Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 21/11/14 09:02, Siavash Eliasi wrote: On 11/10/2014 04:28 AM, Emil Velikov wrote: I'm not sure did you just said that you've checked it, or that's what it ought to do ? There is a reason why I'm so picky - this bizarre (as one might call it) setup is just the tip of the iceberg when it comes to people building mesa themselves. Would be nice to get your patch in as long as it does not break stuff :) -Emil Hello Emil, David Heidelberg kindly tested the patch on his machines and it was okay. Please consider pushing the patch upstream if it looks good to you guys too. Considering Ian's supporting comment and confirming that it does not break the unique builds I've added David's r-b and pushed this to master. Thanks -Emil ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 11/10/2014 04:28 AM, Emil Velikov wrote: I'm not sure did you just said that you've checked it, or that's what it ought to do ? There is a reason why I'm so picky - this bizarre (as one might call it) setup is just the tip of the iceberg when it comes to people building mesa themselves. Would be nice to get your patch in as long as it does not break stuff :) -Emil Hello Emil, David Heidelberg kindly tested the patch on his machines and it was okay. Please consider pushing the patch upstream if it looks good to you guys too. Best regards, Siavash Eliasi. On 11/15/2014 08:58 PM, David Heidelberg wrote: Hello, I tested it directly on TK-55 laptop, crosscompiled (from my AMD A3870 to TK-55) and on my computer (A3870). Everything worked ok. Tested-by: David Heidelberg da...@ixit.cz On 11/10/2014 04:53 AM, Siavash Eliasi wrote: Hello sir, I've sent a patch to mesa-dev which removes the runtime checks for CPU features which are known to be supported by that target at compile time. Just to make sure that this patch won't break your machine (Athlon TK-55?) with your usual build flags, can you please try this patch and see if that's ok for you? Patch: http://patchwork.freedesktop.org/patch/36488/ Related bug: https://bugs.freedesktop.org/show_bug.cgi?id=71547#c3 Best regards, Siavash Eliasi. ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 11/10/2014 04:28 AM, Emil Velikov wrote: On 09/11/14 04:37, Siavash Eliasi wrote: On 11/08/2014 09:55 PM, Emil Velikov wrote: [...] Can you confirm that it does not cause issues with interesting setups such as https://bugs.freedesktop.org/show_bug.cgi?id=71547 Challenge accepted! What my patch is doing is to check for provided compile flags (-msse, ...) on compile time (__SSE__, ...) and set cpu_has_sse macro to 1 which allows any sane compiler to turn this pieces of code: I'm not sure did you just said that you've checked it, or that's what it ought to do ? There is a reason why I'm so picky - this bizarre (as one might call it) setup is just the tip of the iceberg when it comes to people building mesa themselves. Would be nice to get your patch in as long as it does not break stuff :) -Emil I tried building with same CFLAGS in that bug report, no build failures; ofcourse after applying this patch: http://patchwork.freedesktop.org/patch/36869/ Best regards, Siavash Eliasi. Best regards, Siavash Eliasi. ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On Sun, 2014-11-09 at 14:24 +1100, Timothy Arceri wrote: On Sat, 2014-11-08 at 18:25 +, Emil Velikov wrote: On 08/11/14 08:35, Siavash Eliasi wrote: This will remove the need for unnecessary runtime checks for CPU features if already supported by target CPU, resulting in smaller and less branchy code. A comment I could not withheld based on your earlier post - We require micro-benchmark for this code. It will take me hours to find why mesa is so slow now :P Hehe, you are technically correct, the best kind of correct. Ideally mesa should have an infrastructure/farm that handles regressions - be that performance or otherwise. Pretty sure some companies have such features but those seem to be hidden behind locked doors :'( I considered trying to work on a solution for this where anyone could volunteer their machine to run such tests and send the results back to a central server. But in the end its a pretty big project to get something like this working correctly, and making it easy to setup thus participate in. I guess you could leverage something like the oibaf ppa to start with rather than pulling down each commit and building. Anyway a much bigger project than what I have time for at the moment. Maybe it isn't needed for every commit. Depending on the architecture you are testing you would focus on some commits. For example, I have a radeon with