[BACKEND] Fix promoteOperand behavior in AccelerateMatmul for SM < 80 (#7158)

Dot op using MMA for compute capability < 80 has been deprecated. It
falls back to the FMA path. In this path, `promoteOperand` used
`triton::FpToFpOp` unconditionally, which supports `F8 <-> FP16, BF16,
FP32, FP64` conversions.

This change introduces an `ElementType` check in `promoteOperand`: if
the operand’s element type is **not** FP8, it uses `arith::ExtFOp`
instead of `triton::FpToFpOp`.

Author: sBobHuang

lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp

@@ -6,6 +6,7 @@
 #include "mlir/Support/LogicalResult.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 #include "triton/Analysis/Utility.h"
+#include "triton/Conversion/MLIRTypes.h"
 #include "triton/Dialect/Triton/IR/Dialect.h"
 #include "triton/Dialect/Triton/IR/Utility.h"
 #include "triton/Dialect/TritonGPU/IR/Attributes.h"
@@ -757,7 +758,12 @@ static Value promoteOperand(OpBuilder &builder, Location loc, Value operand,
                             Type promotedType) {
   Type tensorPromotedType = cast<RankedTensorType>(operand.getType())
                                 .cloneWith(std::nullopt, promotedType);
-  return builder.create<FpToFpOp>(loc, tensorPromotedType, operand);
+  Type operandElType =
+      cast<RankedTensorType>(operand.getType()).getElementType();
+  if (type::isFloat8(operandElType)) {
+    return builder.create<FpToFpOp>(loc, tensorPromotedType, operand);
+  }
+  return builder.create<arith::ExtFOp>(loc, tensorPromotedType, operand);
 }
 
 // promote operands of dot op if the existing combination is not natively

test/TritonGPU/accelerate-matmul.mlir

@@ -549,3 +549,20 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 8 : i32, ttg.targ
     tt.return
   }
 }
+
+// -----
+
+#blocked = #ttg.blocked<{sizePerThread = [4, 4], threadsPerWarp = [1, 32], warpsPerCTA = [4, 2], order = [1, 0]}>
+module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 8 : i32, ttg.target = "cuda:75", "ttg.threads-per-warp" = 32 : i32} {
+  // CHECK-LABEL: dot_fall_back_fma_before_ampere
+  tt.func public @dot_fall_back_fma_before_ampere(%arg0: tensor<128x64xf16, #ttg.dot_op<{opIdx = 0, parent = #blocked}>>, %arg1: tensor<64x256xf16, #ttg.dot_op<{opIdx = 1, parent = #blocked}>>, %arg2: tensor<128x256x!tt.ptr<f32>, #blocked>) {
+    %cst = arith.constant dense<0.000000e+00> : tensor<128x256xf32, #blocked>
+    // CHECK:   %[[EXT0:.*]] = arith.extf %arg0
+    // CHECK:   %[[EXT1:.*]] = arith.extf %arg1
+    // CHECK:   %[[DOT:.*]] = tt.dot %[[EXT0]], %[[EXT1]]
+    %0 = tt.dot %arg0, %arg1, %cst, inputPrecision = tf32 : tensor<128x64xf16, #ttg.dot_op<{opIdx = 0, parent = #blocked}>> * tensor<64x256xf16, #ttg.dot_op<{opIdx = 1, parent = #blocked}>> -> tensor<128x256xf32, #blocked>
+    // CHECK:   tt.store %arg2, %[[DOT]]
+    tt.store %arg2, %0 : tensor<128x256x!tt.ptr<f32>, #blocked>
+    tt.return
+  }
+}