Re-enable float8 tensor memory

peterbell10 · peterbell10 · commit b3f7ce4e4f63 · 2025-08-05T21:37:51.000+01:00
diff --git a/lib/Dialect/TritonGPU/IR/Types.cpp b/lib/Dialect/TritonGPU/IR/Types.cpp
@@ -120,11 +120,11 @@ LogicalResult MemDescType::verify(function_ref<InFlightDiagnostic()> emitError,
       return emitError() << "rank must be 2 or 3";
     }
     auto bitwidth = elementType.getIntOrFloatBitWidth();
-    if (!enc.getUnpacked() && bitwidth != 16) {
-      return emitError() << "bitwidth must be 16 for packed tensor memory";
+    if (!enc.getUnpacked() && bitwidth > 16) {
+      return emitError() << "bitwidth must be <= 16 for packed tensor memory";
     }
-    if (bitwidth != 16 && bitwidth != 32) {
-      return emitError() << "bitwidth must be 16 or 32";
+    if (bitwidth > 32) {
+      return emitError() << "bitwidth must be <= 32";
     }
     shape = shape.take_back(2);
     allocShape = allocShape.take_back(2);

Original file line number	Diff line number	Diff line change
`@@ -120,11 +120,11 @@ LogicalResult MemDescType::verify(function_ref<InFlightDiagnostic()> emitError,`
`120`	`120`	`return emitError() << "rank must be 2 or 3";`
`121`	`121`	`}`
`122`	`122`	`auto bitwidth = elementType.getIntOrFloatBitWidth();`
`123`		`- if (!enc.getUnpacked() && bitwidth != 16) {`
`124`		`- return emitError() << "bitwidth must be 16 for packed tensor memory";`
	`123`	`+ if (!enc.getUnpacked() && bitwidth > 16) {`
	`124`	`+ return emitError() << "bitwidth must be <= 16 for packed tensor memory";`
`125`	`125`	`}`
`126`		`- if (bitwidth != 16 && bitwidth != 32) {`
`127`		`- return emitError() << "bitwidth must be 16 or 32";`
	`126`	`+ if (bitwidth > 32) {`
	`127`	`+ return emitError() << "bitwidth must be <= 32";`
`128`	`128`	`}`
`129`	`129`	`shape = shape.take_back(2);`
`130`	`130`	`allocShape = allocShape.take_back(2);`