the r600 driver, so I don't need to check all the intel and nvidia related commits. Even if it's not feasible for every commit, we may try to do it for individual series. Or even, we may send some commits with some kind of 'request for test for performance'. But we would need to have some kind of standarized benchmark/tests for this, à la piglit. Just my opinion :) - Bruno But on a more mature note, currently only cpu_has_xmm (_tnl_generate_sse_emit) and cpu_has_sse4_1(vbo_get_minmax_index) are actually useful, with the former of questionable amount :P Can you confirm that it does not cause issues with interesting setups such as https://bugs.freedesktop.org/show_bug.cgi?id=71547 I think this patch should be find in this case as the solution there was to wrap the code with #ifdef __SSE4_1__ which is what makes this patch work. Thanks Emil V2: - Removed the SSSE3 related part for the not yet merged patch. - Avoiding redefinition of macros. --- src/mesa/x86/common_x86_features.h | 26 ++ 1 file changed, 26 insertions(+) diff --git a/src/mesa/x86/common_x86_features.h b/src/mesa/x86/common_x86_features.h index 66f2cf6..65634aa 100644 --- a/src/mesa/x86/common_x86_features.h +++ b/src/mesa/x86/common_x86_features.h @@ -59,13 +59,39 @@ #define X86_CPUEXT_3DNOW_EXT (130) #define X86_CPUEXT_3DNOW (131) +#ifdef __MMX__ +#define cpu_has_mmx 1 +#else #define cpu_has_mmx (_mesa_x86_cpu_features X86_FEATURE_MMX) +#endif + #define cpu_has_mmxext (_mesa_x86_cpu_features X86_FEATURE_MMXEXT) + +#ifdef __SSE__ +#define cpu_has_xmm 1 +#else #define cpu_has_xmm (_mesa_x86_cpu_features X86_FEATURE_XMM) +#endif + +#ifdef __SSE2__ +#define cpu_has_xmm2 1 +#else #define cpu_has_xmm2 (_mesa_x86_cpu_features X86_FEATURE_XMM2) +#endif + +#ifdef __3dNOW__ +#define cpu_has_3dnow1 +#else #define cpu_has_3dnow(_mesa_x86_cpu_features X86_FEATURE_3DNOW) +#endif + #define cpu_has_3dnowext (_mesa_x86_cpu_features X86_FEATURE_3DNOWEXT) + +#ifdef __SSE4_1__ +#define cpu_has_sse4_1 1 +#else #define cpu_has_sse4_1 (_mesa_x86_cpu_features X86_FEATURE_SSE4_1) +#endif #endif ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 09/11/14 04:37, Siavash Eliasi wrote: On 11/08/2014 09:55 PM, Emil Velikov wrote: [...] Can you confirm that it does not cause issues with interesting setups such as https://bugs.freedesktop.org/show_bug.cgi?id=71547 Challenge accepted! What my patch is doing is to check for provided compile flags (-msse, ...) on compile time (__SSE__, ...) and set cpu_has_sse macro to 1 which allows any sane compiler to turn this pieces of code: I'm not sure did you just said that you've checked it, or that's what it ought to do ? There is a reason why I'm so picky - this bizarre (as one might call it) setup is just the tip of the iceberg when it comes to people building mesa themselves. Would be nice to get your patch in as long as it does not break stuff :) -Emil Best regards, Siavash Eliasi. ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 08/11/14 08:35, Siavash Eliasi wrote: This will remove the need for unnecessary runtime checks for CPU features if already supported by target CPU, resulting in smaller and less branchy code. A comment I could not withheld based on your earlier post - We require micro-benchmark for this code. It will take me hours to find why mesa is so slow now :P Ideally mesa should have an infrastructure/farm that handles regressions - be that performance or otherwise. Pretty sure some companies have such features but those seem to be hidden behind locked doors :'( But on a more mature note, currently only cpu_has_xmm (_tnl_generate_sse_emit) and cpu_has_sse4_1(vbo_get_minmax_index) are actually useful, with the former of questionable amount :P Can you confirm that it does not cause issues with interesting setups such as https://bugs.freedesktop.org/show_bug.cgi?id=71547 Thanks Emil V2: - Removed the SSSE3 related part for the not yet merged patch. - Avoiding redefinition of macros. --- src/mesa/x86/common_x86_features.h | 26 ++ 1 file changed, 26 insertions(+) diff --git a/src/mesa/x86/common_x86_features.h b/src/mesa/x86/common_x86_features.h index 66f2cf6..65634aa 100644 --- a/src/mesa/x86/common_x86_features.h +++ b/src/mesa/x86/common_x86_features.h @@ -59,13 +59,39 @@ #define X86_CPUEXT_3DNOW_EXT (130) #define X86_CPUEXT_3DNOW (131) +#ifdef __MMX__ +#define cpu_has_mmx 1 +#else #define cpu_has_mmx (_mesa_x86_cpu_features X86_FEATURE_MMX) +#endif + #define cpu_has_mmxext (_mesa_x86_cpu_features X86_FEATURE_MMXEXT) + +#ifdef __SSE__ +#define cpu_has_xmm 1 +#else #define cpu_has_xmm (_mesa_x86_cpu_features X86_FEATURE_XMM) +#endif + +#ifdef __SSE2__ +#define cpu_has_xmm2 1 +#else #define cpu_has_xmm2 (_mesa_x86_cpu_features X86_FEATURE_XMM2) +#endif + +#ifdef __3dNOW__ +#define cpu_has_3dnow1 +#else #define cpu_has_3dnow(_mesa_x86_cpu_features X86_FEATURE_3DNOW) +#endif + #define cpu_has_3dnowext (_mesa_x86_cpu_features X86_FEATURE_3DNOWEXT) + +#ifdef __SSE4_1__ +#define cpu_has_sse4_1 1 +#else #define cpu_has_sse4_1 (_mesa_x86_cpu_features X86_FEATURE_SSE4_1) +#endif #endif ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On Sat, 2014-11-08 at 18:25 +, Emil Velikov wrote: On 08/11/14 08:35, Siavash Eliasi wrote: This will remove the need for unnecessary runtime checks for CPU features if already supported by target CPU, resulting in smaller and less branchy code. A comment I could not withheld based on your earlier post - We require micro-benchmark for this code. It will take me hours to find why mesa is so slow now :P Hehe, you are technically correct, the best kind of correct. Ideally mesa should have an infrastructure/farm that handles regressions - be that performance or otherwise. Pretty sure some companies have such features but those seem to be hidden behind locked doors :'( I considered trying to work on a solution for this where anyone could volunteer their machine to run such tests and send the results back to a central server. But in the end its a pretty big project to get something like this working correctly, and making it easy to setup thus participate in. I guess you could leverage something like the oibaf ppa to start with rather than pulling down each commit and building. Anyway a much bigger project than what I have time for at the moment. But on a more mature note, currently only cpu_has_xmm (_tnl_generate_sse_emit) and cpu_has_sse4_1(vbo_get_minmax_index) are actually useful, with the former of questionable amount :P Can you confirm that it does not cause issues with interesting setups such as https://bugs.freedesktop.org/show_bug.cgi?id=71547 I think this patch should be find in this case as the solution there was to wrap the code with #ifdef __SSE4_1__ which is what makes this patch work. Thanks Emil V2: - Removed the SSSE3 related part for the not yet merged patch. - Avoiding redefinition of macros. --- src/mesa/x86/common_x86_features.h | 26 ++ 1 file changed, 26 insertions(+) diff --git a/src/mesa/x86/common_x86_features.h b/src/mesa/x86/common_x86_features.h index 66f2cf6..65634aa 100644 --- a/src/mesa/x86/common_x86_features.h +++ b/src/mesa/x86/common_x86_features.h @@ -59,13 +59,39 @@ #define X86_CPUEXT_3DNOW_EXT (130) #define X86_CPUEXT_3DNOW (131) +#ifdef __MMX__ +#define cpu_has_mmx1 +#else #define cpu_has_mmx(_mesa_x86_cpu_features X86_FEATURE_MMX) +#endif + #define cpu_has_mmxext (_mesa_x86_cpu_features X86_FEATURE_MMXEXT) + +#ifdef __SSE__ +#define cpu_has_xmm1 +#else #define cpu_has_xmm(_mesa_x86_cpu_features X86_FEATURE_XMM) +#endif + +#ifdef __SSE2__ +#define cpu_has_xmm2 1 +#else #define cpu_has_xmm2 (_mesa_x86_cpu_features X86_FEATURE_XMM2) +#endif + +#ifdef __3dNOW__ +#define cpu_has_3dnow 1 +#else #define cpu_has_3dnow (_mesa_x86_cpu_features X86_FEATURE_3DNOW) +#endif + #define cpu_has_3dnowext (_mesa_x86_cpu_features X86_FEATURE_3DNOWEXT) + +#ifdef __SSE4_1__ +#define cpu_has_sse4_1 1 +#else #define cpu_has_sse4_1 (_mesa_x86_cpu_features X86_FEATURE_SSE4_1) +#endif #endif ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH V2] mesa: Permanently enable features supported by target CPU at compile time.
On 11/08/2014 09:55 PM, Emil Velikov wrote:
A comment I could not withheld based on your earlier post - We require
micro-benchmark for this code. It will take me hours to find why mesa is
so slow now :P
Which brings the question why didn't you post to that thread/topic in
first place instead :P
Ideally mesa should have an infrastructure/farm that handles regressions
- be that performance or otherwise. Pretty sure some companies have such
features but those seem to be hidden behind locked doors :'(
Yes, that's unfortunate. But atleast we have Phoronix :)
Can you confirm that it does not cause issues with interesting setups
such as https://bugs.freedesktop.org/show_bug.cgi?id=71547
Challenge accepted! What my patch is doing is to check for provided
compile flags (-msse, ...) on compile time (__SSE__, ...) and set
cpu_has_sse macro to 1 which allows any sane compiler to turn this
pieces of code:
#ifdef USE_SSE
if (cpu_has_sse)
{
/* SSE code path */
}
else
#endif
{
/* C fallback */
}
into this:
/* SSE code path */
by using compile time information by target CPU.
Best regards,
Siavash Eliasi.
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
