!8146 Improve performance for GPU-ScatterUpdate, add int32 support

Merge pull request !8146 from 34bunny/GPU-ScatterUpdateFix

mindspore/ccsrc/backend/kernel_compiler/gpu/arrays/scatter_update_gpu_kernel.cc

@@ -32,5 +32,12 @@ MS_REG_GPU_KERNEL_ONE(ScatterUpdate,
                         .AddInputAttr(kNumberTypeFloat16)
                         .AddOutputAttr(kNumberTypeFloat16),
                       ScatterUpdateKernel, half)
+MS_REG_GPU_KERNEL_ONE(ScatterUpdate,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddInputAttr(kNumberTypeInt32)
+                        .AddOutputAttr(kNumberTypeInt32),
+                      ScatterUpdateKernel, int)
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/gpu/arrays/scatter_update_gpu_kernel.h

@@ -40,8 +40,10 @@ class ScatterUpdateKernel : public GpuKernel {
     int *indices = GetDeviceAddress<int>(inputs, 1);
     T *updates = GetDeviceAddress<T>(inputs, 2);
     T *output = GetDeviceAddress<T>(outputs, 0);
-    CalScatterUpdate(input_size_, inner_size_, indices_size_, input, indices, updates, output,
-                     reinterpret_cast<cudaStream_t>(stream_ptr));
+    CHECK_CUDA_RET_WITH_EXCEPT(cudaMemcpyAsync(&output[0], &input[0], input_size_ * sizeof(T), cudaMemcpyDeviceToDevice,
+                                               reinterpret_cast<cudaStream_t>(stream_ptr)),
+                               "cudaMemcpyAsync output failed");
+    CalScatterUpdate(inner_size_, indices_size_, indices, updates, output, reinterpret_cast<cudaStream_t>(stream_ptr));
     return true;
   }
 

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

@@ -17,29 +17,27 @@
 #include "backend/kernel_compiler/gpu/cuda_impl/scatter_update_impl.cuh"
 
 template <typename T>
-__global__ void ScatterUpdate(const int input_size, const int inner_size, const int indices_size, const T *input,
-                              const int *indices, const T *updates, T *output) {
-  for (int pos = blockIdx.x * blockDim.x + threadIdx.x; pos < input_size; pos += blockDim.x * gridDim.x) {
-    output[pos] = input[pos];
+__global__ void ScatterUpdate(const int inner_size, const int updates_size, const int *indices, const T *updates,
+                              T *output) {
+  for (int pos = blockIdx.x * blockDim.x + threadIdx.x; pos < updates_size; pos += blockDim.x * gridDim.x) {
     const int index = pos / inner_size;
     const int offset = pos % inner_size;
-    for (int i = 0; i < indices_size; i++) {
-      const int update_pos = i * inner_size + offset;
-      output[pos] = (indices[i] == index ? updates[update_pos] : output[pos]);
-    }
+    const int current_pos = indices[index] * inner_size + offset;
+    output[current_pos] = updates[pos];
   }
 }
 
 template <typename T>
-void CalScatterUpdate(const int &input_size, const int &inner_size, const int &indices_size, const T *input,
-                      const int *indices, const T *updates, T *output, cudaStream_t cuda_stream) {
-  ScatterUpdate<<<GET_BLOCKS(input_size), GET_THREADS, 0, cuda_stream>>>(input_size, inner_size, indices_size, input,
-                                                                         indices, updates, output);
+void CalScatterUpdate(const int &inner_size, const int &indices_size, const int *indices, const T *updates, T *output,
+                      cudaStream_t cuda_stream) {
+  const int updates_size = inner_size * indices_size;
+  ScatterUpdate<<<GET_BLOCKS(updates_size), GET_THREADS, 0, cuda_stream>>>(inner_size, updates_size, indices, updates,
+                                                                           output);
 }
 
-template void CalScatterUpdate<float>(const int &input_size, const int &inner_size, const int &indices_size,
-                                      const float *input, const int *indices, const float *updates, float *output,
-                                      cudaStream_t cuda_stream);
-template void CalScatterUpdate<half>(const int &input_size, const int &inner_size, const int &indices_size,
-                                     const half *input, const int *indices, const half *updates, half *output,
-                                     cudaStream_t cuda_stream);
+template void CalScatterUpdate<float>(const int &inner_size, const int &indices_size, const int *indices,
+                                      const float *updates, float *output, cudaStream_t cuda_stream);
+template void CalScatterUpdate<half>(const int &inner_size, const int &indices_size, const int *indices,
+                                     const half *updates, half *output, cudaStream_t cuda_stream);
+template void CalScatterUpdate<int>(const int &inner_size, const int &indices_size, const int *indices,
+                                    const int *updates, int *output, cudaStream_t cuda_stream);

