Merge pull request #1695 from ROCm/2nd-backport-6.3

Backport bug fixes to release 6.3

mlir/include/mlir/Dialect/MIGraphX/IR/MIGraphX.td

@@ -98,7 +98,7 @@ def MIGraphX_ClipOp :
 // Keep that logic here.
 def MIGraphX_WhereOp :
     MIGraphX_Op<"where">,
-    Arguments<(ins MIXRShapedOf<[I8]>:$cond,
+    Arguments<(ins MIXRShapedOf<[I8, SI8, UI8]>:$cond,
       AnyMIXRShaped:$inA,
       AnyMIXRShaped:$inB)>,
     Results<(outs AnyMIXRShaped:$output)> {
@@ -117,18 +117,14 @@ def MIGraphX_WhereOp :
 
 def MIGraphX_ConvertOp :
     MIGraphX_Op<"convert">,
-    Arguments<(ins AnyMIXRShaped:$inA, UnitAttr:$zeroExtend)>,
+    Arguments<(ins AnyMIXRShaped:$inA)>,
 	  Results<(outs AnyMIXRShaped:$output)> {
   let summary = "Elementwise type conversion";
   let description = [{
     Type conversion. Due to impedance mismatches between MIGraphX and Tosa,
     currently only supports float to float conversions
-
-    If zeroExtend is set, the input is treated as an unsigned integer.
-    This is MLIR-specific, since MIGraphX encodes integer signedness in types
-    but MLIR generally uses signless integers.
   }];
-  let assemblyFormat = "(`zero_extend` $zeroExtend^)? $inA attr-dict `:` type($inA) `to` type($output)";
+  let assemblyFormat = "$inA attr-dict `:` type($inA) `to` type($output)";
 }
 
 class MIGraphX_ElementwiseUnaryOp<string name, list<Trait> traits=[]> :
@@ -181,10 +177,9 @@ def MIGraphX_TanhOp :
 
 // int4 operations
 def MIGraphX_UnpackOp : MIGraphX_Op<"unpack">,
-    Arguments<(ins MIXRShapedOf<[I8, I<4>]>:$in,
-                   I64Attr:$axis,
-                   BoolAttr:$isUnsigned)>,
-    Results<(outs MIXRShapedOf<[I8, I<4>]>:$out)> {
+    Arguments<(ins MIXRShapedOf<[I8, UI8, SI8, I<4>, SI<4>, UI<4>]>:$in,
+                   I64Attr:$axis)>,
+    Results<(outs MIXRShapedOf<[I8, UI8, SI8, I<4>, SI<4>, UI<4>]>:$out)> {
   let summary = "Unpack int4 vaules stored as bytes";
   let description = [{
     Given a shaped tensor of bytes, double the length of `axis` by
@@ -201,9 +196,6 @@ def MIGraphX_UnpackOp : MIGraphX_Op<"unpack">,
     the corresponding tensor of i8 (in which case, the `i4` are exposed as an
     extra dimension and not flattened) or another tensor of i4. This allows us to
     progressively move unpack up to function boundaries.
-
-    If `isUnsigned` is true, the inputs are a buffer of unsigned 4-bit ints,
-    otherwise they are signed.
   }];
 
   let assemblyFormat = [{ $in attr-dict `:` type($in) `->` type($out) }];
@@ -333,7 +325,7 @@ class MIGraphX_ConvOpBase<string mnemonic, list<Type> inputTypes=[], list<Type>
 }
 
 def MIGraphX_QuantConvolutionOp :
-    MIGraphX_ConvOpBase<"quant_convolution", [F8E4M3FNUZ, F8E5M2FNUZ, F8E5M2, F8E4M3FN, I8], [F32, I32]> {
+    MIGraphX_ConvOpBase<"quant_convolution", [F8E4M3FNUZ, F8E5M2FNUZ, F8E5M2, F8E4M3FN, I8, SI8], [F32, I32, SI32]> {
   let summary = "quantized convolution forward";
   let description = [{
     The `migraphx.quant_convolution` op computes quantized convolution forward.
@@ -510,7 +502,7 @@ class MIGraphX_DotOpBase<string mnemonic, list<Type> inputTypes=[], list<Type> o
 }
 
 def MIGraphX_QuantDotOp :
-    MIGraphX_DotOpBase<"quant_dot", [F8E4M3FNUZ, F8E5M2FNUZ, F8E4M3FN, F8E5M2, I8], [F32, I32]>{
+    MIGraphX_DotOpBase<"quant_dot", [F8E4M3FNUZ, F8E5M2FNUZ, F8E4M3FN, F8E5M2, I8, SI8], [F32, I32, SI32]>{
   let summary = "Dot product of quantized tensors";
   let description = [{
     The `migraphx.quant_dot` op computes the dot product of two tensors.

mlir/include/mlir/Dialect/Rock/IR/RockAttrDefs.td

@@ -472,11 +472,6 @@ def Rock_EnableSplitKForTuning : Rock_Attr<"EnableSplitKForTuning"> {
   let mnemonic = "enable_splitk_for_tuning";
 }
 
-// Used when converting 1-dimensional migraphx.convolution to tosa.conv2d
-def Rock_ExpandedFrom1D : Rock_Attr<"ExpandedFrom1D"> {
-  let mnemonic = "expanded_from_1d";
-}
-
 def Rock_PrefillAttr : Rock_Attr<"Prefill"> {
   let mnemonic = "rock.prefill";