[Mlir-commits] [mlir] [mlir][TOSA] restore unrealized casts when lowering rescale ops (PR #141096)

[Thomas Preud'homme]

================
@@ -1230,19 +1298,52 @@ func.func @rescale_dyn(%arg0 : tensor<1x?x?x32xi32>) -> () {
 }
 
 // -----
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)>
 
+// CHECK-LABEL: @rescale_i8_unsigned_input_explicit
+// CHECK-SAME: (%[[ARG0:[0-9a-zA-Z_]*]]:
+func.func @rescale_i8_unsigned_input_explicit(%arg0 : tensor<2xui8>) -> () {
+  // CHECK: [[C0:%.+]] = arith.constant 19689
+  // CHECK: [[C1:%.+]] = arith.constant 15
+  // CHECK: [[INIT:%.+]] = tensor.empty()
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel"]} ins(%[[ARG0]] : tensor<2xui8>) outs([[INIT]] : tensor<2xi8>)
+  // CHECK: ^bb0([[IN:%.+]]: ui8, [[UNUSED:%.+]]: i8):
+  // CHECK-DAG: [[C17:%.+]] = arith.constant 17
+  // CHECK-DAG: [[C22:%.+]] = arith.constant 22
+  // CHECK-DAG: [[IN_UTOI:%.+]] = builtin.unrealized_conversion_cast [[IN]] : ui8 to i8
+  // CHECK-DAG: [[IN32:%.+]] = arith.extui [[IN_UTOI]]
+  // CHECK-DAG: [[IN_ZEROED:%.+]] = arith.subi [[IN32]], [[C17]]
+  // CHECK-DAG: [[SCALED:%.+]] = tosa.apply_scale [[IN_ZEROED]], [[C0]], [[C1]] {rounding_mode = "SINGLE_ROUND"}
+  // CHECK-DAG: [[SCALED_ZEROED:%.+]] = arith.addi [[SCALED]], [[C22]]
+  // CHECK-DAG: [[CMIN:%.+]] = arith.constant -128
+  // CHECK-DAG: [[CMAX:%.+]] = arith.constant 127
+  // CHECK-DAG: [[LOWER:%.+]] = arith.maxsi [[CMIN]], [[SCALED_ZEROED]]
+  // CHECK: [[BOUNDED:%.+]] = arith.minsi [[CMAX]], [[LOWER]]
+  // CHECK: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
+  // CHECK: linalg.yield [[TRUNC]]
+  %multiplier = "tosa.const"() {values = dense<19689> : tensor<1xi16> } : () -> tensor<1xi16>
+  %shift = "tosa.const"() {values = dense<15> : tensor<1xi8> } : () -> tensor<1xi8>
+  %input_zp = "tosa.const"() {values = dense<17> : tensor<1xi8>} : () -> tensor<1xi8>
+  %output_zp = "tosa.const"() {values = dense<22> : tensor<1xi8>} : () -> tensor<1xi8>
+  %0 = tosa.rescale %arg0, %multiplier, %shift, %input_zp, %output_zp {scale32 = false, rounding_mode = "SINGLE_ROUND", per_channel = false, input_unsigned = true, output_unsigned = false} : (tensor<2xui8>, tensor<1xi16>, tensor<1xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<2xi8>
+
+  return
+}
+
+// -----
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)>
 
-// CHECK-LABEL: @rescale_i8_unsigned_input
+// CHECK-LABEL: @rescale_i8_unsigned_input_implicit
 // CHECK-SAME: (%[[ARG0:[0-9a-zA-Z_]*]]:
-func.func @rescale_i8_unsigned_input(%arg0 : tensor<2xi8>) -> () {
+func.func @rescale_i8_unsigned_input_implicit(%arg0 : tensor<2xi8>) -> () {
   // CHECK: [[C0:%.+]] = arith.constant 19689
   // CHECK: [[C1:%.+]] = arith.constant 15
   // CHECK: [[INIT:%.+]] = tensor.empty()
   // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel"]} ins(%[[ARG0]] : tensor<2xi8>) outs([[INIT]] : tensor<2xi8>)
   // CHECK: ^bb0([[IN:%.+]]: i8, [[UNUSED:%.+]]: i8):
-  // CHECK: [[C128:%.+]] = arith.constant 128
-  // CHECK: [[C22:%.+]] = arith.constant 22
+  // CHECK-NOT: builtin.unrealized_conversion_cast [[IN]] : i8 to i8
----------------
RoboTux wrote:

Likewise, remove the type info.

https://github.com/llvm/llvm-project/pull/141096