!2338 Gpu Minimum & Maximum kernels support int32

Merge pull request !2338 from chenweifeng/nezha

mindspore/ccsrc/kernel/gpu/cuda_impl/broadcast_grad_impl.cu

@@ -110,7 +110,13 @@ void NoBroadcastGrad(const int &nums, enum BroadcastGradOpType op, const T *x1,
 
 template void NoBroadcastGrad(const int &nums, enum BroadcastGradOpType op, const float *x1, const float *x2,
                               const float *dy, float *dx1, float *dx2, cudaStream_t stream);
+template void NoBroadcastGrad(const int &nums, enum BroadcastGradOpType op, const int *x1, const int *x2,
+                              const int *dy, int *dx1, int *dx2, cudaStream_t stream);
 template void BroadcastGrad(const int &l0, const int &l1, const int &l2, const int &l3, const int &r0, const int &r1,
                             const int &r2, const int &r3, const int &d0, const int &d1, const int &d2, const int &d3,
                             enum BroadcastGradOpType op, const float *x1, const float *x2, const float *dy, float *dx1,
                             float *dx2, cudaStream_t stream);
+template void BroadcastGrad(const int &l0, const int &l1, const int &l2, const int &l3, const int &r0, const int &r1,
+                            const int &r2, const int &r3, const int &d0, const int &d1, const int &d2, const int &d3,
+                            enum BroadcastGradOpType op, const int *x1, const int *x2, const int *dy, int *dx1,
+                            int *dx2, cudaStream_t stream);

mindspore/ccsrc/kernel/gpu/math/broadcast_gpu_kernel.cc

@@ -90,5 +90,11 @@ MS_REG_GPU_KERNEL_TWO(
 MS_REG_GPU_KERNEL_TWO(
   TensorAdd, KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
   BroadcastOpGpuKernel, int, int)
+MS_REG_GPU_KERNEL_TWO(
+  Minimum, KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
+  BroadcastOpGpuKernel, int, int)
+MS_REG_GPU_KERNEL_TWO(
+  Maximum, KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
+  BroadcastOpGpuKernel, int, int)
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/gpu/math/broadcast_grad_gpu_kernel.cc

@@ -34,5 +34,21 @@ MS_REG_GPU_KERNEL_ONE(MaximumGrad,
                         .AddOutputAttr(kNumberTypeFloat32)
                         .AddOutputAttr(kNumberTypeFloat32),
                       BroadcastOpGradGpuKernel, float)
+MS_REG_GPU_KERNEL_ONE(MinimumGrad,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddOutputAttr(kNumberTypeInt32)
+                        .AddOutputAttr(kNumberTypeInt32),
+                      BroadcastOpGradGpuKernel, int)
+MS_REG_GPU_KERNEL_ONE(MaximumGrad,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddOutputAttr(kNumberTypeInt32)
+                        .AddOutputAttr(kNumberTypeInt32),
+                      BroadcastOpGradGpuKernel, int)
 }  // namespace kernel
 }  // namespace mindspore

tests/st/ops/gpu/test_maximum_op.py

@@ -222,3 +222,27 @@ def test_broadcast_diff_dims():
     output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np))
     assert np.allclose(output_ms[0].asnumpy(), expect_dx1)
     assert np.allclose(output_ms[1].asnumpy(), expect_dx2)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_maximum_int():
+    x = Tensor(np.array([[1, 2, 3]]).astype(np.int32))
+    y = Tensor(np.array([[2]]).astype(np.int32))
+    expect = [[2, 2, 3]]
+    error = np.ones(shape=[1, 3]) * 1.0e-5
+
+    context.set_context(mode=context.PYNATIVE_MODE, device_target="GPU")
+    max_op = Net()
+    output = max_op(x, y)
+    diff = output.asnumpy() - expect
+    assert np.all(diff < error)
+    assert np.all(-diff < error)
+
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+    max_op_2 = Net()
+    output = max_op_2(x, y)
+    diff = output.asnumpy() - expect
+    assert np.all(diff < error)
+    assert np.all(-diff < error)

tests/st/ops/gpu/test_minimum_op.py

@@ -218,3 +218,21 @@ def test_broadcast_diff_dims():
     output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np))
     assert np.allclose(output_ms[0].asnumpy(), expect_dx1)
     assert np.allclose(output_ms[1].asnumpy(), expect_dx2)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_broadcast_int32():
+    context.set_context(mode=context.GRAPH_MODE, save_graphs=True, device_target='GPU')
+
+    x1_np = np.random.rand(3, 4).astype(np.int32)
+    x2_np = np.random.rand(3, 4).astype(np.int32)
+    dy_np = np.random.rand(3, 4).astype(np.int32)
+
+    net = Grad(MinimumNet())
+    output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np))
+    output0_np = np.where(x1_np < x2_np, dy_np, 0)
+    output1_np = np.where(x1_np < x2_np, 0, dy_np)
+    assert np.allclose(output_ms[0].asnumpy(), output0_np)
+    assert np.allclose(output_ms[1].asnumpy(), output1_np)