diff options
| author | Jan Zielinski <[email protected]> | 2020-03-18 13:36:53 +0100 |
|---|---|---|
| committer | Jan Zielinski <[email protected]> | 2020-03-23 09:20:51 +0000 |
| commit | 8b3b07afc0b97ecff0431486ca57031150985268 (patch) | |
| tree | 61a416d3fd579b9b7f66a47f98e23948f94d5162 | |
| parent | de550805c5d96b17e7b7db4a0c62b29db354fd74 (diff) | |
gallium/gallivm: Remove workaround disabling AVX code for newer CPUs
The change enables using full 256-bit AVX and AVX2 instructions
on newer platforms.
Reviewed-by: Alok Hota <[email protected]>
Reviewed-by: Adam Jackson <[email protected]>
Reviewed-by: Jose Fonseca <[email protected]>
Reviewed-by: Roland Scheidegger <[email protected]>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4225>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4225>
| -rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_init.c | 12 |
1 files changed, 3 insertions, 9 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c index 4047f2bd781..fd5a35114f4 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_init.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c @@ -435,15 +435,7 @@ lp_build_init(void) } #endif - /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using - * 8-wide vector needs more floating ops than 4-wide (due to padding), it is - * actually more efficient to use 4-wide vectors on this processor. - * - * See also: - * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2 - */ - if (util_cpu_caps.has_avx && - util_cpu_caps.has_intel) { + if (util_cpu_caps.has_avx2 || util_cpu_caps.has_avx) { lp_native_vector_width = 256; } else { /* Leave it at 128, even when no SIMD extensions are available. @@ -455,6 +447,7 @@ lp_build_init(void) lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH", lp_native_vector_width); +#if LLVM_VERSION_MAJOR < 4 if (lp_native_vector_width <= 128) { /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by * "util_cpu_caps.has_avx" predicate, and lack the @@ -468,6 +461,7 @@ lp_build_init(void) util_cpu_caps.has_f16c = 0; util_cpu_caps.has_fma = 0; } +#endif #ifdef PIPE_ARCH_PPC_64 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as |