Philippe Mathieu-Daudé <f4...@amsat.org> writes: > On 05/17/2018 05:30 PM, Richard Henderson wrote: >> On 05/17/2018 10:47 AM, Alex Bennée wrote: >>> This runs through the usual float to float conversions and crucially >>> also runs with ARM Alternative Half Precision Format. >>> >>> Signed-off-by: Alex Bennée <alex.ben...@linaro.org> >>> [rth: tweak vcvtb.f16.f32/vctb.f32.f16 code and regen] >>> Signed-off-by: Richard Henderson <richard.hender...@linaro.org> >>> >>> --- >>> v4 >>> - add fcvt.ref and check results against it >>> - fix single_to_half, single_to_double conversions >>> - properly toggle AHP mode (fpsr->fpcr) >>> - more values around the AHP margins >>> - add INF/NAN/SNAN inputs >>> - build for ARM and AArch64 >>> - fix bug for hex literals >>> - add float-to-int >>> - checkpatch fix >>> --- >> >> Reviewed-by: Richard Henderson <richard.hender...@linaro.org> > > Thanks Richard, I was not very motivated to review each line of the > fcvt.ref files =) > >> diff --git a/tests/tcg/arm/Makefile.target b/tests/tcg/arm/Makefile.target >> index 9d2b551732..7bb777f442 100644 >> --- a/tests/tcg/arm/Makefile.target >> +++ b/tests/tcg/arm/Makefile.target >> @@ -8,7 +8,9 @@ ARM_SRC=$(SRC_PATH)/tests/tcg/arm >> # Set search path for all sources >> VPATH += $(ARM_SRC) >> >> -TESTS += hello-arm test-arm-iwmmxt >> +ARM_TESTS=hello-arm test-arm-iwmmxt >> + >> +TESTS += $(ARM_TESTS) fcvt >> >> hello-arm: CFLAGS+=-marm -ffreestanding >> hello-arm: LDFLAGS+=-nostdlib >> @@ -24,3 +26,14 @@ run-test-mmap: test-mmap >> $(call quiet-command, $(QEMU) -p 8192 $< 8192 > test-mmap-8192.out, > "TEST", "$< (8k pages) on $(TARGET_NAME)") >> $(call quiet-command, $(QEMU) -p 16384 $< 16384 > > test-mmap-16384.out, "TEST", "$< (16k pages) on $(TARGET_NAME)") >> endif >> + >> +ifeq ($(TARGET_NAME), arm) >> +fcvt: LDFLAGS+=-lm >> +# fcvt: CFLAGS+=-march=armv8.2-a+fp16 -mfpu=neon-fp-armv8 > > Alex, What is your idea here, enable this later? > Maybe add some TODO comment around..
How about: # AArch32 doesn't support all combinations of FP16<->FOO in it's base # profile. Once the AArch32 profile is caught up we should build an # additional binary and reference to test the additional combinations. I backed the change out when I realised the stretch compiler didn't support the same range as my local (18.04) gcc's. > > Except this nit, for both Makefile.target: > Reviewed-by: Philippe Mathieu-Daudé <f4...@amsat.org> > >> + >> +run-fcvt: fcvt >> + $(call quiet-command, \ >> + $(QEMU) $< > fcvt.out && \ >> + diff -u $(ARM_SRC)/fcvt.ref fcvt.out, \ >> + "TEST", "$< (default) on $(TARGET_NAME)") >> +endif > > $ make -j1 run-tcg-tests-aarch64-linux-user > BUILD debian9 > BUILD debian-arm64-cross > CROSS-BUILD aarch64 guest-tests with docker qemu:debian-arm64-cross > BUILD debian9 > BUILD debian-arm64-cross > CROSS-BUILD aarch64 guest-tests with docker qemu:debian-arm64-cross > RUN-TESTS for aarch64 > TEST test-mmap (default) on aarch64 > TEST test-mmap (8k pages) on aarch64 > TEST test-mmap (16k pages) on aarch64 > TEST test-mmap (32k pages) on aarch64 > TEST sha1 on aarch64 > TEST linux-test on aarch64 > TEST testthread on aarch64 > TEST fcvt (default) on aarch64 > TEST sysregs on aarch64 > > \o/ > > Tested-by: Philippe Mathieu-Daudé <f4...@amsat.org> -- Alex Bennée
