Merge pull request #13 from Xilinx/matthias.tosa_bitint

TOSA: Test 2 bit integer matmul/cast and 3 bit unsigned cast

Author: mgehre-amd

mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td

@@ -38,8 +38,6 @@ class Tosa_QuantizedType<string n, list<int> params, bit signed>
 // Used to express accumulator results or compare results.
 //===----------------------------------------------------------------------===//
 
-def Tosa_UInt8 : UI<8>;
-
 def Tosa_Int8 : I<8>;
 def Tosa_Int16 : I<16>;
 def Tosa_Int32 : I<32>;
@@ -54,9 +52,11 @@ def Tosa_SignedInt : AnyTypeOf<[Tosa_Int8,
 
 def Tosa_Bool : I<1>;
 
-// No unsigned unquantized int types.
 def Tosa_Int : AnyTypeOf<[Tosa_Bool,
-                          Tosa_UInt8,
+                          AnyUnsignedInteger,
+                          AnySignlessInteger,
+// TODO: For backwards compatibility, keep Tosa_SignedInt, which is actually
+// a set of signless types.
                           Tosa_SignedInt]>;
 
 def Tosa_Int32Or64 : AnyTypeOf<[Tosa_Int32,

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

@@ -469,7 +469,8 @@ createLinalgBodyCalculationForElementwiseOp(Operation *op, ValueRange args,
                                             args.front(), zero);
     }
 
-    if (arith::FPToSIOp::areCastCompatible(srcTy, dstTy)) {
+    if (dstTy.isSignlessInteger() &&
+        arith::FPToSIOp::areCastCompatible(srcTy, dstTy)) {
       auto intMin = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getF32FloatAttr(
                    APInt::getSignedMinValue(dstTy.getIntOrFloatBitWidth())
@@ -487,6 +488,30 @@ createLinalgBodyCalculationForElementwiseOp(Operation *op, ValueRange args,
       return rewriter.create<arith::FPToSIOp>(loc, dstTy, clamped);
     }
 
+    if (dstTy.isUnsignedInteger() &&
+        arith::FPToUIOp::areCastCompatible(srcTy, dstTy)) {
+      auto intMin = rewriter.create<arith::ConstantOp>(
+          loc, rewriter.getF32FloatAttr(
+                   APInt::getMinValue(dstTy.getIntOrFloatBitWidth())
+                       .getZExtValue()));
+
+      auto intMax = rewriter.create<arith::ConstantOp>(
+          loc, rewriter.getF32FloatAttr(
+                   APInt::getMaxValue(dstTy.getIntOrFloatBitWidth())
+                       .getZExtValue()));
+
+      auto rounded = rewriter.create<math::RoundEvenOp>(loc, args[0]);
+
+      auto clamped = clampFloatHelper(loc, rounded, intMin, intMax, rewriter);
+
+      auto cast = rewriter.create<arith::FPToUIOp>(
+          loc, rewriter.getIntegerType(dstTy.getIntOrFloatBitWidth()), clamped);
+      // arith is signless, so temporarily cast back to being unsigned.
+      return rewriter
+          .create<UnrealizedConversionCastOp>(loc, dstTy, cast->getResult(0))
+          .getResult(0);
+    }
+
     // Casting to boolean, integers need to only be checked as not-equal to
     // zero.
     if (srcTy.isa<IntegerType>() && dstTy.isInteger(1)) {

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-i2.mlir

@@ -0,0 +1,20 @@
+// RUN: mlir-opt --split-input-file -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" %s -verify-diagnostics -o -| FileCheck %s
+
+func.func @test_cast(%arg0: tensor<1xf32>) -> tensor<1xf32> {
+  // CHECK: linalg.generic
+  // CHECK: arith.constant -2.000000e+00
+  // CHECK: arith.constant 1.000000e+00
+  // CHECK: math.roundeven
+  // CHECK: arith.minf
+  // CHECK: arith.maxf
+  // CHECK: arith.fptosi
+  %1 = "tosa.cast"(%arg0) : (tensor<1xf32>) -> tensor<1xi2>
+  
+  // CHECK: linalg.generic
+  // CHECK: arith.sitofp
+  %2 = "tosa.cast"(%1) : (tensor<1xi2>) -> tensor<1xf32>
+
+  return %2 : tensor<1xf32>
+}
+
+// -----

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named-i2.mlir

@@ -0,0 +1,25 @@
+// RUN: mlir-opt --split-input-file -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named))" %s -verify-diagnostics -o -| FileCheck %s
+
+// CHECK-LABEL: @matmul
+func.func @matmul(%arg0: tensor<1x5x3xi2>, %arg1: tensor<1x3x6xi2>) -> (tensor<1x5x6xi2>) {
+  // CHECK: [[C0:%.+]] = arith.constant 0 : i2
+  // CHECK: [[INIT:%.+]] = tensor.empty()
+  // CHECK: [[FILLED:%.+]] = linalg.fill ins([[C0]] : i2) outs([[INIT]] : tensor<1x5x6xi2>) -> tensor<1x5x6xi2>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xi2>, tensor<1x3x6xi2>) outs([[FILLED]] : tensor<1x5x6xi2>) -> tensor<1x5x6xi2>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xi2>, tensor<1x3x6xi2>)  -> (tensor<1x5x6xi2>)
+  return %0 : tensor<1x5x6xi2>
+}
+
+// -----
+
+// CHECK-LABEL: @matmul
+func.func @matmul(%arg0: tensor<1x5x3xi2>, %arg1: tensor<1x3x6xi2>) -> (tensor<1x5x6xi4>) {
+  // CHECK: [[C0:%.+]] = arith.constant 0 : i4
+  // CHECK: [[INIT:%.+]] = tensor.empty()
+  // CHECK: [[FILLED:%.+]] = linalg.fill ins([[C0]] : i4) outs([[INIT]] : tensor<1x5x6xi4>) -> tensor<1x5x6xi4>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xi2>, tensor<1x3x6xi2>) outs([[FILLED]] : tensor<1x5x6xi4>) -> tensor<1x5x6xi4>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xi2>, tensor<1x3x6xi2>)  -> (tensor<1x5x6xi4>)
+  return %0 : tensor<1x5x6xi4>
+}
+
+// -----

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-ui3.mlir

@@ -0,0 +1,16 @@
+// RUN: mlir-opt --split-input-file -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" %s | FileCheck %s
+
+func.func @test_cast(%arg0: tensor<1xf32>) -> tensor<1xui3> {
+  // CHECK: linalg.generic
+  // CHECK: arith.constant 0.000000e+00
+  // CHECK: arith.constant 7.000000e+00
+  // CHECK: math.roundeven
+  // CHECK: arith.minf
+  // CHECK: arith.maxf
+  // CHECK: arith.fptoui {{.*}} : f32 to i3
+  // CHECK: builtin.unrealized_conversion_cast
+  %1 = "tosa.cast"(%arg0) : (tensor<1xf32>) -> tensor<1xui3>
+
+  return %1 : tensor<1xui3>
+}
+