agx: Model interpolation for iter instructions

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23480>

src/asahi/compiler/agx_compile.c

@@ -376,6 +376,8 @@ agx_emit_load_vary(agx_builder *b, agx_index dest, nir_intrinsic_instr *instr)
    /* TODO: Interpolation modes */
    assert(bary != NULL);
    assert(bary->intrinsic == nir_intrinsic_load_barycentric_pixel);
+   enum agx_interpolation interp = AGX_INTERPOLATION_CENTER;
+   agx_index sample_index = agx_zero();
 
    bool perspective =
       nir_intrinsic_interp_mode(bary) != INTERP_MODE_NOPERSPECTIVE;
@@ -395,9 +397,9 @@ agx_emit_load_vary(agx_builder *b, agx_index dest, nir_intrinsic_instr *instr)
    /* For perspective interpolation, we project (multiply by 1/W) */
    if (perspective) {
       agx_index J = agx_get_cf(b->shader, true, false, VARYING_SLOT_POS, 3, 1);
-      agx_iterproj_to(b, dest, I, J, components);
+      agx_iterproj_to(b, dest, I, J, sample_index, components, interp);
    } else {
-      agx_iter_to(b, dest, I, components);
+      agx_iter_to(b, dest, I, sample_index, components, interp);
    }
 
    agx_emit_cached_split(b, dest, components);
@@ -644,7 +646,7 @@ agx_emit_load_frag_coord(agx_builder *b, agx_index dst,
             agx_get_cf(b->shader, true, false, VARYING_SLOT_POS, i, 1);
 
          dests[i] = fp32;
-         agx_iter_to(b, fp32, cf, 1);
+         agx_iter_to(b, fp32, cf, agx_zero(), 1, AGX_INTERPOLATION_CENTER);
       }
    }
 

src/asahi/compiler/agx_compiler.h

@@ -305,6 +305,9 @@ typedef struct {
    bool shadow            : 1;
    enum agx_gather gather : 3;
 
+   /* TODO: Handle iter ops more efficient */
+   enum agx_interpolation interpolation : 2;
+
    /* Final st_vary op */
    bool last : 1;
 

src/asahi/compiler/agx_lower_uniform_sources.c

@@ -42,6 +42,8 @@ should_lower(enum agx_opcode op, agx_index uniform, unsigned src_index)
    case AGX_OPCODE_ST_VARY:
    case AGX_OPCODE_LOCAL_ATOMIC:
    case AGX_OPCODE_SAMPLE_MASK:
+   case AGX_OPCODE_ITER:
+   case AGX_OPCODE_ITERPROJ:
       return true;
    default:
       return false;

src/asahi/compiler/agx_opcodes.py

@@ -153,6 +153,14 @@ ATOMIC_OPC = enum("atomic_opc", {
 	10: 'xor',
 })
 
+INTERPOLATION = enum("interpolation", {
+    0: 'center',
+    1: 'sample',
+    2: 'centroid',
+    # We translate sample -> sample_register at pack time for simplicity
+    3: 'sample_register',
+})
+
 FUNOP = lambda x: (x << 28)
 FUNOP_MASK = FUNOP((1 << 14) - 1)
 
@@ -327,8 +335,14 @@ for is_float in [False, True]:
 
 op("bitop", (0x7E, 0x7F, 6, _), srcs = 2, imms = [TRUTH_TABLE])
 op("convert", (0x3E | L, 0x7F | L | (0x3 << 38), 6, _), srcs = 2, imms = [ROUND]) 
-op("iter", (0x21, 0xBF, 8, _), srcs = 1, imms = [CHANNELS])
+
-op("iterproj", (0x21, 0xBF, 8, _), srcs = 2, imms = [CHANNELS])
+# Sources are the coeffient register and the sample index (if applicable)
+op("iter", (0x21, 0xBF, 8, _), srcs = 2, imms = [CHANNELS, INTERPOLATION])
+
+# Sources are the coeffient register for the varying, the coefficient register
+# for W, and the sample index (if applicable)
+op("iterproj", (0x21, 0xBF, 8, _), srcs = 3, imms = [CHANNELS, INTERPOLATION])
+
 op("ldcf", (0xA1, 0xBF, 8, _), srcs = 1, imms = [CHANNELS])
 op("st_vary", None, dests = 0, srcs = 2, can_eliminate = False)
 op("no_varyings", (0x80000051, 0xFFFFFFFF, 4, _), dests = 0, can_eliminate = False)

src/asahi/compiler/agx_pack.c

@@ -547,14 +547,28 @@ agx_pack_instr(struct util_dynarray *emission, struct util_dynarray *fixups,
       assert(cf_I < 0x100);
       assert(cf_J < 0x100);
 
+      enum agx_interpolation interp = I->interpolation;
+      agx_index sample_index = flat ? agx_null() : I->src[perspective ? 2 : 1];
+
+      /* Fix up the interpolation enum to distinguish the sample index source */
+      if (interp == AGX_INTERPOLATION_SAMPLE) {
+         if (sample_index.type == AGX_INDEX_REGISTER)
+            interp = AGX_INTERPOLATION_SAMPLE_REGISTER;
+         else
+            assert(sample_index.type == AGX_INDEX_IMMEDIATE);
+      } else {
+         sample_index = agx_zero();
+      }
+
       bool kill = false; // TODO: optimize
 
       uint64_t raw =
          0x21 | (flat ? (1 << 7) : 0) | (perspective ? (1 << 6) : 0) |
          ((D & 0xFF) << 7) | (1ull << 15) | /* XXX */
          ((cf_I & BITFIELD_MASK(6)) << 16) | ((cf_J & BITFIELD_MASK(6)) << 24) |
-         (((uint64_t)channels) << 30) | (!flat ? (1ull << 46) : 0) | /* XXX */
+         (((uint64_t)channels) << 30) | (((uint64_t)sample_index.value) << 32) |
-         (kill ? (1ull << 52) : 0) |                                 /* XXX */
+         (!flat ? (1ull << 46) : 0) |                             /* XXX */
+         (((uint64_t)interp) << 48) | (kill ? (1ull << 52) : 0) | /* XXX */
          (((uint64_t)(D >> 8)) << 56) | ((uint64_t)(cf_I >> 6) << 58) |
          ((uint64_t)(cf_J >> 6) << 60);