Re: [FFmpeg-devel] [PATCH] libswscale/riscv: Fix syntax of vsetvli
Le maanantaina 3. heinäkuuta 2023, 18.52.13 EEST Khem Raj a écrit : > > *Hopefully* LLVM gets their act together by release 16, and ship a usable > > assembler, rather than tell us to use automatic RVV vectorisation (which > > *is* a release 16 feature, though it was half-baked last time I tried). > I am on clang/llvm trunk ( future 17.0 release ), I will also open an > issue with llvm on github regarding this. If I were you, I would first open an issue on the riscv-v-spec Github to seek clarification (though I'm not sure if it's open for bug submission?). > > >Fixes building with clang > > > > More like bug-compatible work-around than fix, AFAIU. > > I can do it although I did not find a relevant section in spec about > these being optional, I did see examples omitting them though. > > > >| src/libswscale/riscv/rgb2rgb_rvv.S:88:25: error: operand must be > > >| e[8|16|32|64|128|256|512|1024],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu]> > > Do you have a reference to the Github RVV spec to validate this, that I > > overlooked, or it's just misled and misleading spew from LLVM? > Perhaps the latter but I also did not see reference to the former. > Here is a small example > > a.S > #.option arch, +zve32x > # OK > #vsetvli t4, t3, e8, m1, ta, ma > # BAD > vsetvli t4, t3, e8, ta, ma > > clang -target riscv64 -march=rv64izve32x1p0 a.S -c Well, yes but the idea is to keep V disabled in the target flags, and do run- time detection... Especially so far, while RVV hardware remains unobtainium, the compiler can't assume that V is supported, or else... And so LLVM AS has been so far unusable for both FFmpeg and kernel alike. -- 雷米‧德尼-库尔蒙 http://www.remlab.net/ ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] [PATCH] libswscale/riscv: Fix syntax of vsetvli
On Mon, Jul 3, 2023 at 1:00 AM Rémi Denis-Courmont wrote: > > Hi, > > The diff is ok. I even have a slight incline in favour thereof for the sake > of consistency. Yet, I may be vainly pedantic but I do have problems with the > description. > > Le 3 juillet 2023 05:44:12 GMT+03:00, Khem Raj a écrit : > >Add missing operand > > The spec does explicitly make the mask and tail policy mandatory, but I can't > see any such requirements for the group multiplier. To the contrary, there > are several examples in the spec _without_ explicit multiplier. > > In reality, `vsetvli` (and `vsetivli`) merely transfers an immediate value to > the vector configuration register. Missing fields are to be left at zero for > forward compatibility, and `m1` happens to encode as 0b000. > I just replicated what GAS is doing here implicitly. > > Which clang complains about but gcc assumes it to be > >'m1' if not specifiied. > > The technical term for the RISC-V support in LLVM AS version 15 and earlier > is "useless junk". If you want to compile FFmpeg on RISC-V with Clang, you > *must* disable the integrated AS and use binutils GNU/as instead. > > The Linux RISC-V kernel altogether gave up on LLVM entirely, requiring GCC > for RVV, so they're even stricter (and without kernel support, FFmpeg support > is obviously useless). > > *Hopefully* LLVM gets their act together by release 16, and ship a usable > assembler, rather than tell us to use automatic RVV vectorisation (which *is* > a release 16 feature, though it was half-baked last time I tried). > I am on clang/llvm trunk ( future 17.0 release ), I will also open an issue with llvm on github regarding this. > > > >Fixes building with clang > > More like bug-compatible work-around than fix, AFAIU. I can do it although I did not find a relevant section in spec about these being optional, I did see examples omitting them though. > > >| src/libswscale/riscv/rgb2rgb_rvv.S:88:25: error: operand must be > >e[8|16|32|64|128|256|512|1024],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu] > > Do you have a reference to the Github RVV spec to validate this, that I > overlooked, or it's just misled and misleading spew from LLVM? Perhaps the latter but I also did not see reference to the