radeonsi: enable nir_group_loads for better performance

The best case I have is one viewperf subtest getting +9% performance.

56979 shaders in 34726 tests
Totals:
SGPRS: 2667522 -> 2669178 (0.06 %)
VGPRS: 1543608 -> 1553472 (0.64 %)
Spilled SGPRs: 4090 -> 4100 (0.24 %)
Spilled VGPRs: 1600 -> 1791 (11.94 %)
Private memory VGPRs: 256 -> 256 (0.00 %)
Scratch size: 1872 -> 2076 (10.90 %) dwords per thread
Code Size: 59443980 -> 59479804 (0.06 %) bytes
Max Waves: 867280 -> 865634 (-0.19 %)

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>

v2: No change in pixels but the hash changed.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13604>

src/gallium/drivers/radeonsi/ci/traces-radeonsi.yml

@@ -37,7 +37,7 @@ traces:
   - path: gputest/pixmark-piano.trace
     expectations:
       - device: gl-radeonsi-stoney
-        checksum: 58a86d233d03e2a174cb79c16028f916
+        checksum: a7317d54d452d19ce630c7f554f2279b
   - path: gputest/triangle.trace
     expectations:
       - device: gl-radeonsi-stoney

src/gallium/drivers/radeonsi/si_shader.c

@@ -1426,6 +1426,11 @@ struct nir_shader *si_get_nir_shader(struct si_shader_selector *sel,
                                                        nir_var_shader_out);
    }
 
+   /* This helps LLVM form VMEM clauses and thus get more GPU cache hits.
+    * 200 is tuned for Viewperf. It should be done last.
+    */
+   NIR_PASS_V(nir, nir_group_loads, nir_group_same_resource_only, 200);
+
    return nir;
 }