spriv: Produce correct result for GLSLstd450Step with NaN

NOTE: This commit needs "nir: All set-on-comparison opcodes can take all
float types" or regressions will occur in other Vulkan SPIR-V tests.

No shader-db changes on any Intel platform.

NOTE: This commit depends on "nir: All set-on-comparison opcodes can
take all float types".

v2: Fix handling 16-bit (and presumably 64-bit) values.

About 280 shaders in Talos are hurt by a few instructions, and a couple
shaders in Doom 2016 are hurt by a few instructions.

Tiger Lake
Instructions in all programs: 159893290 -> 159895026 (+0.0%)
SENDs in all programs: 6936431 -> 6936431 (+0.0%)
Loops in all programs: 38385 -> 38385 (+0.0%)
Cycles in all programs: 7019260087 -> 7019254134 (-0.0%)
Spills in all programs: 101389 -> 101389 (+0.0%)
Fills in all programs: 131532 -> 131532 (+0.0%)

Ice Lake
Instructions in all programs: 143624235 -> 143625691 (+0.0%)
SENDs in all programs: 6980289 -> 6980289 (+0.0%)
Loops in all programs: 38383 -> 38383 (+0.0%)
Cycles in all programs: 8440083238 -> 8440090702 (+0.0%)
Spills in all programs: 102246 -> 102246 (+0.0%)
Fills in all programs: 131908 -> 131908 (+0.0%)

Skylake
Instructions in all programs: 134185495 -> 134186618 (+0.0%)
SENDs in all programs: 6938790 -> 6938790 (+0.0%)
Loops in all programs: 38356 -> 38356 (+0.0%)
Cycles in all programs: 8222366923 -> 8222365826 (-0.0%)
Spills in all programs: 98821 -> 98821 (+0.0%)
Fills in all programs: 125218 -> 125218 (+0.0%)

Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Fixes: 1feeee9cf47 ("nir/spirv: Add initial support for GLSL 4.50 builtins")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13999>

diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c
index d6ab4c0..d596c25 100644 (file)
--- a/src/compiler/spirv/vtn_glsl450.c
+++ b/src/compiler/spirv/vtn_glsl450.c
@@ -352,9 +352,28 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
       break;
    }
 
-   case GLSLstd450Step:
-      dest->def = nir_sge(nb, src[1], src[0]);
+   case GLSLstd450Step: {
+      /* The SPIR-V Extended Instructions for GLSL spec says:
+       *
+       *    Result is 0.0 if x < edge; otherwise result is 1.0.
+       *
+       * Here src[1] is x, and src[0] is edge.  The direct implementation is
+       *
+       *    bcsel(src[1] < src[0], 0.0, 1.0)
+       *
+       * This is effectively b2f(!(src1 < src0)).  Previously this was
+       * implemented using sge(src1, src0), but that produces incorrect
+       * results for NaN.  Instead, we use the identity b2f(!x) = 1 - b2f(x).
+       */
+      const bool exact = nb->exact;
+      nb->exact = true;
+
+      nir_ssa_def *cmp = nir_slt(nb, src[1], src[0]);
+
+      nb->exact = exact;
+      dest->def = nir_fsub(nb, nir_imm_floatN_t(nb, 1.0f, cmp->bit_size), cmp);
       break;
+   }
 
    case GLSLstd450Length:
       dest->def = nir_fast_length(nb, src[0]);