[mlir][dxsa] Add dcl_max_output_vertex_count instruction

mlir/include/mlir/Dialect/DXSA/IR/DXSAOps.td

@@ -857,4 +857,25 @@ def DXSA_DclGsInstanceCount : DXSA_Op<"dcl_gs_instance_count"> {
   let assemblyFormat = [{ $count attr-dict }];
 }
 
+def DXSA_DclMaxOutputVertexCount :
+    DXSA_Op<"dcl_max_output_vertex_count"> {
+  let summary = "declares the geometry shader maximum output vertex count";
+  let description = [{
+    The `dxsa.dcl_max_output_vertex_count` operation declares the maximum
+    number of vertices that a single invocation of the geometry shader will
+    emit.
+
+    The count must be in [1, 1024].
+
+    Example:
+
+    ```mlir
+    dxsa.dcl_max_output_vertex_count 42
+    ```
+  }];
+  let arguments = (ins ConfinedAttr<I32Attr,
+      [IntPositive, IntMaxValue<1024>]>:$count);
+  let assemblyFormat = [{ $count attr-dict }];
+}
+
 #endif // DXSA_OPS

mlir/lib/Target/DXSA/BinaryParser.cpp

@@ -159,8 +159,8 @@ static void initInstructionInfo(MutableArrayRef<InstructionInfo> instructions) {
       D3D10_SB_DCL_OP);
   SET(D3D10_SB_OPCODE_DCL_GS_OUTPUT_PRIMITIVE_TOPOLOGY, "dcl_outputtopology", 0,
       0x00, D3D10_SB_DCL_OP);
-  SET(D3D10_SB_OPCODE_DCL_MAX_OUTPUT_VERTEX_COUNT, "dcl_maxout", 0, 0x00,
-      D3D10_SB_DCL_OP);
+  SET(D3D10_SB_OPCODE_DCL_MAX_OUTPUT_VERTEX_COUNT,
+      "dcl_max_output_vertex_count", 0, 0x00, D3D10_SB_DCL_OP);
   SET(D3D10_SB_OPCODE_DCL_INPUT_PS, "dcl_input_ps", 1, 0x00, D3D10_SB_DCL_OP);
   SET(D3D10_SB_OPCODE_DCL_CONSTANT_BUFFER, "dcl_constantbuffer", 1, 0x00,
       D3D10_SB_DCL_OP);
@@ -591,6 +591,10 @@ class DXBuilder {
                                             builder.getI32IntegerAttr(count));
   }
 
+  Instruction buildDclMaxOutputVertexCount(uint32_t count, Location loc) {
+    return dxsa::DclMaxOutputVertexCount::create(builder, loc, count);
+  }
+
   Instruction buildDclInputPs(dxsa::InterpolationMode interpolationMode,
                               Operand operand, Location loc) {
     auto interpolationModeAttr = dxsa::InterpolationModeAttr::get(
@@ -1101,6 +1105,19 @@ class Parser {
     return builder.buildDclGsInstanceCount(count, loc);
   }
 
+  FailureOr<Instruction> parseDclMaxOutputVertexCount(Location loc) {
+    auto countToken = parseToken();
+    FAILURE_IF_FAILED(countToken);
+    auto count = *countToken;
+    if (count == 0)
+      return emitError(getLocation(), "max output vertex count cannot be zero");
+    if (count > 1024)
+      return emitError(getLocation(),
+                       "max output vertex count must be <= 1024, got ")
+             << count;
+    return builder.buildDclMaxOutputVertexCount(count, loc);
+  }
+
   FailureOr<dxsa::InterpolationMode>
   parseInterpolationMode(uint32_t opcodeToken, Location loc) {
     auto rawInterpolationMode =
@@ -1301,6 +1318,9 @@ class Parser {
     case D3D11_SB_OPCODE_DCL_GS_INSTANCE_COUNT:
       result = parseDclGsInstanceCount(loc);
       break;
+    case D3D10_SB_OPCODE_DCL_MAX_OUTPUT_VERTEX_COUNT:
+      result = parseDclMaxOutputVertexCount(loc);
+      break;
     case D3D10_SB_OPCODE_DCL_INPUT_PS:
       result = parseDclInputPs(opcodeToken, loc);
       break;

mlir/test/Target/DXSA/dcl_max_output_vertex_count.mlir

@@ -0,0 +1,6 @@
+// RUN: mlir-translate --import-dxsa-bin %S/inputs/dcl_max_output_vertex_count.bin | FileCheck %s
+
+// CHECK:      module {
+// CHECK-NEXT:   dxsa.dcl_max_output_vertex_count 1
+// CHECK-NEXT:   dxsa.dcl_max_output_vertex_count 1024
+// CHECK-NEXT: }

mlir/test/Target/DXSA/dcl_max_output_vertex_count_invalid.mlir

@@ -0,0 +1,14 @@
+// RUN: mlir-opt %s -split-input-file -verify-diagnostics
+
+// expected-error@+1 {{attribute 'count' failed to satisfy constraint: 32-bit signless integer attribute whose value is positive whose maximum value is 1024}}
+dxsa.dcl_max_output_vertex_count -1
+
+// -----
+
+// expected-error@+1 {{attribute 'count' failed to satisfy constraint: 32-bit signless integer attribute whose value is positive whose maximum value is 1024}}
+dxsa.dcl_max_output_vertex_count 0
+
+// -----
+
+// expected-error@+1 {{attribute 'count' failed to satisfy constraint: 32-bit signless integer attribute whose value is positive whose maximum value is 1024}}
+dxsa.dcl_max_output_vertex_count 1025