Merge pull request #5454 from ChipKerchner/addRISCV64_BFLOAT16

Add support for BF16 sbgemm on RISCV.

Author: martin-frbg

Makefile.riscv64

@@ -6,23 +6,20 @@ ifeq ($(CORE), x280)
 CCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_zvl512b -mabi=lp64d 
 FCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d -static
 endif
-ifeq ($(CORE), RISCV64_ZVL256B)
+RISCV64_OPT  = rv64imafdcv
 ifeq ($(BUILD_HFLOAT16), 1)
-CCOMMON_OPT += -march=rv64imafdcv_zvl256b_zvfh_zfh -mabi=lp64d
-FCOMMON_OPT += -march=rv64imafdcv_zvfh_zfh -mabi=lp64d
-else
-CCOMMON_OPT += -march=rv64imafdcv_zvl256b -mabi=lp64d
-FCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d
+RISCV64_OPT := $(RISCV64_OPT)_zvfh_zfh
 endif
+ifeq ($(BUILD_BFLOAT16), 1)
+RISCV64_OPT := $(RISCV64_OPT)_zfbfmin_zvfbfmin_zvfbfwma
 endif
-ifeq ($(CORE), RISCV64_ZVL128B)
-ifeq ($(BUILD_HFLOAT16), 1)
-CCOMMON_OPT += -march=rv64imafdcv_zvfh_zfh -mabi=lp64d
-FCOMMON_OPT += -march=rv64imafdcv_zvfh_zfh -mabi=lp64d
-else
-CCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d
-FCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d
+ifeq ($(CORE), RISCV64_ZVL256B)
+CCOMMON_OPT += -march=$(RISCV64_OPT)_zvl256b -mabi=lp64d
+FCOMMON_OPT += -march=$(RISCV64_OPT) -mabi=lp64d
 endif
+ifeq ($(CORE), RISCV64_ZVL128B)
+CCOMMON_OPT += -march=$(RISCV64_OPT) -mabi=lp64d
+FCOMMON_OPT += -march=$(RISCV64_OPT) -mabi=lp64d
 endif
 ifeq ($(CORE), RISCV64_GENERIC)
 CCOMMON_OPT += -march=rv64imafdc -mabi=lp64d

kernel/riscv64/KERNEL.RISCV64_ZVL128B

@@ -256,7 +256,7 @@ ifndef ZGEMM_BETA
 ZGEMM_BETA = zgemm_beta_rvv.c
 endif
 
-ifeq ($(BUILD_BFLOAT16), 1)
+ifeq ($(BUILD_HFLOAT16), 1)
 SHGEMMKERNEL    =  shgemm_kernel_$(SHGEMM_UNROLL_M)x$(SHGEMM_UNROLL_N)_zvl128b.c
 SHGEMMONCOPY    =  ../generic/gemm_ncopy_$(SHGEMM_UNROLL_N).c
 SHGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SHGEMM_UNROLL_N).c
@@ -267,5 +267,16 @@ SHGEMM_BETA =  gemm_beta_rvv.c
 endif
 endif
 
+ifeq ($(BUILD_BFLOAT16), 1)
+SBGEMMKERNEL    =  sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_zvl128b.c
+SBGEMMONCOPY    =  ../generic/gemm_ncopy_$(SBGEMM_UNROLL_N).c
+SBGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SBGEMM_UNROLL_N).c
+SBGEMMONCOPYOBJ =  sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SBGEMMOTCOPYOBJ =  sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ifndef SBGEMM_BETA
+SBGEMM_BETA =  gemm_beta_rvv.c
+endif
+endif
+
 DOMATCOPY_CT = omatcopy_ct_rvv.c
 SOMATCOPY_CT = omatcopy_ct_rvv.c

kernel/riscv64/KERNEL.RISCV64_ZVL256B

@@ -223,7 +223,7 @@ DOMATCOPY_CT = omatcopy_ct_rvv.c
 SOMATCOPY_CT = omatcopy_ct_rvv.c
 
 
-ifeq ($(BUILD_BFLOAT16), 1)
+ifeq ($(BUILD_HFLOAT16), 1)
 SHGEMMKERNEL    =  shgemm_kernel_$(SHGEMM_UNROLL_M)x$(SHGEMM_UNROLL_N)_zvl256b.c
 ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
 SHGEMMINCOPY    =  ../generic/gemm_ncopy_$(SHGEMM_UNROLL_M).c
@@ -240,5 +240,22 @@ SHGEMM_BETA =  gemm_beta_rvv.c
 endif
 endif
 
+ifeq ($(BUILD_BFLOAT16), 1)
+SBGEMMKERNEL    =  sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_zvl256b.c
+ifneq ($(SBGEMM_UNROLL_M), $(SBGEMM_UNROLL_N))
+SBGEMMINCOPY    =  ../generic/gemm_ncopy_$(SBGEMM_UNROLL_M).c
+SBGEMMITCOPY    =  ../generic/gemm_tcopy_$(SBGEMM_UNROLL_M).c
+SBGEMMINCOPYOBJ =  sbgemm_incopy$(TSUFFIX).$(SUFFIX)
+SBGEMMITCOPYOBJ =  sbgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+SBGEMMONCOPY    =  ../generic/gemm_ncopy_$(SBGEMM_UNROLL_N).c
+SBGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SBGEMM_UNROLL_N).c
+SBGEMMONCOPYOBJ =  sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SBGEMMOTCOPYOBJ =  sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ifndef SBGEMM_BETA
+SBGEMM_BETA =  gemm_beta_rvv.c
+endif
+endif
+
 SAXPBYKERNEL  = axpby_vector_v2.c
 DAXPBYKERNEL  = axpby_vector_v2.c