jay: Implement sample mask writes

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41688>

src/compiler/nir/nir_intrinsics.py

@@ -2662,7 +2662,7 @@ intrinsic("load_reloc_const_intel", dest_comp=1, bit_sizes=[32],
 
 # Write a render target
 # src[] = { color, src0_alpha, omask, depth, stencil, predicate }
-intrinsic("store_render_target_intel", [4, 1, 1, 1, 1, 1], indices=[TARGET], bit_sizes=[32, 32, 16, 32, 32, 1])
+intrinsic("store_render_target_intel", [4, 1, 1, 1, 1, 1], indices=[TARGET], bit_sizes=[32, 32, 32, 32, 32, 1])
 
 # Shuffle with an offset in bytes instead of a lane index.
 # src[] = { payload, lane offset in bytes }

src/intel/compiler/jay/jay_from_nir.c

@@ -809,7 +809,16 @@ jay_emit_fb_write(jay_builder *b, nir_intrinsic_instr *intr)
    if (!jay_is_null(src0_alpha))
       srcs[len++] = jay_as_gpr(b, src0_alpha);
 
-   assert(jay_is_null(omask) && "TODO: samplemask");
+   if (!jay_is_null(omask)) {
+      jay_def packed = jay_alloc_def(b, UGPR, jay_ugpr_per_grf(b->shader));
+      jay_WORD_PACK(b, packed, omask);
+
+      for (unsigned i = 0; i < jay_num_values(packed); i++)
+         srcs[len++] = jay_extract(packed, i);
+
+      /* Split send after omask due to file difference */
+      split = len;
+   }
 
    for (unsigned i = 0; i < 4; i++)
       srcs[len++] = jay_as_gpr(b, jay_extract(colour, i));
@@ -821,7 +830,8 @@ jay_emit_fb_write(jay_builder *b, nir_intrinsic_instr *intr)
       jay_def packed = jay_alloc_def(b, UGPR, jay_ugpr_per_grf(b->shader));
       jay_BYTE_PACK(b, packed, jay_as_gpr(b, stencil));
 
-      /* Split send before stencil */
+      /* Split send before stencil due to file difference */
+      assert(split == -1 && "TODO: samplemask and stencil outputs together");
       split = len;
 
       for (unsigned i = 0; i < jay_num_values(packed); i++)

src/intel/compiler/jay/jay_nir.c

@@ -153,7 +153,6 @@ collect_fragment_output(nir_builder *b, nir_intrinsic_instr *intr, void *ctx_)
    } else if (loc == FRAG_RESULT_STENCIL) {
       out = &ctx->stencil;
    } else if (loc == FRAG_RESULT_SAMPLE_MASK) {
-      UNREACHABLE("todo");
       out = &ctx->sample_mask;
    } else {
       UNREACHABLE("invalid location");

src/intel/compiler/jay/jay_opcodes.py

@@ -172,6 +172,9 @@ op('and_u32_u16', 2, 'u32')
 # This is used for stencil outputs in render target write messages.
 op('byte_pack', 1, 'u32')
 
+# Similar to byte_pack, but for words
+op('word_pack', 1, 'u32')
+
 # Pixel coord calculations. expand_quad replicates out the per-2x2 values from
 # its source g0.[10...13] and - in the case of SIMD32 - g1.[10...13] into a
 # per-lane value. Then offset_packed_pixel_coords adds the appropriate packed

src/intel/compiler/jay/jay_opt_propagate.c

@@ -155,7 +155,8 @@ propagate_forwards(jay_function *f)
       /* Don't propagate into phis yet - TODO: File awareness */
       if (I->op == JAY_OPCODE_PHI_SRC ||
           I->op == JAY_OPCODE_SEND ||
-          I->op == JAY_OPCODE_BYTE_PACK)
+          I->op == JAY_OPCODE_BYTE_PACK ||
+          I->op == JAY_OPCODE_WORD_PACK)
          continue;
 
       jay_foreach_ssa_src(I, s) {

src/intel/compiler/jay/jay_to_binary.c

@@ -493,6 +493,11 @@ emit(struct brw_codegen *p,
               stride(retype(SRC(0), BRW_TYPE_UB), 4, 1, 0));
       break;
 
+   case JAY_OPCODE_WORD_PACK:
+      brw_set_default_exec_size(p, util_logbase2(2 * exec_size));
+      brw_MOV(p, retype(dst, BRW_TYPE_UW), subscript(SRC(0), BRW_TYPE_UW, 0));
+      break;
+
    case JAY_OPCODE_SHR_ODD_SUBSPANS_BY_4:
       brw_SHR(p, dst, SRC(0), brw_imm_uv(0x44440000));
       break;