fix(mgb): fix profile skip condition

GitOrigin-RevId: f196eabc9810b5ac32744a10a1473a6b445f7f4a

imperative/src/impl/transformations/dtype_promote.cpp

@@ -186,6 +186,15 @@ ValueRefList convolution_rule(const OpDef& op, Span<ValueRef> inputs) {
 ValueRefList matmul_rule(const OpDef& op, Span<ValueRef> inputs) {
     auto&& conv_op = const_cast<MatrixMul&>(op.cast_final_safe<MatrixMul>());
     SmallVector<DType> dtypes = get_value_dtypes(inputs);
+
+    // skip dtype promotion when inputs are quantized
+    if (dtypes[0].category() == megdnn::DTypeCategory::QUANTIZED) {
+        mgb_assert(
+                dtypes[0].category() == dtypes[1].category(),
+                "inputs of matmul should have same quantized dtype.");
+        return imperative::apply(op, inputs);
+    }
+
     mgb::DType target_dtype;
 
     if (DTypePromoteCfg::amp_dtype_autocast_enabled) {
@@ -212,6 +221,15 @@ ValueRefList batch_matmul_rule(const OpDef& op, Span<ValueRef> inputs) {
     auto&& conv_op =
             const_cast<BatchedMatrixMul&>(op.cast_final_safe<BatchedMatrixMul>());
     SmallVector<DType> dtypes = get_value_dtypes(inputs);
+
+    // skip dtype promotion when inputs are quantized
+    if (dtypes[0].category() == megdnn::DTypeCategory::QUANTIZED) {
+        mgb_assert(
+                dtypes[0].category() == dtypes[1].category(),
+                "inputs of batched matmul should have same quantized dtype.");
+        return imperative::apply(op, inputs);
+    }
+
     mgb::DType target_dtype;
 
     if (DTypePromoteCfg::amp_dtype_autocast_enabled) {

src/rdnn/impl/algo_chooser.cpp

@@ -600,7 +600,9 @@ typename AlgoChooser<Opr>::ImplExecutionPolicy AlgoChooser<Opr>::AlgoChooserHelp
             auto&& megdnn_opr = opr::intl::create_megdnn_opr<_Opr>(m_cn);
             // skip different sub opr, for example:
             // skip matmul algo when profiling convolution
-            if (m_dnn_opr->get_opr_type() != megdnn_opr->get_opr_type())
+            if ((m_cn.device_type() == mgb::CompNode::DeviceType::CUDA ||
+                 m_cn.device_type() == mgb::CompNode::DeviceType::ROCM) &&
+                m_dnn_opr->get_opr_type() != megdnn_opr->get_opr_type())
                 continue;
             megdnn_opr->param() =
                     Algorithm::deserialize_read_pod<typename _Opr::Param>(_item.param);