test(dnn/bn): add compatible configs for NHWC BN

GitOrigin-RevId: ac757ca307

dnn/test/common/bn.h

@@ -53,6 +53,18 @@ std::vector<TestArg> get_args() {
                           TensorShape{1, 3, 1, 1}, dtype::Float16());
     }
 
+    // case 3: 1 x 1 x 1 x C
+
+    for (size_t i = 4; i < 257; i *= 4) {
+        param::BN param;
+        param.fwd_mode = param::BN::FwdMode::TRAINING;
+        param.param_dim = param::BN::ParamDim::DIM_111C;
+        args.emplace_back(param, TensorShape{3, i, i, 3},
+                          TensorShape{1, 1, 1, 3}, dtype::Float32());
+        args.emplace_back(param, TensorShape{3, i, i, 3},
+                          TensorShape{1, 1, 1, 3}, dtype::Float16());
+    }
+
     return args;
 }
 
@@ -60,4 +72,4 @@ std::vector<TestArg> get_args() {
 }  // namespace test
 }  // namespace megdnn
 
-// vim: syntax=cpp.doxygen
+// vim: syntax=cpp.doxygen

dnn/test/common/checker.cpp

@@ -419,7 +419,7 @@ void CheckerHelper::copy_tensors_from_device(const TensorValueArray& dest,
 void CheckerHelper::check_tensors(const TensorValueArray& expected,
                                   const TensorValueArray& computed) {
     for (size_t i = 0; i < expected.size(); ++i) {
-        if (expected[i].layout.ndim == 0)
+        if (expected[i].layout.ndim == 0 || m_bypass.find(i) != m_bypass.end())
             continue;
         if (m_allow_invalid_check) {
             MEGDNN_ASSERT_TENSOR_EQ_EPS_AVG_ALLOW_INVALID(

dnn/test/common/checker.h

@@ -69,6 +69,7 @@ protected:
     std::unordered_map<size_t, RNG*> m_rng;
     std::unordered_map<size_t, DType> m_dtype;
     std::unordered_map<size_t, TensorFormat> m_fmt;
+    std::set<size_t> m_bypass;
     float_t m_epsilon = 1e-3, m_max_avg_error = 1e-3,
             m_max_avg_biased_error = 1e-3;
     float_t m_perf_check_threshold = -1;
@@ -184,6 +185,10 @@ public:
         m_rng[idx] = rng;
         return *this;
     }
+    Checker& set_bypass(size_t idx) {
+        m_bypass.insert(idx);
+        return *this;
+    }
     //! max error of a single element
     Checker& set_epsilon(dt_float32 epsilon) {
         m_epsilon = epsilon;

dnn/test/common/deduce_layout_proxy.h

@@ -82,6 +82,15 @@ struct DeduceLayoutProxy<Opr, 8, true> {
     }
 };
 
+template <typename Opr>
+struct DeduceLayoutProxy<Opr, 9, true> {
+    static void deduce_layout(Opr* opr, TensorLayoutArray& layouts) {
+        megdnn_assert(layouts.size() == 9);
+        opr->deduce_layout(layouts[0], layouts[1], layouts[2], layouts[3],
+                           layouts[4], layouts[5], layouts[6], layouts[7],
+                           layouts[8]);
+    }
+};
 }  // namespace test
 }  // namespace megdnn
 

dnn/test/common/exec_proxy.h

@@ -22,6 +22,23 @@ namespace test {
 template <typename Opr, size_t Arity, bool has_workspace>
 struct ExecProxy;
 
+template <typename Opr>
+struct ExecProxy<Opr, 9, true> {
+    WorkspaceWrapper W;
+    void exec(Opr* opr, const TensorNDArray& tensors) {
+        if (!W.valid()) {
+            W = WorkspaceWrapper(opr->handle(), 0);
+        }
+        W.update(opr->get_workspace_in_bytes(
+                tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                tensors[3].layout, tensors[4].layout, tensors[5].layout,
+                tensors[6].layout, tensors[7].layout, tensors[8].layout));
+        opr->exec(tensors[0], tensors[1], tensors[2], tensors[3], tensors[4],
+                  tensors[5], tensors[6], tensors[7], tensors[8],
+                  W.workspace());
+    }
+};
+
 template <typename Opr>
 struct ExecProxy<Opr, 8, true> {
     WorkspaceWrapper W;

dnn/test/common/rng.cpp

@@ -211,6 +211,10 @@ void IIDRNG::gen(const TensorND& tensor) {
         }
         return;
     }
+    if (tensor.layout.dtype.enumv() == DTypeEnum::Byte) {
+        memset(tensor.raw_ptr, 0, tensor.layout.access_bytes());
+        return;
+    }
     megdnn_assert(0, "IIDRNG does not know how to generate value for DType %s",
                   tensor.layout.dtype.name());
 }

dnn/test/cuda/bn.cpp

@@ -6,10 +6,13 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 #include "test/cuda/fixture.h"
 
+#include "src/cuda/batch_normalization/opr_impl.h"
+#include "src/cuda/utils.h"
 #include "megdnn/opr_param_defs.h"
 #include "megdnn/oprs.h"
 #include "test/common/bn.h"
@@ -21,15 +24,26 @@
 namespace megdnn {
 namespace test {
 
