Improve argsort performance. (#21267)

* Improve argsort performance.

- Give 200000 data to compute argsort on v100,
can speed up ~190x
before opt cost: 0.53s
after opt cost:0.0027s

- Add fp16 support

* Refine error message
* Refine code

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

paddle/fluid/operators/argsort_op.cu

@@ -14,82 +14,133 @@ limitations under the License. */
 
 #include <thrust/execution_policy.h>
 #include <thrust/sort.h>
+#include "cub/cub.cuh"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/argsort_op.h"
+#include "paddle/fluid/operators/transpose_op.h"
 #include "paddle/fluid/platform/cuda_device_function.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
 
+// set cub base traits in order to handle float16
+namespace cub {
+template <>
+struct NumericTraits<paddle::platform::float16>
+    : BaseTraits<FLOATING_POINT, true, false, uint16_t,
+                 paddle::platform::float16> {};
+}  // namespace cub
+
 namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
-using platform::PADDLE_CUDA_NUM_THREADS;
-
-const int kMaxRank = 9;  // The max rank of a tensor allowed in Fluid
-
-__global__ void ComputeTargetIdx(const int64_t* in_dims, int dims_size,
-                                 int axis, int64_t n, int64_t* trg_idx,
-                                 int64_t* med_ids) {
-  int64_t index = threadIdx.x + blockDim.x * blockIdx.x;
-  if (index < n) {
-    int64_t shape_out_axis[kMaxRank - 1] = {0};
-    int64_t dims_out_axis[kMaxRank - 1] = {0};
-    int64_t tmp = index;
-    int64_t pos_in_axis = 0;
-    int64_t i = dims_size - 2;
-    int64_t dim_axis = 0;
-    for (int64_t j = dims_size - 1; j >= 0; --j) {
-      int64_t dim = in_dims[j];
-      if (j != axis) {
-        shape_out_axis[i] = tmp % dim;
-        dims_out_axis[i] = dim;
-        i--;
-      } else {
-        dim_axis = dim;
-        pos_in_axis = tmp % dim_axis;
-      }
-      tmp /= dim;
-    }
-    int64_t group = (dims_size > 1) ? shape_out_axis[0] : 0;
-    for (int64_t j = 0; j < dims_size - 2; ++j) {
-      group = group * dims_out_axis[j + 1] + shape_out_axis[j + 1];
-    }
 
-    int64_t traget_idx = group * dim_axis + pos_in_axis;
-    trg_idx[index] = traget_idx;
-    med_ids[traget_idx] = pos_in_axis;
-  }
-}
+// Iter for move to next row
+struct SegmentOffsetIter {
+  EIGEN_DEVICE_FUNC
+  explicit SegmentOffsetIter(int num_cols) : num_cols_(num_cols) {}
 
-template <typename T>
-__global__ void PermuteInData(const T* in, const int64_t* trg_idx, int64_t n,
-                              T* med_out) {
-  int index = threadIdx.x + blockDim.x * blockIdx.x;
-  if (index < n) {
-    med_out[trg_idx[index]] = in[index];
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator()(int idx) const {
+    return idx * num_cols_;
   }
-}
+
+  int num_cols_;
+};
 
 template <typename T>
-__global__ void Sort(int64_t axis_dim, int64_t groups, T* med_out,
-                     int64_t* med_ids) {
-  int index = threadIdx.x + blockDim.x * blockIdx.x;
-  if (index < groups) {
-    thrust::sort_by_key(thrust::device, med_out + index * axis_dim,
-                        med_out + axis_dim * (1 + index),
-                        med_ids + index * axis_dim);
+static __global__ void FillIndex(T* indices, T num_rows, T num_cols) {
+  int col_id = threadIdx.x;
+  int row_id = blockIdx.x;
+
+  for (T j = row_id; j < num_rows; j += gridDim.x) {
+    for (T i = col_id; i < num_cols; i += blockDim.x) {
+      indices[j * num_cols + i] = i;
+    }
   }
 }
 
