[Triton-MLIR] Support FP8 (#864)

Co-authored-by: Superjomn <yanchunwei@outlook.com>

lib/Dialect/TritonGPU/IR/Dialect.cpp

@@ -44,7 +44,9 @@ namespace gpu {
 
 // TODO: Inheritation of layout attributes
 unsigned getElemsPerThread(Type type) {
-  if (type.isIntOrIndexOrFloat() || type.isa<triton::PointerType>())
+  if (type.isIntOrIndexOrFloat() ||
+      type.isa<triton::Float8Type>() ||
+      type.isa<triton::PointerType>())
     return 1;
   auto tensorType = type.cast<RankedTensorType>();
   auto layout = tensorType.getEncoding();

lib/Dialect/TritonGPU/Transforms/Coalesce.cpp

@@ -32,7 +32,10 @@ struct CoalescePass : public TritonGPUCoalesceBase<CoalescePass> {
     // Thread tile size depends on memory alignment
     SmallVector<unsigned, 4> sizePerThread(rank, 1);
     PointerType ptrType = origType.getElementType().cast<PointerType>();
-    unsigned numBits = ptrType.getPointeeType().getIntOrFloatBitWidth();
+    auto pointeeType = ptrType.getPointeeType();
+    unsigned numBits =
+        pointeeType.isa<triton::Float8Type>() ?
+        8 : pointeeType.getIntOrFloatBitWidth();
     unsigned maxMultiple = info.getDivisibility(order[0]);
     unsigned maxContig = info.getContiguity(order[0]);
     unsigned alignment = std::min(maxMultiple, maxContig);