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/scatter_update_impl.cuh

@@ -20,7 +20,7 @@
 #include "runtime/device/gpu/cuda_common.h"
 
 template <typename T>
-void CalScatterUpdate(const int &input_size, const int &inner_size, const int &indices_size, const T *input,
-                      const int *indices, const T *updates, T *output, cudaStream_t cuda_stream);
+void CalScatterUpdate(const int &inner_size, const int &indices_size, const int *indices, const T *updates, T *output,
+                      cudaStream_t cuda_stream);
 
 #endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_CUDA_IMPL_SCATTER_UPDATE_IMPL_CUH_

tests/st/ops/gpu/test_scatter_update_op.py

@@ -75,7 +75,19 @@ def test_scatter_update_float16():
     updates = Tensor(np.arange(6).reshape((2, 3)).astype(np.float16))
     output = scatter_update_net(inputx, indices, updates)
     expected = np.array([[0., 1., 2.],
-                         [3., 4., 5.]])
+                         [3., 4., 5.]]).astype(np.float16)
+    np.testing.assert_array_almost_equal(output.asnumpy(), expected)
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_scatter_update_int32():
+    inputx = Tensor(np.zeros((2, 3)).astype(np.int32))
+    indices = Tensor(np.array([0, 1]).astype(np.int32))
+    updates = Tensor(np.arange(6).reshape((2, 3)).astype(np.int32))
+    output = scatter_update_net(inputx, indices, updates)
+    expected = np.array([[0., 1., 2.],
+                         [3., 4., 5.]]).astype(np.int32)
     np.testing.assert_array_almost_equal(output.asnumpy(), expected)
 
 @pytest.mark.level0
@@ -89,7 +101,7 @@ def test_scatter_update_large_float16():
     expected = np.array([[69., 70., 71.],
                          [66., 67., 68.],
                          [63., 64., 65.],
-                         [72., 73., 74.]])
+                         [72., 73., 74.]]).astype(np.float16)
     np.testing.assert_array_almost_equal(output.asnumpy(), expected)
 
 @pytest.mark.level0
@@ -102,5 +114,52 @@ def test_scatter_update_disordered_float16():
     output = scatter_update_net(inputx, indices, updates)
     expected = np.array([[45., 44., 43., 42.],
                          [63., 64., 65., 66.],
-                         [67., 68., 69., 70.]])
+                         [67., 68., 69., 70.]]).astype(np.float16)
+    np.testing.assert_array_almost_equal(output.asnumpy(), expected)
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_scatter_update_disordered_int32():
+    inputx = Tensor(np.flip(np.arange(34, 46).reshape(3, 4).astype(np.int32)))
+    indices = Tensor(np.array([1, 2]).astype(np.int32))
+    updates = Tensor(np.arange(63, 71).reshape((2, 4)).astype(np.int32))
+    output = scatter_update_net(inputx, indices, updates)
+    expected = np.array([[45., 44., 43., 42.],
+                         [63., 64., 65., 66.],
+                         [67., 68., 69., 70.]]).astype(np.int32)
+    np.testing.assert_array_almost_equal(output.asnumpy(), expected)
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_scatter_update_large_shape_float16():
+    inputx = Tensor(np.arange(96).reshape((4, 2, 3, 4)).astype(np.float16))
+    indices = Tensor(np.array([1, 0]).astype(np.int32))
+    updates = Tensor(np.flip(np.arange(48).reshape((2, 2, 3, 4)).astype(np.float16)))
+    output = scatter_update_net(inputx, indices, updates)
+    expected = np.array([[[[23., 22., 21., 20.],
+                           [19., 18., 17., 16.],
+                           [15., 14., 13., 12.]],
+                          [[11., 10., 9., 8.],
+                           [7., 6., 5., 4.],
+                           [3., 2., 1., 0.]]],
+                         [[[47., 46., 45., 44.],
+                           [43., 42., 41., 40.],
+                           [39., 38., 37., 36.]],
+                          [[35., 34., 33., 32.],
+                           [31., 30., 29., 28.],
+                           [27., 26., 25., 24.]]],
+                         [[[48., 49., 50., 51.],
+                           [52., 53., 54., 55.],
+                           [56., 57., 58., 59.]],
+                          [[60., 61., 62., 63.],
+                           [64., 65., 66., 67.],
+                           [68., 69., 70., 71.]]],
+                         [[[72., 73., 74., 75.],
+                           [76., 77., 78., 79.],
+                           [80., 81., 82., 83.]],
+                          [[84., 85., 86., 87.],
+                           [88., 89., 90., 91.],
+                           [92., 93., 94., 95.]]]]).astype(np.float16)
     np.testing.assert_array_almost_equal(output.asnumpy(), expected)