fix(dnn/sfotmax): call cpu dispatch for softmax opr

GitOrigin-RevId: a606e66101

dnn/src/fallback/softmax/opr_impl.cpp

@@ -6,35 +6,19 @@
 
 namespace megdnn {
 namespace fallback {
-void SoftmaxForwardImpl::exec(
-        _megdnn_tensor_in src, _megdnn_tensor_out dst, _megdnn_workspace workspace) {
-    auto axis = param().axis;
-    if (axis < 0)
-        axis += src.layout.ndim;
-    megdnn_assert(axis >= 0);
-    check_exec(src.layout, dst.layout, workspace.size);
-
-    if (!usable(src.layout)) {
-        naive::SoftmaxForwardImpl::exec(src, dst, workspace);
-        return;
-    }
-
-    typedef DTypeTrait<dtype::Float32>::ctype Float32;
-    auto sptr = src.ptr<Float32>();
-    auto dptr = dst.ptr<Float32>();
-
-    constexpr auto float_min = std::numeric_limits<Float32>::min();
-    constexpr auto step = GI_SIMD_LEN_BYTE / sizeof(Float32);
-
-    size_t A, B, C;
-    reduce::get_ABC(src.layout, A, B, C, axis);
 
+static void do_softmax(
+        const float* sptr, float* dptr, size_t A, size_t B, size_t C,
+        _megdnn_workspace workspace) {
+    constexpr auto float_min = std::numeric_limits<float>::min();
+    constexpr auto step = GI_SIMD_LEN_BYTE / sizeof(float);
     // TODO: When C=2,3,4..., src_ptr span is relatively large, the performance may
     // be poor
+
     if (C != 1) {
         WorkspaceBundle workspace_bundle{
-                workspace.raw_ptr, {A * C * sizeof(Float32), A * C * sizeof(Float32)}};
-        Float32* max = workspace_bundle.get_workspace(0).raw_ptr->as<Float32>();
+                workspace.raw_ptr, {A * C * sizeof(float), A * C * sizeof(float)}};
+        float* max = workspace_bundle.get_workspace(0).raw_ptr->as<float>();
         GI_FLOAT32_t v_max = GiBroadcastFloat32(float_min);
         size_t i = 0;
         for (; i + step <= A * C; i += step)
@@ -60,8 +44,8 @@ void SoftmaxForwardImpl::exec(
             }
         }
 
-        Float32* sum = workspace_bundle.get_workspace(1).raw_ptr->as<Float32>();
-        memset(sum, 0, A * C * sizeof(Float32));
+        float* sum = workspace_bundle.get_workspace(1).raw_ptr->as<float>();
+        memset(sum, 0, A * C * sizeof(float));
         for (size_t a = 0; a < A; a++) {
             for (size_t b = 0; b < B; b++) {
                 auto max_ptr = max + a * C;
@@ -157,6 +141,28 @@ void SoftmaxForwardImpl::exec(
     }
 }
 
+void SoftmaxForwardImpl::exec(
+        _megdnn_tensor_in src, _megdnn_tensor_out dst, _megdnn_workspace workspace) {
+    auto axis = param().axis;
+    if (axis < 0)
+        axis += src.layout.ndim;
+    megdnn_assert(axis >= 0);
+    check_exec(src.layout, dst.layout, workspace.size);
+
+    if (!usable(src.layout)) {
+        naive::SoftmaxForwardImpl::exec(src, dst, workspace);
+        return;
+    }
+
+    typedef DTypeTrait<dtype::Float32>::ctype Float32;
+    auto sptr = src.ptr<Float32>();
+    auto dptr = dst.ptr<Float32>();
+
+    size_t A, B, C;
+    reduce::get_ABC(src.layout, A, B, C, axis);
+    MEGDNN_DISPATCH_CPU_KERN_OPR(do_softmax(sptr, dptr, A, B, C, workspace));
+}
+
 }  // namespace fallback
 }  // namespace megdnn
 

imperative/python/test/unit/functional/test_functional.py

@@ -1653,3 +1653,18 @@ def test_conv_transpose3d():
     np.testing.assert_equal(
         output_shape.numpy(), np.array([20, 33, 32, 96, 197], dtype=np.int32)
     )
+
+
+@pytest.mark.skip(reason="pytest aborted")
+def test_softmax():
+    def np_softmax(x):
+        return np.exp(x) / np.sum(np.exp(x), axis=1, keepdims=True)
+
+    data = (np.random.random(size=(1, 16, 224, 224)).astype(np.float32) - 0.5) * 100
+    desired = np_softmax(data[:, :3, 0, 0])
+
+    data = Tensor(data)
+    data = data[:, :3, 0, 0]
+    actual = F.softmax(data)
+
+    np.testing.assert_allclose(actual.numpy(), desired, rtol=1e-5)