fix(dnn/test): fix some bug when force_deduce_layout is off

GitOrigin-RevId: d7ccc397dfdbd9e2c599a24dd072242f69d19027

dnn/src/common/fake_quant.cpp

@@ -16,7 +16,7 @@
 namespace megdnn {
 
 void FakeQuantBase::deduce_layout_fwd(const TensorLayout& input, TensorLayout& output) {
-    output = TensorLayout(input, input.dtype);
+    output = TensorLayout(input);
 }
 
 void FakeQuantBase::check_layout_fwd(

dnn/src/common/lsq.cpp

@@ -16,7 +16,7 @@
 namespace megdnn {
 
 void LSQBase::deduce_layout_fwd(const TensorLayout& input, TensorLayout& output) {
-    output = TensorLayout(input, input.dtype);
+    output = TensorLayout(input);
 }
 
 void LSQBase::check_layout_fwd(

dnn/src/common/tqt.cpp

@@ -16,7 +16,7 @@
 namespace megdnn {
 
 void TQTBase::deduce_layout_fwd(const TensorLayout& input, TensorLayout& output) {
-    output = TensorLayout(input, input.dtype);
+    output = TensorLayout(input);
 }
 
 void TQTBase::check_layout_fwd(

dnn/test/common/checker.h

@@ -84,7 +84,7 @@ protected:
     TensorsConstriant m_tensor_constraint;
     bool m_no_naive_and_check = false;
     bool m_stable_check = false;
-    bool m_force_deduce_dst = true;
+    bool m_force_deduce_dst = false;
     bool m_allow_invalid_check = false;
     /**
      * the offset from the start of malloc memory

dnn/test/common/elemwise.cpp

@@ -756,6 +756,8 @@ DEF_TEST(all_modes) {
         for (size_t i = 0; i < shapes.size() - 1; ++i) {
             shapes[i] = {3, 9, 7};
         }
+        //! NOTE: force set output layout to  empty to trigger layout deduce
+        shapes[shapes.size() - 1] = {};
         auto do_run = [&](DType dtype, float eps = 1e-3) {
             // limit value ranges for some modes
             if (mode == Mode::LOG || mode == Mode::LOG1P) {

dnn/test/cuda/check_non_finite.cpp

@@ -22,16 +22,17 @@ TEST_F(CUDA, CHECK_NON_FINITE_BASIC) {
     const auto nan = std::numeric_limits<float>::quiet_NaN();
     UniformFloatWithValueRNG rng(-1.0f, 1.0f, 0.1f, inf);
     checker.set_rng(0, &rng);
-    checker.execs({{512 * 4}, {4}, {1}});
+    //! while deduce layout, dst tensor dtype will be set to Int32
+    checker.execs({{512 * 4}, {4}, {}});
     rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 1.f, inf);
     checker.set_rng(0, &rng);
-    checker.execs({{4}, {512 * 4}, {1}});
+    checker.execs({{4}, {512 * 4}, {}});
     rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 1.f, nan);
     checker.set_rng(0, &rng);
-    checker.execs({{32}, {256}, {1}});
+    checker.execs({{32}, {256}, {}});
     rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 0.f, nan);
     checker.set_rng(0, &rng);
-    checker.execs({{16}, {16}, {2}, {1}});
+    checker.execs({{16}, {16}, {2}, {}});
 }
 
 }  // namespace test

dnn/test/cuda/diag.cpp

@@ -26,10 +26,10 @@ TEST_F(CUDA, DIAG) {
             checker.set_dtype(1, dtype);
             size_t absk = static_cast<size_t>(std::abs(k));
             checker.exec(TensorShapeArray{{8}, {8 + absk, 8 + absk}});
-
+            //! NOTE: diag for vector or matrix is a vector
             auto oshape = [&](int n, int m) -> TensorShape {
-                size_t o = (k >= 0 ? std::min(n - k, m) : std::min(m + k, n));
-                return {o, o};
+                size_t o = (k >= 0 ? std::min(m - k, n) : std::min(n + k, m));
+                return {o};
             };
             checker.exec(TensorShapeArray{{8, 6}, oshape(8, 6)});
             checker.exec(TensorShapeArray{{6, 8}, oshape(6, 8)});