gallivm: Use 8 wide AoS sampling on AVX2.

v2: Make sure that with num_lods > 1 and min_filter != mag_filter we
still enter the splitting path. So this case would still use 4-wide aos
path (as a side note, the 4-wide aos sampling path could actually be
improved quite a bit if we have avx2, by just doing the filtering with
256bit vectors).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index 1b48fc2a91ea38870ee8bc7968bd223f5209286f..1477a72d6017e90f0c6ee1ba66053b8f98c5fd3a 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -2860,12 +2860,13 @@ lp_build_sample_soa_code(struct gallivm_state *gallivm,
       }
 
       /*
-       * we only try 8-wide sampling with soa as it appears to
-       * be a loss with aos with AVX (but it should work, except
-       * for conformance if min_filter != mag_filter if num_lods > 1).
-       * (It should be faster if we'd support avx2)
+       * we only try 8-wide sampling with soa or if we have AVX2
+       * as it appears to be a loss with just AVX)
        */
-      if (num_quads == 1 || !use_aos) {
+      if (num_quads == 1 || !use_aos ||
+          (util_cpu_caps.has_avx2 &&
+           (bld.num_lods == 1 ||
+            derived_sampler_state.min_img_filter == derived_sampler_state.mag_img_filter))) {
          if (use_aos) {
             /* do sampling/filtering with fixed pt arithmetic */
             lp_build_sample_aos(&bld, sampler_index,