-TEST_F(CUDA, BN_FORWARD) {
+TEST_F(CUDA, BN_FORWARD_BACKWARD) {
     using namespace batch_normalization;
+    using cuda::cudnn_handle;
+    using cuda::batch_normalization::BNTensorDescHolder;
+    using cuda::batch_normalization::get_reserve_size;
     std::vector<TestArg> args = get_args();
     Checker<BNForward> checker(handle_cuda());
+    Checker<BNBackward> checker_bwd(handle_cuda());
     for (auto&& arg : args) {
-        for (int i = 0; i < 8; ++i) {
+        auto tensor_desc = BNTensorDescHolder({arg.src, arg.dtype}, arg.param.param_dim,
+                                         arg.param.fwd_mode);
+        auto reserve = get_reserve_size(cudnn_handle(handle_cuda()), tensor_desc);
+        // Forward
+        for (int i = 0; i < 9; ++i) {
             checker.set_dtype(i, dtype::Float32());
         }
         checker.set_dtype(0, arg.dtype);
+        checker.set_dtype(7, dtype::Byte());
+        checker.set_dtype(8, arg.dtype);
+        checker.set_bypass(7);
         checker.set_epsilon(1e-3).set_param(arg.param);
         for (bool need_statistic : {false, true})
             checker.exec({
@@ -40,27 +54,26 @@ TEST_F(CUDA, BN_FORWARD) {
                                    : TensorShape({0}),  // mean
                     need_statistic ? arg.param_shape
                                    : TensorShape({0}),  // variance
-                    arg.param_shape,                 // batch_mean
-                    arg.param_shape,                 // batch_inv_variance
-                    {}                               // dst
+                    arg.param_shape,                    // batch_mean
+                    arg.param_shape,                    // batch_inv_variance
+                    {reserve},                            // reserve
+                    arg.src                             // dst
             });
-    }
-}
 
-TEST_F(CUDA, BN_BACKWARD) {
-    using namespace batch_normalization;
-    std::vector<TestArg> args = get_args();
-    Checker<BNBackward> checker(handle_cuda());
-    for (auto&& arg : args) {
-        for (int i = 0; i < 8; ++i) {
-            checker.set_dtype(i, dtype::Float32());
+        // Backward
+        for (int i = 0; i < 9; ++i) {
+            checker_bwd.set_dtype(i, dtype::Float32());
         }
-        checker.set_dtype(0, arg.dtype)    // x
-                .set_dtype(1, arg.dtype)   // dy
-                .set_dtype(7, arg.dtype);  // dx
-        checker.set_epsilon(1e-3).set_param(arg.param).exec(
+        checker_bwd
+                .set_dtype(0, arg.dtype)      // x
+                .set_dtype(1, arg.dtype)      // dy
+                .set_dtype(5, dtype::Byte())  // reserve
+                .set_dtype(8, arg.dtype)      // dx
+                .set_bypass(5);
+        checker_bwd.set_epsilon(1e-3).set_param(arg.param).exec(
                 {arg.src, arg.src, arg.param_shape, arg.param_shape,
-                 arg.param_shape, arg.param_shape, arg.param_shape, arg.src});
+                 arg.param_shape, {reserve}, arg.param_shape, arg.param_shape,
+                 arg.src});
     }
 }
 

dnn/test/rocm/bn.cpp

@@ -31,6 +31,7 @@ TEST_F(ROCM, BN_FORWARD) {
             checker.set_dtype(i, dtype::Float32());
         }
         checker.set_dtype(0, arg.dtype);
+        checker.set_dtype(8, arg.dtype);
         checker.set_epsilon(1e-3).set_param(arg.param);
         for (bool need_statistic : {false, true})
             checker.exec({
@@ -43,7 +44,8 @@ TEST_F(ROCM, BN_FORWARD) {
                                    : TensorShape({0}),  // variance
                     arg.param_shape,                    // batch_mean
                     arg.param_shape,                    // batch_inv_variance
-                    {}                                  // dst
+                    {0},                                // reserve
+                    arg.src                             // dst
             });
     }
 }
@@ -53,15 +55,16 @@ TEST_F(ROCM, BN_BACKWARD) {
     std::vector<TestArg> args = get_args();
     Checker<BNBackward> checker(handle_rocm());
     for (auto&& arg : args) {
-        for (int i = 0; i < 8; ++i) {
+        for (int i = 0; i < 9; ++i) {
             checker.set_dtype(i, dtype::Float32());
         }
         checker.set_dtype(0, arg.dtype)    // x
                 .set_dtype(1, arg.dtype)   // dy
-                .set_dtype(7, arg.dtype);  // dx
+                .set_dtype(8, arg.dtype);  // dx
         checker.set_epsilon(1e-3).set_param(arg.param).exec(
                 {arg.src, arg.src, arg.param_shape, arg.param_shape,
-                 arg.param_shape, arg.param_shape, arg.param_shape, arg.src});
+                 arg.param_shape, {0}, arg.param_shape, arg.param_shape,
+                 arg.src});
     }
 }