brw: Properly handle cooperative matrices created with constants

Expand constant sources to cover the region read by DPAS, and also
use NULL register as accumulator when possible.

Reviewed-by: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34373>

src/intel/compiler/brw_from_nir.cpp

@@ -4697,16 +4697,59 @@ brw_from_nir_emit_cs_intrinsic(nir_to_brw_state &ntb,
       const brw_reg_type src_type =
          brw_type_for_nir_type(devinfo, nir_intrinsic_src_type(instr));
 
-      dest = retype(dest, dest_type);
-      brw_reg src0 = retype(get_nir_src(ntb, instr->src[0], 0), dest_type);
+      brw_reg src[3] = {};
+      for (unsigned i = 0; i < ARRAY_SIZE(src); i++) {
+         nir_src nsrc = instr->src[i];
 
-      brw_builder bld16 = bld.exec_all().group(16, 0);
-      brw_builder bldn = devinfo->ver >= 20 ? bld16 : bld.exec_all().group(8, 0);
+         if (!nir_src_is_const(nsrc)) {
+            src[i] = get_nir_src(ntb, nsrc, 0);
+            continue;
+         }
 
-      bldn.DPAS(dest,
-                src0,
-                retype(get_nir_src(ntb, instr->src[2], 0), src_type),
-                retype(get_nir_src(ntb, instr->src[1], 0), src_type),
+         /* A single constant value can be used to fill the entire
+          * cooperative matrix.  In this case get_nir_src() would give a
+          * uniform value (with stride 0), but DPAS can't use regioning,
+          * it needs the full data available in the register.
+          *
+          * So when a source is a constant, allocate the space necessary
+          * and fill it with the constant value.  Except for
+          *
+          *     When Src0 is specified as null, it is treated as an
+          *     immediate value of +0.
+          *
+          * documented in ACM PRM, Vol 2a, "Dot Product Accumulate Systolic".
+          */
+         const unsigned num_components = nir_src_num_components(nsrc);
+         const unsigned bit_size = nir_src_bit_size(nsrc);
+         const nir_const_value *nval = nir_src_as_const_value(instr->src[0]);
+
+         assert(bit_size <= 32);
+         for (unsigned j = 1; j < num_components; j++)
+            assert(nval[0].u32 == nval[j].u32);
+         uint32_t val = nval[0].u32;
+
+         if (i == 0 && val == 0) {
+            src[i] = brw_null_reg();
+
+         } else {
+            unsigned size = bit_size * num_components;
+            unsigned count = size / 32;
+            assert(size % 32 == 0);
+
+            src[i] = bld.vgrf(BRW_TYPE_UD, count);
+            for (unsigned j = 0; j < count; j++)
+               bld.exec_all().MOV(offset(src[i], bld, j), brw_imm_ud(val));
+         }
+      }
+
+      const unsigned dpas_exec_size = devinfo->ver >= 20 ? 16 : 8;
+      brw_builder bldn = bld.exec_all().group(dpas_exec_size, 0);
+
+      /* DPAS uses a different source order: Accumulator, B, A. */
+      bldn.DPAS(retype(dest,   dest_type),
+                retype(src[0], dest_type),
+                retype(src[2], src_type),
+                retype(src[1], src_type),
                 sdepth,
                 rcount)
          ->saturate = nir_intrinsic_saturate(instr);