Re: [Mesa-dev] [PATCH 1/2] gallium: Enable aarch64 NEON CPU detection.

2019-01-23 Thread Matt Turner 
On Wed, Jan 23, 2019 at 3:26 PM Eric Anholt  wrote:
>
> Matt Turner  writes:
>
> > NEON (now called ASIMD) is available on all aarch64 CPUs. It seems that
> > our code was missing an aarch64 path, leading to util_cpu_caps.has_neon
> > always being false on aarch64. I think that means that the NEON tiling
> > code in vc4 would not be enabled on aarch64 (vc4_load_lt_image_neon,
> > etc).
> > ---
> > I have very little clue about aarch64 ABIs, so I don't know if there's
> > another case that needs to be handled -- aarch32 maybe? Does
> > PIPE_ARCH_AARCH64 just mean ARMv8 and so we should check something else
> > for the ABI and choose Elf{32,64} based on that?
>
> Do we actually need to do runtime detection?  It sounds like "standard"
> armv8 is guaranteed NEON:
>
> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0024a/CEGDJGGC.html

I think that's right.
https://gcc.gnu.org/onlinedocs/gcc/AArch64-Options.html seems to
agree:

‘simd’

Enable Advanced SIMD instructions. This also enables floating-point
instructions. This is on by default for all possible values for
options -march and -mcpu.

I'll send a new patch that just sets util_cpu_caps.has_neon = true for aarch64.
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Re: [Mesa-dev] [PATCH 1/2] gallium: Enable aarch64 NEON CPU detection.

2019-01-23 Thread Eric Anholt 
Matt Turner  writes:

> NEON (now called ASIMD) is available on all aarch64 CPUs. It seems that
> our code was missing an aarch64 path, leading to util_cpu_caps.has_neon
> always being false on aarch64. I think that means that the NEON tiling
> code in vc4 would not be enabled on aarch64 (vc4_load_lt_image_neon,
> etc).
> ---
> I have very little clue about aarch64 ABIs, so I don't know if there's
> another case that needs to be handled -- aarch32 maybe? Does
> PIPE_ARCH_AARCH64 just mean ARMv8 and so we should check something else
> for the ABI and choose Elf{32,64} based on that?

Do we actually need to do runtime detection?  It sounds like "standard"
armv8 is guaranteed NEON:

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0024a/CEGDJGGC.html

For vc4, the aarch64 build uses neon in the _base version of the tiling
implementation.


signature.asc
Description: PGP signature
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

[Mesa-dev] [PATCH 1/2] gallium: Enable aarch64 NEON CPU detection.

2019-01-22 Thread Matt Turner 
NEON (now called ASIMD) is available on all aarch64 CPUs. It seems that
our code was missing an aarch64 path, leading to util_cpu_caps.has_neon
always being false on aarch64. I think that means that the NEON tiling
code in vc4 would not be enabled on aarch64 (vc4_load_lt_image_neon,
etc).
---
I have very little clue about aarch64 ABIs, so I don't know if there's
another case that needs to be handled -- aarch32 maybe? Does
PIPE_ARCH_AARCH64 just mean ARMv8 and so we should check something else
for the ABI and choose Elf{32,64} based on that?

Also, Android is not handled.

 src/util/u_cpu_detect.c | 22 +-
 1 file changed, 17 insertions(+), 5 deletions(-)

diff --git a/src/util/u_cpu_detect.c b/src/util/u_cpu_detect.c
index 52b9ae547d4..e9cdb78e458 100644
--- a/src/util/u_cpu_detect.c
+++ b/src/util/u_cpu_detect.c
@@ -328,7 +328,7 @@ PIPE_ALIGN_STACK static inline boolean sse2_has_daz(void)
 
 #endif /* X86 or X86_64 */
 
-#if defined(PIPE_ARCH_ARM)
+#if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
 static void
 check_os_arm_support(void)
 {
@@ -348,24 +348,36 @@ check_os_arm_support(void)
  util_cpu_caps.has_neon = 1;
}
 #elif defined(PIPE_OS_LINUX)
-Elf32_auxv_t aux;
+
+#if defined(PIPE_ARCH_ARM)
+#define Elf_auxv_t Elf32_auxv_t
+#elif defined(PIPE_ARCH_AARCH64)
+#define Elf_auxv_t Elf64_auxv_t
+#endif
+
+Elf_auxv_t aux;
 int fd;
 
 fd = open("/proc/self/auxv", O_RDONLY | O_CLOEXEC);
 if (fd >= 0) {
-   while (read(fd, , sizeof(Elf32_auxv_t)) == sizeof(Elf32_auxv_t)) {
+   while (read(fd, , sizeof(Elf_auxv_t)) == sizeof(Elf_auxv_t)) {
   if (aux.a_type == AT_HWCAP) {
  uint32_t hwcap = aux.a_un.a_val;
 
+#if defined(PIPE_ARCH_ARM)
  util_cpu_caps.has_neon = (hwcap >> 12) & 1;
+#elif defined(PIPE_ARCH_AARCH64)
+ util_cpu_caps.has_neon = (hwcap >> 1) & 1;
+#endif
  break;
   }
}
close (fd);
 }
+#undef Elf_auxv_t
 #endif /* PIPE_OS_LINUX */
 }
-#endif /* PIPE_ARCH_ARM */
+#endif /* PIPE_ARCH_ARM || PIPE_ARCH_AARCH64 */
 
 static void
 get_cpu_topology(void)
@@ -534,7 +546,7 @@ util_cpu_detect_once(void)
}
 #endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */
 
-#if defined(PIPE_ARCH_ARM)
+#if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
check_os_arm_support();
 #endif
 
-- 
2.19.2

___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev