Change definition of dequantizelinear to match MIGraphX, ONNX (#1567)

While debugging accuracy issues related to an MIGraphX change that
caused the emission of ops like

    migraphx.dequantizelinear %x, %bias, %scale : <...xi8>, <...xi8>, <...xf32>

we discovered that our understanding of MIGraphX's semantics for
dequnatizelinear were incorrect or out of date. The intended
semantics, as seen in MIGraphX's quantiziation simplifier and, more
importantly, in the ONNX reference at
https://github.com/onnx/onnx/blob/c7717bb39c684a6d86c82a5bb1d4c0e5d90353fe/onnx/reference/ops/op_dequantize_linear.py#L42
are that the input and bias have to be cast to the
scale/output (they're required to match) element type before
computation.

We didn't do this, causing validitaion errors due to 8-bit integer
over/unferflow

This commit fixes is the issue. It isn't known if it needs a backport.

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp

@@ -837,8 +837,11 @@ struct SoftmaxConverter final
 };
 } // namespace
 
-// MIGraphX pseudo code:
-// output[i] = static_cast<T>(input[i] - zero_pts[i]) * scales[i];
+// MIGraphX implements:
+// Let T = scale element type
+// output[i] = (convert<T>(input[i]) - convert<T>(zero_pts[i])) * scales[i];
+// For f32, this matches ONNX reference, dequantizing to f16, if it's ever done
+// will be less precise than the reference but that's probably fine.
 LogicalResult DeQuantizeLinearConverter::matchAndRewrite(
     migraphx::DeQuantizeLinearOp op, OpAdaptor adaptor,
     ConversionPatternRewriter &rewriter) const {
@@ -847,34 +850,25 @@ LogicalResult DeQuantizeLinearConverter::matchAndRewrite(
   Value output = op.getOutput();
   Location loc = op->getLoc();
 
-  Value shifted = input;
+  Type outputType = getShapedElementTy(output);
+  Value upcastInput = createCastOp(rewriter, loc, outputType, input);
+
+  Value shifted = upcastInput;
   if (auto bias = adaptor.getBias()) {
-    Type inElemTy = getShapedElementTy(input);
-    Type biasElemTy = getShapedElementTy(bias);
-    Type elementType =
-        inElemTy.getIntOrFloatBitWidth() <= biasElemTy.getIntOrFloatBitWidth()
-            ? biasElemTy
-            : inElemTy;
-    if (inElemTy != elementType)
-      input = createCastOp(rewriter, loc, elementType, shifted);
-    if (biasElemTy != elementType)
-      bias = createCastOp(rewriter, loc, elementType, bias);
-    shifted =
-        createOpAndInfer<tosa::SubOp>(rewriter, loc, elementType, input, bias);
+    Value upcastBias = createCastOp(rewriter, loc, outputType, bias);
+    shifted = createOpAndInfer<tosa::SubOp>(rewriter, loc, outputType,
+                                            upcastInput, upcastBias);
   }
 
-  Type outputType = getShapedElementTy(output);
-  Value upCast = createCastOp(rewriter, loc, outputType, shifted);
-
   Value scaled = createOpAndInfer<tosa::MulOp>(rewriter, loc, outputType,
-                                               upCast, scale, /*shift=*/0);
+                                               shifted, scale, /*shift=*/0);
 
   rewriter.replaceOp(op, scaled);
   return success();
 }
 
 // MIGraphX pseudo code:
-// int64_t quantized = static_cast<int32>(
+// int32_t quantized = static_cast<int32>(
 //      std::round(input[i] / scales[i])) + zero_pts[i];
 // output[i] = std::max(-128, std::min(127, quantized));
 LogicalResult QuantizeLinearConverter::matchAndRewrite(

mlir/test/Conversion/MIGraphXToTosa/mixr-to-tosa-ops.mlir

@@ -28,28 +28,29 @@ module  {
   }
 
   // CHECK-LABEL: func @dequantize_scale_bias
-  // CHECK: tosa.sub
   // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.sub
   // CHECK: tosa.mul
   func.func @dequantize_scale_bias(%arg: !migraphx.shaped<1x112x112x64xi32, 802816x7168x64x1>, %scale: !migraphx.shaped<64xf32, 1>, %bias: !migraphx.shaped<64xi32, 1>) -> !migraphx.shaped<1x112x112x64xf32, 802816x7168x64x1> attributes {kernel = "mixr"} {
     %1 = migraphx.dequantizelinear %arg, %scale, %bias : <1x112x112x64xi32, 802816x7168x64x1>, <64xf32, 1>, !migraphx.shaped<64xi32, 1> -> <1x112x112x64xf32, 802816x7168x64x1>
     return %1 : !migraphx.shaped<1x112x112x64xf32, 802816x7168x64x1>
   }
 
   // CHECK-LABEL: func @dequantize_wide_bias
-  // CHECK: tosa.cast{{.*}}i32
-  // CHECK: tosa.sub{{.*}}i32
   // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.sub{{.*}}f32
   // CHECK: tosa.mul
   func.func @dequantize_wide_bias(%arg: !migraphx.shaped<1x112x112x64xi8, 802816x7168x64x1>, %scale: !migraphx.shaped<64xf32, 1>, %bias: !migraphx.shaped<64xi32, 1>) -> !migraphx.shaped<1x112x112x64xf32, 802816x7168x64x1> attributes {kernel = "mixr"} {
     %1 = migraphx.dequantizelinear %arg, %scale, %bias : <1x112x112x64xi8, 802816x7168x64x1>, <64xf32, 1>, !migraphx.shaped<64xi32, 1> -> <1x112x112x64xf32, 802816x7168x64x1>
     return %1 : !migraphx.shaped<1x112x112x64xf32, 802816x7168x64x1>
   }
 
   // CHECK-LABEL: func @dequantize_wide_input
-  // CHECK: tosa.cast{{.*}}i32
-  // CHECK: tosa.sub{{.*}}i32
   // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.cast{{.*}}f32
+  // CHECK: tosa.sub{{.*}}f32
   // CHECK: tosa.mul
   func.func @dequantize_wide_input(%arg: !migraphx.shaped<1x112x112x64xi32, 802816x7168x64x1>, %scale: !migraphx.shaped<64xf32, 1>, %bias: !migraphx.shaped<64xi8, 1>) -> !migraphx.shaped<1x112x112x64xf32, 802816x7168x64x1> attributes {kernel = "mixr"} {
     %1 = migraphx.dequantizelinear %arg, %scale, %bias : <1x112x112x64xi32, 802816x7168x64x1>, <64xf32, 1>, !migraphx.shaped<64xi8, 1> -> <1x112x112x64xf32, 802816x7168x64x1>
@@ -142,8 +143,9 @@ module  {
 
   // CHECK-LABEL: func @conv_with_quant
   // CHECK: tosa.conv2d{{.*}} quantization_info
-  // CHECK: tosa.sub
   // CHECK: tosa.cast
+  // CHECK: tosa.cast
+  // CHECK: tosa.sub
   // CHECK: tosa.mul
   // CHECK: tosa.reciprocal
   // CHECK: tosa.mul