-template <typename T>
-__global__ void PermuteMediateData(const T* med_out, const int64_t* med_ids,
-                                   const int64_t* trg_idx, int64_t n, T* out,
-                                   int64_t* indices) {
-  int index = threadIdx.x + blockDim.x * blockIdx.x;
-  if (index < n) {
-    out[index] = med_out[trg_idx[index]];
-    indices[index] = med_ids[trg_idx[index]];
-  }
+// Default use ascending sort
+template <typename T, typename IndType>
+void ArgFullSortAscending(const platform::CUDADeviceContext& ctx,
+                          const Tensor* input, Tensor* output, Tensor* indices,
+                          const IndType num_rows, const IndType num_cols) {
+  auto cu_stream = ctx.stream();
+
+  Tensor input_indices;
+
+  const std::vector<IndType> dims = {num_rows, num_cols};
+  auto dim = framework::make_ddim(dims);
+  input_indices.Resize(dim);
+  input_indices.mutable_data<IndType>(ctx.GetPlace());
+
+  size_t temp_storage_bytes = -1;
+
+  auto ComputeBlockSize = [](IndType col) {
+    if (col > 512)
+      return 1024;
+    else if (col > 256 && col <= 512)
+      return 512;
+    else if (col > 128 && col <= 256)
+      return 256;
+    else if (col > 64 && col <= 128)
+      return 128;
+    else
+      return 64;
+  };
+
+  int block_size = ComputeBlockSize(num_cols);
+
+  int maxGridDimX = ctx.GetCUDAMaxGridDimSize().x;
+  // actually, int num_rows < max_grid_size
+  int grid_size = num_rows < maxGridDimX ? num_rows : maxGridDimX;
+  // Init a index array
+  FillIndex<<<grid_size, block_size, 0, cu_stream>>>(
+      input_indices.data<IndType>(), num_rows, num_cols);
+
+  T* sorted_out_ptr;
+  IndType* sorted_indices_ptr;
+
+  const T* inp = input->data<T>();
+  T* out = output->mutable_data<T>(ctx.GetPlace());
+  IndType* ind = indices->mutable_data<IndType>(ctx.GetPlace());
+
+  sorted_out_ptr = out;
+  sorted_indices_ptr = ind;
+
+  // create iter for counting input
+  cub::CountingInputIterator<IndType> counting_iter(0);
+  // segment_offset is used for move to next row
+  cub::TransformInputIterator<IndType, SegmentOffsetIter,
+                              cub::CountingInputIterator<IndType>>
+      segment_offsets_t(counting_iter, SegmentOffsetIter(num_cols));
+
+  auto err = cub::DeviceSegmentedRadixSort::SortPairs(
+      nullptr, temp_storage_bytes, inp, sorted_out_ptr,
+      input_indices.data<IndType>(), sorted_indices_ptr, num_cols * num_rows,
+      num_rows, segment_offsets_t, segment_offsets_t + 1, 0, sizeof(T) * 8,
+      cu_stream);
+
+  PADDLE_ENFORCE_CUDA_SUCCESS(
+      err,
+      "ArgSortOP failed as could not launch "
+      "cub::DeviceSegmentedRadixSort::SortPairsDescending to calculate"
+      "temp_storage_bytes, status:%s.",
+      temp_storage_bytes, cudaGetErrorString(err));
+
+  Tensor temp_storage;
+  temp_storage.mutable_data<uint8_t>(ctx.GetPlace(), temp_storage_bytes);
+
+  err = cub::DeviceSegmentedRadixSort::SortPairs(
+      temp_storage.data<uint8_t>(), temp_storage_bytes, inp, sorted_out_ptr,
+      input_indices.data<IndType>(), sorted_indices_ptr, num_cols * num_rows,
+      num_rows, segment_offsets_t, segment_offsets_t + 1, 0, sizeof(T) * 8,
+      cu_stream);
+
+  PADDLE_ENFORCE_CUDA_SUCCESS(
+      err,
+      "ArgSortOP failed as could not launch "
+      "cub::DeviceSegmentedRadixSort::SortPairsDescending to sort input, "
+      "temp_storage_bytes:%d status:%s.",
+      temp_storage_bytes, cudaGetErrorString(err));
 }
 
 template <typename T>
@@ -104,47 +155,91 @@ class ArgsortOpCUDAKernel : public framework::OpKernel<T> {
     auto in_dims = input->dims();
     axis = (axis < 0) ? (in_dims.size() + axis) : axis;
 
