pan/bi: Constant fold texturing lowerings

This ensures we can constant fold the ALU ops used to lower:

* explicit LOD calculations
* array textures
* texture offsets
* multisample indices

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12205>

diff --git a/src/panfrost/bifrost/bi_opt_constant_fold.c b/src/panfrost/bifrost/bi_opt_constant_fold.c
index bf4cae2..f69866c 100644 (file)
--- a/src/panfrost/bifrost/bi_opt_constant_fold.c
+++ b/src/panfrost/bifrost/bi_opt_constant_fold.c
@@ -44,6 +44,8 @@ bi_fold_constant(bi_instr *I, bool *unsupported)
         /* Grab the sources */
         uint32_t a = bi_apply_swizzle(I->src[0].value, I->src[0].swizzle);
         uint32_t b = bi_apply_swizzle(I->src[1].value, I->src[1].swizzle);
+        uint32_t c = bi_apply_swizzle(I->src[2].value, I->src[2].swizzle);
+        uint32_t d = bi_apply_swizzle(I->src[3].value, I->src[3].swizzle);
 
         /* Evaluate the instruction */
         switch (I->op) {
@@ -53,10 +55,30 @@ bi_fold_constant(bi_instr *I, bool *unsupported)
         case BI_OPCODE_MKVEC_V2I16:
                 return (b << 16) | (a & 0xFFFF);
 
+        case BI_OPCODE_MKVEC_V4I8:
+                return (d << 24) | ((c & 0xFF) << 16) | ((b & 0xFF) << 8) | (a & 0xFF);
+
+        case BI_OPCODE_LSHIFT_OR_I32:
+                if (I->not_result || I->src[0].neg || I->src[1].neg)
+                        break;
+
+                return (a << c) | b;
+
+        case BI_OPCODE_F32_TO_U32:
+                if (I->round == BI_ROUND_NONE) {
+                        /* Explicitly clamp to prevent undefined behaviour and
+                         * match hardware rules */
+                        float f = uif(a);
+                        return (f >= 0.0) ? (uint32_t) f : 0;
+                } else
+                        break;
+
         default:
-                *unsupported = true;
-                return 0;
+                break;
         }
+
+        *unsupported = true;
+        return 0;
 }
 
 void

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index d480431..fbe59f6 100644 (file)
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -2297,6 +2297,14 @@ bi_emit_texc_array_index(bi_builder *b, bi_index idx, nir_alu_type T)
 static bi_index
 bi_emit_texc_lod_88(bi_builder *b, bi_index lod, bool fp16)
 {
+        /* Precompute for constant LODs to avoid general constant folding */
+        if (lod.type == BI_INDEX_CONSTANT) {
+                uint32_t raw = lod.value;
+                float x = fp16 ? _mesa_half_to_float(raw) : uif(raw);
+                int32_t s32 = CLAMP(x, -16.0f, 16.0f) * 256.0f;
+                return bi_imm_u32(s32 & 0xFFFF);
+        }
+
         /* Sort of arbitrary. Must be less than 128.0, greater than or equal to
          * the max LOD (16 since we cap at 2^16 texture dimensions), and
          * preferably small to minimize precision loss */
@@ -3689,6 +3697,7 @@ bifrost_compile_shader_nir(nir_shader *nir,
         }
 
         if (likely(optimize)) {
+                bi_opt_copy_prop(ctx);
                 bi_opt_constant_fold(ctx);
                 bi_opt_copy_prop(ctx);
                 bi_opt_mod_prop_forward(ctx);