Small fixes

python/xmos_ai_tools/xinterpreters/src/dll_interpreter.cc

@@ -34,6 +34,7 @@ void add_lib_vision_ops(
   resolver->AddResizeBilinear();
   resolver->AddResizeNearestNeighbor();
   resolver->AddRound();
+  resolver->AddRsqrt();
   resolver->AddStridedSlice();
   resolver->AddSlice();
   resolver->AddSub();

xformer/Transforms/ConvPatterns.td

@@ -8,6 +8,13 @@ include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"
 include "IR/XCoreOps.td"
 include "Utils/Utils.td"
 
+// Unfuse Relu from Conv if output zero point is not -128 for QI8 output
+// XC Conv assumes there is no RELU, or that if there is one, it is being
+// clamped to -128 Add benefit so that this pattern runs first
+def:
+Pat<(TFL_Conv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, AnyTypeOf<[TensorOf<[I32,QI32]>, NoneType]>:$b, $dh, $dw, TFL_AF_Relu, $wf, $sh, $sw),
+  (TFL_ReluOp (TFL_Conv2DOp $input, $f, $b, $dh, $dw, TFL_AF_None, $wf, $sh, $sw, (returnType $output))), [(IsZeroPointNotEqualTo<-128> $output)], [], (addBenefit 20)>;
+
 def CreateNoneValue : NativeCodeCall<"$_builder.create<TFL::NoValueOp>($0."
                                      "getLoc(), $_builder.getUnitAttr())">;
 

xformer/Transforms/ReplaceAddSub.cpp

@@ -39,7 +39,7 @@ LogicalResult replaceAddorSub(T addOp, PatternRewriter &rewriter,
   auto lhsZeroPoint = lhsQType.getZeroPoint();
 
   auto rhsQType = utils::getQType(addOp.getRhs());
-  auto rhsScale =  rhsQType.getScale();
+  auto rhsScale = rhsQType.getScale();
   auto rhsZeroPoint = rhsQType.getZeroPoint();
 
   auto outputQType = utils::getQType(addOp.getOutput());
@@ -56,7 +56,7 @@ LogicalResult replaceAddorSub(T addOp, PatternRewriter &rewriter,
   int shift = int(floor(log2(pow(2, 14) / maxR)));
 
   // For doing subtraction with add op
-  rhsRatio = negateForSub? -rhsRatio: rhsRatio;
+  rhsRatio = negateForSub ? -rhsRatio : rhsRatio;
 
   // Multipliers are converted to fixed-point
   int m1 = round(lhsRatio * pow(2, shift));

xformer/Transforms/ReplaceConcat.cpp

@@ -62,7 +62,8 @@ struct SplitConcatPattern : public OpRewritePattern<TFL::ConcatenationOp> {
     auto outputType = concatOp.getOutput().getType().cast<RankedTensorType>();
     Type elementType = outputType.getElementType();
     ArrayRef<int64_t> outputShape = outputType.getShape();
-    const int axis = concatOp.getAxis();
+    const int axis = concatOp.getAxis() == -1 ? outputType.getRank() - 1
+                                              : concatOp.getAxis();
 
     int axisShape = 0;
     for (int i = 0; i < CONCAT_OP_MAX_INPUTS; i++) {

xformer/Transforms/TFLPatterns.td

@@ -56,8 +56,9 @@ def : Pat<(TFL_ReluOp(TFL_MinimumOp $lhs, $rhs)), (TFL_ReluOp $lhs),
           [(IsSplatAndEqualTo<127> $rhs)]>;
 
 // Merge Relu with Conv
-def : Pat<(TFL_ReluOp(TFL_Conv2DOp $input, $f, $b, $dh, $dw, TFL_AF_None, $p,
-                      $sh, $sw)),
+def : Pat<(TFL_ReluOp
+           : $output(TFL_Conv2DOp $input, $f, $b, $dh, $dw, TFL_AF_None, $p,
+                     $sh, $sw)),
           (TFL_Conv2DOp $input, $f, $b, $dh, $dw, TFL_AF_Relu, $p, $sh, $sw)>;
 
 // Unfuse activation functions from binary ops

xformer/Utils/Utils.td

@@ -57,6 +57,11 @@ class IsSplatAndEqualTo<int n>
                 "dyn_cast<TFL::QConstOp>($0.getDefiningOp()).getValue()."
                 "cast<DenseElementsAttr>().getSplatValue<int8_t>() == " #n>>;
 
+class IsZeroPointNotEqualTo<int n>
+    : Constraint<
+          CPred<"dyn_cast<mlir::quant::UniformQuantizedType>($0.getType().cast<"
+                "ShapedType>().getElementType()).getZeroPoint() != " #n>>;
+
 // Get the dimension size as integer attr.
 class GetDimAsI32<int n>
     : NativeCodeCall<