-    const T* in_data = input->data<T>();
-    T* out_data = output->mutable_data<T>(ctx.GetPlace());
-    int64_t* ids_data = indices->mutable_data<int64_t>(ctx.GetPlace());
-
     int64_t numel = input->numel();
     int64_t groups = numel / in_dims[axis];
 
-    std::vector<int64_t> in_dims_vec = vectorize(in_dims);
-    thrust::device_vector<int64_t> in_dims_dev(in_dims_vec.begin(),
-                                               in_dims_vec.end());
-    int64_t* in_dims_data = thrust::raw_pointer_cast(in_dims_dev.data());
-    // Mediate tensor for sorting data and indices
-    Tensor mediate_output, mediate_indices;
-    T* med_out_data =
-        mediate_output.mutable_data<T>(input->dims(), ctx.GetPlace());
-    int64_t* med_ids_data =
-        mediate_indices.mutable_data<int64_t>(in_dims, ctx.GetPlace());
-    // Target index of each element along the given axis in the mediate tensors
-    Tensor trg_idx_t;
-    int64_t* trg_idx = trg_idx_t.mutable_data<int64_t>(in_dims, ctx.GetPlace());
-
-    auto stream = ctx.cuda_device_context().stream();
-    const int num_threads = PADDLE_CUDA_NUM_THREADS;
-
-    ComputeTargetIdx<<<(numel - 1) / num_threads + 1, num_threads, 0, stream>>>(
-        in_dims_data, in_dims.size(), axis, numel, trg_idx, med_ids_data);
-
-    PermuteInData<<<(numel - 1) / num_threads + 1, num_threads, 0, stream>>>(
-        in_data, trg_idx, numel, med_out_data);
-
-    Sort<<<(groups - 1) / num_threads + 1, num_threads, 0, stream>>>(
-        in_dims[axis], groups, med_out_data, med_ids_data);
-
-    PermuteMediateData<<<(numel - 1) / num_threads + 1, num_threads, 0,
-                         stream>>>(med_out_data, med_ids_data, trg_idx, numel,
-                                   out_data, ids_data);
+    // Special case for full sort, speedup ~190x.
+    if (axis == -1 || axis + 1 == in_dims.size()) {
+      const int64_t input_height = framework::product(
+          framework::slice_ddim(in_dims, 0, in_dims.size() - 1));
+      const int64_t input_width = in_dims[in_dims.size() - 1];
+      const auto& dev_ctx = ctx.cuda_device_context();
+      if (input_width < INT_MAX && input_height < INT_MAX) {
+        ArgFullSortAscending<T, int>(dev_ctx, input, output, indices,
+                                     static_cast<int>(input_height),
+                                     static_cast<int>(input_width));
+      } else {
+        ArgFullSortAscending<T, int64_t>(dev_ctx, input, output, indices,
+                                         input_height, input_width);
+      }
+    } else {
+      // if not full sort, do transpose first
+      std::vector<int> trans;
+      for (int i = 0; i < axis; i++) {
+        trans.push_back(i);
+      }
+      trans.push_back(in_dims.size() - 1);
+      for (int i = axis + 1; i < in_dims.size() - 1; i++) {
+        trans.push_back(i);
+      }
+      trans.push_back(axis);
+      framework::DDim trans_dims(in_dims);
+      for (int i = 0; i < trans.size(); i++) {
+        trans_dims[i] = in_dims[trans[i]];
+      }
+
+      Tensor trans_inp;
+      T* trans_inp_data = trans_inp.mutable_data<T>(trans_dims, ctx.GetPlace());
+      int ndims = trans.size();
+      const auto& dev_ctx = ctx.cuda_device_context();
+      // Do transpose
+      TransCompute<platform::CUDADeviceContext, T>(ndims, dev_ctx, *input,
+                                                   &trans_inp, trans);
+
+      const int64_t input_height = framework::product(
+          framework::slice_ddim(trans_dims, 0, trans_dims.size() - 1));
+      const int64_t input_width = trans_dims[trans_dims.size() - 1];
+
+      Tensor tmp_out;
+      tmp_out.mutable_data<T>(trans_dims, ctx.GetPlace());
+      T* out_data = output->mutable_data<T>(ctx.GetPlace());
+
+      Tensor tmp_indices;
+      if (input_height < INT_MAX && input_width < INT_MAX) {
+        // temp indices for sorting
+        tmp_indices.mutable_data<int>(trans_dims, ctx.GetPlace());
+        indices->mutable_data<int>(ctx.GetPlace());
+
+        ArgFullSortAscending<T, int>(
+            dev_ctx, &trans_inp, &tmp_out, &tmp_indices,
+            static_cast<int>(input_height), static_cast<int>(input_width));
+
+        TransCompute<platform::CUDADeviceContext, int>(
+            ndims, dev_ctx, tmp_indices, indices, trans);
+      } else {
+        // temp indices for sorting
+        tmp_indices.mutable_data<int64_t>(trans_dims, ctx.GetPlace());
+        indices->mutable_data<int64_t>(ctx.GetPlace());
+
+        ArgFullSortAscending<T, int64_t>(dev_ctx, &trans_inp, &tmp_out,
+                                         &tmp_indices, input_height,
+                                         input_width);
+
+        TransCompute<platform::CUDADeviceContext, int64_t>(
+            ndims, dev_ctx, tmp_indices, indices, trans);
+      }
+      // transpose back
+      TransCompute<platform::CUDADeviceContext, T>(ndims, dev_ctx, tmp_out,
+                                                   output, trans);
+      return;
+    }
   }
 };
 