former. Here is a small example a.S #.option arch, +zve32x # OK #vsetvli t4, t3, e8, m1, ta, ma # BAD vsetvli t4, t3, e8, ta, ma clang -target riscv64 -march=rv64izve32x1p0 a.S -c > > >| vsetvli t4, t3, e8, ta, ma > >| ^ > > > >Signed-off-by: Khem Raj > >--- > > libswscale/riscv/rgb2rgb_rvv.S | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > >diff --git a/libswscale/riscv/rgb2rgb_rvv.S b/libswscale/riscv/rgb2rgb_rvv.S > >index 5626d906eb..bbdfdbebbc 100644 > >--- a/libswscale/riscv/rgb2rgb_rvv.S > >+++ b/libswscale/riscv/rgb2rgb_rvv.S > >@@ -85,7 +85,7 @@ func ff_interleave_bytes_rvv, zve32x > > mv t3, a3 > > addia4, a4, -1 > > 2: > >-vsetvlit4, t3, e8, ta, ma > >+vsetvlit4, t3, e8, m1, ta, ma > > subt3, t3, t4 > > vle8.v v8, (t0) > > addt0, t4, t0 ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] [PATCH] libswscale/riscv: Fix syntax of vsetvli
Hi, The diff is ok. I even have a slight incline in favour thereof for the sake of consistency. Yet, I may be vainly pedantic but I do have problems with the description. Le 3 juillet 2023 05:44:12 GMT+03:00, Khem Raj a écrit : >Add missing operand The spec does explicitly make the mask and tail policy mandatory, but I can't see any such requirements for the group multiplier. To the contrary, there are several examples in the spec _without_ explicit multiplier. In reality, `vsetvli` (and `vsetivli`) merely transfers an immediate value to the vector configuration register. Missing fields are to be left at zero for forward compatibility, and `m1` happens to encode as 0b000. > Which clang complains about but gcc assumes it to be >'m1' if not specifiied. The technical term for the RISC-V support in LLVM AS version 15 and earlier is "useless junk". If you want to compile FFmpeg on RISC-V with Clang, you *must* disable the integrated AS and use binutils GNU/as instead. The Linux RISC-V kernel altogether gave up on LLVM entirely, requiring GCC for RVV, so they're even stricter (and without kernel support, FFmpeg support is obviously useless). *Hopefully* LLVM gets their act together by release 16, and ship a usable assembler, rather than tell us to use automatic RVV vectorisation (which *is* a release 16 feature, though it was half-baked last time I tried). > >Fixes building with clang More like bug-compatible work-around than fix, AFAIU. >| src/libswscale/riscv/rgb2rgb_rvv.S:88:25: error: operand must be >e[8|16|32|64|128|256|512|1024],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu] Do you have a reference to the Github RVV spec to validate this, that I overlooked, or it's just misled and misleading spew from LLVM? >| vsetvli t4, t3, e8, ta, ma >| ^ > >Signed-off-by: Khem Raj >--- > libswscale/riscv/rgb2rgb_rvv.S | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > >diff --git a/libswscale/riscv/rgb2rgb_rvv.S b/libswscale/riscv/rgb2rgb_rvv.S >index 5626d906eb..bbdfdbebbc 100644 >--- a/libswscale/riscv/rgb2rgb_rvv.S >+++ b/libswscale/riscv/rgb2rgb_rvv.S >@@ -85,7 +85,7 @@ func ff_interleave_bytes_rvv, zve32x > mv t3, a3 > addia4, a4, -1 > 2: >-vsetvlit4, t3, e8, ta, ma >+vsetvlit4, t3, e8, m1, ta, ma > subt3, t3, t4 > vle8.v v8, (t0) > addt0, t4, t0 ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-devel] [PATCH] libswscale/riscv: Fix syntax of vsetvli
Add missing operand which clang complains about but gcc assumes it to be 'm1' if not specifiied. Fixes building with clang | src/libswscale/riscv/rgb2rgb_rvv.S:88:25: error: operand must be e[8|16|32|64|128|256|512|1024],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu] | vsetvli t4, t3, e8, ta, ma | ^ Signed-off-by: Khem Raj --- libswscale/riscv/rgb2rgb_rvv.S | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/libswscale/riscv/rgb2rgb_rvv.S b/libswscale/riscv/rgb2rgb_rvv.S index 5626d906eb..bbdfdbebbc 100644 --- a/libswscale/riscv/rgb2rgb_rvv.S +++ b/libswscale/riscv/rgb2rgb_rvv.S @@ -85,7 +85,7 @@ func ff_interleave_bytes_rvv, zve32x mv t3, a3 addia4, a4, -1 2: -vsetvlit4, t3, e8, ta, ma +vsetvlit4, t3, e8, m1, ta, ma subt3, t3, t4 vle8.v v8, (t0) addt0, t4, t0 -- 2.41.0 ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".