr = 8;
Expand Down Expand Up @@ -2711,10 +2711,10 @@ static void init(void) {
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_4x8__asm_aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_4x8__asm_aarch64_neonfma_prfm_cortex_a53;
#endif
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_cortex_a53;
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_1x8__asm_aarch64_neonfma_prfm_cortex_a53;
#if XNN_ENABLE_JIT
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a53;
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.generator.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_igemm_code_generator_fn) xnn_generate_f32_igemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53;
#if XNN_ENABLE_GEMM_M_SPECIALIZATION
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53;
Expand All @@ -2726,10 +2726,10 @@ static void init(void) {
} else if (mr == 4 && nr == 8 && log2_sr == 0) {
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_4x8__asm_aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_4x8__asm_aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_cortex_a53;
xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_fn) xnn_f32_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_fn) xnn_f32_igemm_minmax_ukernel_1x8__asm_aarch64_neonfma_prfm_cortex_a53;
#if XNN_ENABLE_JIT
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a53;
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.generator.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_jit_igemm_code_generator_fn) xnn_generate_f32_igemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.generator.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_jit_gemm_code_generator_fn) xnn_generate_f32_gemm_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53;
xnn_params.f32.gemm.generator.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_jit_igemm_code_generator_fn) xnn_generate_f32_igemm_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53;
Expand Down

0 comments on commit 14d5fcc

Please sign in to comment.