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_CUDA_KERNEL(argsort, paddle::operators::ArgsortOpCUDAKernel<float>,
-                        paddle::operators::ArgsortOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    argsort, paddle::operators::ArgsortOpCUDAKernel<float>,
+    paddle::operators::ArgsortOpCUDAKernel<double>,
+    paddle::operators::ArgsortOpCUDAKernel<paddle::platform::float16>);

python/paddle/fluid/tests/unittests/test_argsort_op.py

@@ -17,12 +17,14 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle.fluid.core as core
 
 
 class TestArgsortOp(OpTest):
     def setUp(self):
         self.init_axis()
-        x = np.random.random((2, 3, 4, 5, 10)).astype("float32")
+        self.init_datatype()
+        x = np.random.random((2, 3, 4, 5, 10)).astype(self.dtype)
         if self.axis < 0:
             self.axis = self.axis + len(x.shape)
         self.indices = np.argsort(x, kind='quicksort', axis=self.axis)
@@ -35,6 +37,9 @@ class TestArgsortOp(OpTest):
     def init_axis(self):
         self.axis = -1
 
+    def init_datatype(self):
+        self.dtype = "float32"
+
     def test_check_output(self):
         self.check_output()
 
@@ -49,10 +54,54 @@ class TestArgsortOpAxis1(TestArgsortOp):
         self.axis = 1
 
 
+class TestArgsortOpAxis2(TestArgsortOp):
+    def init_axis(self):
+        self.axis = 2
+
+
+class TestArgsortOpAxisNeg1(TestArgsortOp):
+    def init_axis(self):
+        self.axis = -1
+
+
 class TestArgsortOpAxisNeg2(TestArgsortOp):
     def init_axis(self):
         self.axis = -2
 
 
+class TestArgsortOpFP16(TestArgsortOp):
+    def init_datatype(self):
+        if core.is_compiled_with_cuda():
+            self.dtype = 'float16'
+
+    def test_check_output(self):
+        pass
+
+    def test_check_output_with_place(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            self.check_output_with_place(place, atol=1e-5)
+
+
+class TestArgsortOpFP16Axis0(TestArgsortOpFP16):
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestArgsortOpFP16Axis2(TestArgsortOpFP16):
+    def init_axis(self):
+        self.axis = 2
+
+
+class TestArgsortOpFP16AxisNeg2(TestArgsortOpFP16):
+    def init_axis(self):
+        self.axis = -2
+
+
+class TestArgsortOpFP16Axis4Neg4(TestArgsortOpFP16):
+    def init_axis(self):
+        self.axis = -4
+
+
 if __name__ == "__main__":
     unittest.main()