Merge pull request #3475 from wjc404/optimize-A53-dgemm

martin-frbg · web-flow · commit b0a590f4fe3a · 2021-12-12T19:09:08.000+01:00
optimize cgemm on ARM cortex A53 &amp; cortex A55
diff --git a/kernel/arm64/KERNEL.CORTEXA53 b/kernel/arm64/KERNEL.CORTEXA53
@@ -169,7 +169,7 @@ endif
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N)_cortexa53.c
 CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
 ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
 CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
diff --git a/kernel/arm64/KERNEL.CORTEXA55 b/kernel/arm64/KERNEL.CORTEXA55
@@ -169,7 +169,7 @@ endif
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N)_cortexa53.c
 CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
 ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
 CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
diff --git a/kernel/arm64/cgemm_kernel_8x4_cortexa53.c b/kernel/arm64/cgemm_kernel_8x4_cortexa53.c
