!4088 make gpu op Less to support int32

Merge pull request !4088 from TuDouNi/master

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/broadcast_impl.cu

@@ -256,6 +256,10 @@ template void Broadcast(const int &l0, const int &l1, const int &l2, const int &
                         const int &r2, const int &r3, const int &d0, const int &d1, const int &d2, const int &d3,
                         enum BroadcastOpType op, const int *input0, const int *input1, int *output,
                         cudaStream_t stream);
+template void Broadcast(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 BroadcastOpType op, const int *input0, const int *input1, bool *output,
+                        cudaStream_t stream);
 template void NoBroadcast(const int &nums, enum BroadcastOpType op, const float *input0, const float *input1,
                           bool *output, cudaStream_t stream);
 template void NoBroadcast(const int &nums, enum BroadcastOpType op, const float *input0, const float *input1,
@@ -266,6 +270,8 @@ template void NoBroadcast(const int &nums, enum BroadcastOpType op, const half *
                           half *output, cudaStream_t stream);
 template void NoBroadcast(const int &nums, enum BroadcastOpType op, const int *input0, const int *input1,
                           int *output, cudaStream_t stream);
+template void NoBroadcast(const int &nums, enum BroadcastOpType op, const int *input0, const int *input1,
+                          bool *output, cudaStream_t stream);
 template void BroadcastTo(const int &i0, const int &i1, const int &i2, const int &i3, const int &o0, const int &o1,
                           const int &o2, const int &o3, const float *input_addr, float *output_addr,
                           cudaStream_t stream);

mindspore/ccsrc/backend/kernel_compiler/gpu/math/broadcast_gpu_kernel.cc

@@ -103,6 +103,9 @@ MS_REG_GPU_KERNEL_TWO(
   BroadcastOpGpuKernel, half, half)
 
 // int32
+MS_REG_GPU_KERNEL_TWO(
+  Less, KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeBool),
+  BroadcastOpGpuKernel, int, bool)
 MS_REG_GPU_KERNEL_TWO(
   TensorAdd, KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt32).AddOutputAttr(kNumberTypeInt32),
   BroadcastOpGpuKernel, int, int)

tests/st/ops/gpu/test_broadcast_op.py

@@ -29,6 +29,8 @@ def test_nobroadcast():
 
     x1_np = np.random.rand(10, 20).astype(np.float32)
     x2_np = np.random.rand(10, 20).astype(np.float32)
+    x1_np_int32 = np.random.randint(0, 100, (10, 20)).astype(np.int32)
+    x2_np_int32 = np.random.randint(0, 100, (10, 20)).astype(np.int32)
 
     output_ms = P.Minimum()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.minimum(x1_np, x2_np)
@@ -45,6 +47,9 @@ def test_nobroadcast():
     output_ms = P.Less()(Tensor(x1_np), Tensor(x2_np))
     output_np = x1_np < x2_np
     assert np.allclose(output_ms.asnumpy(), output_np)
+    output_ms = P.Less()(Tensor(x1_np_int32), Tensor(x2_np_int32))
+    output_np = x1_np_int32 < x2_np_int32
+    assert np.allclose(output_ms.asnumpy(), output_np)
 
     output_ms = P.Pow()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.power(x1_np, x2_np)
@@ -71,6 +76,8 @@ def test_broadcast():
 
     x1_np = np.random.rand(3, 1, 5, 1).astype(np.float32)
     x2_np = np.random.rand(1, 4, 1, 6).astype(np.float32)
+    x1_np_int32 = np.random.randint(0, 100, (3, 1, 5, 1)).astype(np.int32)
+    x2_np_int32 = np.random.randint(0, 100, (3, 1, 5, 1)).astype(np.int32)
 
     output_ms = P.Minimum()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.minimum(x1_np, x2_np)
@@ -87,6 +94,9 @@ def test_broadcast():
     output_ms = P.Less()(Tensor(x1_np), Tensor(x2_np))
     output_np = x1_np < x2_np
     assert np.allclose(output_ms.asnumpy(), output_np)
+    output_ms = P.Less()(Tensor(x1_np_int32), Tensor(x2_np_int32))
+    output_np = x1_np_int32 < x2_np_int32
+    assert np.allclose(output_ms.asnumpy(), output_np)
 
     output_ms = P.Pow()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.power(x1_np, x2_np)
@@ -113,6 +123,8 @@ def test_broadcast_diff_dims():
 
     x1_np = np.random.rand(2).astype(np.float32)
     x2_np = np.random.rand(2, 1).astype(np.float32)
+    x1_np_int32 = np.random.randint(0, 100, (2)).astype(np.int32)
+    x2_np_int32 = np.random.randint(0, 100, (2, 1)).astype(np.int32)
 
     output_ms = P.Minimum()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.minimum(x1_np, x2_np)
@@ -129,6 +141,9 @@ def test_broadcast_diff_dims():
     output_ms = P.Less()(Tensor(x1_np), Tensor(x2_np))
     output_np = x1_np < x2_np
     assert np.allclose(output_ms.asnumpy(), output_np)
+    output_ms = P.Less()(Tensor(x1_np_int32), Tensor(x2_np_int32))
+    output_np = x1_np_int32 < x2_np_int32
+    assert np.allclose(output_ms.asnumpy(), output_np)
 
     output_ms = P.Pow()(Tensor(x1_np), Tensor(x2_np))
     output_np = np.power(x1_np, x2_np)