Mesa (main): radv: improve LDS alignment check for load/store vectorization

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: d2b9c7e982a5070d77eace22f727f5150ce6f397
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=d2b9c7e982a5070d77eace22f727f5150ce6f397

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Tue Jun  1 11:20:00 2021 +0100

radv: improve LDS alignment check for load/store vectorization

Previously, this could vectorize two scalar 16-bit loads into a u8vec4
load.

No fossil-db changes.

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11113>

---

 src/amd/vulkan/radv_pipeline.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index d5ee79bc627..36019505948 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -3124,13 +3124,23 @@ mem_vectorize_callback(unsigned align_mul, unsigned align_offset, unsigned bit_s
    case nir_intrinsic_load_shared:
    case nir_intrinsic_store_shared:
       if (bit_size * num_components ==
-          96) /* 96 bit loads require 128 bit alignment and are split otherwise */
+          96) { /* 96 bit loads require 128 bit alignment and are split otherwise */
          return align % 16 == 0;
-      else if (bit_size * num_components ==
-               128) /* 128 bit loads require 64 bit alignment and are split otherwise */
-         return align % 8 == 0;
-      else
-         return align % (bit_size == 8 ? 2 : 4) == 0;
+      } else if (bit_size == 16 && (align % 4)) {
+         /* AMD hardware can't do 2-byte aligned f16vec2 loads, but they are useful for ALU
+          * vectorization, because our vectorizer requires the scalar IR to already contain vectors.
+          */
+         return (align % 2 == 0) && num_components <= 2;
+      } else {
+         if (num_components == 3) {
+            /* AMD hardware can't do 3-component loads except for 96-bit loads, handled above. */
+            return false;
+         }
+         unsigned req = bit_size * num_components;
+         if (req == 64 || req == 128) /* 64-bit and 128-bit loads can use ds_read2_b{32,64} */
+            req /= 2u;
+         return align % (req / 8u) == 0;
+      }
    default:
       return false;
    }