Does openBLAS or relative mah libarry support half-precision float point in RISC-V

[bjzhang]

Half-precision float point is included in RISC-V RVA23 profile(AKA zfh, zfhmin, zvfh, zvfhmin). We could see that zfh is enabled in some variants[1][2]. But do not see any relative assembly code generated[3][4].

Is there plan to do it. or is there relative packages support or plan to support it? Thanks

[1] The Makefile.riscv64 we found

ifeq ($(CORE), x280)
CCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_zvl512b -mabi=lp64d
FCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d -static
endif

[2] Another relative commit(add zfh)

commit 968e1f51d80bce07678d60fe5e22de557bd798a3
Author: Xianyi Zhang <[email protected]>
Date:   Mon Feb 28 20:33:11 2022 +0800

    Update RISC-V Intrinsic API.

 32 files changed, 316 insertions(+), 249 deletions(-)

diff --git a/Makefile.prebuild b/Makefile.prebuild
index 399db956f..71ea71d4b 100644
--- a/Makefile.prebuild
+++ b/Makefile.prebuild
@@ -46,7 +46,7 @@ TARGET_FLAGS = -mips64r6
 endif
 
 ifeq ($(TARGET), C910V)
-TARGET_FLAGS = -march=rv64gcvxthead -mabi=lp64v
+TARGET_FLAGS = -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d
 endif
 
 all: getarch_2nd
diff --git a/Makefile.riscv64 b/Makefile.riscv64
index 15d7b059c..ce91e03ec 100644
--- a/Makefile.riscv64
+++ b/Makefile.riscv64
@@ -1,4 +1,4 @@
 ifeq ($(CORE), C910V)
-CCOMMON_OPT += -march=rv64gcvxthead -mabi=lp64v
-FCOMMON_OPT += -march=rv64gcvxthead -mabi=lp64v -static
+CCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920
+FCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920 -static
 endif

[3] our Makefile.riscv64
···
ifeq ($(TARGET), RISCV64_ZVL128B)
TARGET_FLAGS = -march=rv64imafdcv_zfh -mabi=lp64d
endif

[4] our full compiling command:
···
/opt/clang-ubuntu241108/bin/riscv64-unknown-linux-gnu-gfortran -O2 -Wall -frecursive -fno-optimize-sibling-calls  -march=rv64imafdcv_zfh -mabi=lp64d -fno-tree-vectorize  -o xzcblat2 c_zblat2.o c_zblas2.o c_z2chke.o auxiliary.o c_xerbla.o constant.o ../libopenblas_riscv64_zvl128bp-r0.3.29.dev.a -lm -lpthread -lgfortran -lm -lpthread -lgfortran -L/opt/clang-ubuntu241108/bin/../lib/riscv64-unknown-linux-gnu -L/opt/clang-ubuntu241108/lib/clang/18/lib/riscv64-unknown-linux-gnu -L/opt/clang-ubuntu241108/bin/../lib/gcc/riscv64-unknown-linux-gnu/15.0.0 -L/opt/clang-ubuntu241108/bin/../lib/gcc/riscv64-unknown-linux-gnu/15.0.0/../../../../riscv64-unknown-linux-gnu/lib -L/opt/clang-ubuntu241108/bin/../sysroot/lib -L/opt/clang-ubuntu241108/bin/../sysroot/usr/lib  -lc
···

[martin-frbg]

The topic of half-precision float comes up about once per year, but I don't think there's ever been sustained interest.

The old commits you noticed are only about providing the default feature string to the compiler, but without any code for FP16 there is nothing where the compiler could use the zfh extension. (All OpenBLAS currently has is BFLOAT16, but I think it should be fairly easy to clone the generic parts from the initial set of BF16 patches to create a framework for FP16 support if_needed)