[singa] branch dev updated: fix gaussian CURAND_STATUS_LENGTH_NOT_MULTIPLE

[wa...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

wangwei pushed a commit to branch dev
in repository https://gitbox.apache.org/repos/asf/singa.git


The following commit(s) were added to refs/heads/dev by this push:
     new 2d28b11  fix gaussian CURAND_STATUS_LENGTH_NOT_MULTIPLE
     new 8656a4c  Merge pull request #729 from dcslin/gaussian
2d28b11 is described below

commit 2d28b113730f8ac2e00497857ced80e8f4e8aab9
Author: dcslin <13...@users.noreply.github.com>
AuthorDate: Fri Jun 12 01:57:59 2020 +0000

    fix gaussian CURAND_STATUS_LENGTH_NOT_MULTIPLE
---
 src/core/tensor/tensor_math_cuda.h | 11 ++++++++++-
 test/python/test_tensor.py         |  7 +++++++
 2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/src/core/tensor/tensor_math_cuda.h b/src/core/tensor/tensor_math_cuda.h
index bcc97fb..f331cd7 100644
--- a/src/core/tensor/tensor_math_cuda.h
+++ b/src/core/tensor/tensor_math_cuda.h
@@ -802,7 +802,16 @@ void Gaussian<float, lang::Cuda>(const float mean, const float std, Tensor* out,
   auto rgen = ctx->curand_generator;
   float* outPtr = static_cast<float*>(out->block()->mutable_data());
   const size_t num = out->Size();
-  CURAND_CHECK(curandGenerateNormal(rgen, outPtr, num, mean, std));
+
+  // CURAND_STATUS_LENGTH_NOT_MULTIPLE
+  if (num % 2 != 0) {
+    Tensor tmp(Shape{num + 1}, out->device());
+    float* outPtr_tmp = static_cast<float*>(tmp.block()->mutable_data());
+    CURAND_CHECK(curandGenerateNormal(rgen, outPtr_tmp, num + 1, mean, std));
+    CopyDataToFrom(out, tmp, num, 0, 0);
+  } else {
+    CURAND_CHECK(curandGenerateNormal(rgen, outPtr, num, mean, std));
+  }
 }
 
 // =========================Blas operations==================================
diff --git a/test/python/test_tensor.py b/test/python/test_tensor.py
index a3e4b2c..9ba2899 100644
--- a/test/python/test_tensor.py
+++ b/test/python/test_tensor.py
@@ -467,6 +467,13 @@ class TestTensorMethods(unittest.TestCase):
     def test_matmul_transpose_gpu(self):
         self._matmul_transpose_helper(gpu_dev)
 
+    @unittest.skipIf(not singa_api.USE_CUDA, 'CUDA is not enabled')
+    def test_gaussian_gpu(self, dev=gpu_dev):
+        x = tensor.Tensor((3, 5, 3, 5), device=dev)
+        x.gaussian(0, 1)
+        x = tensor.Tensor((4, 5, 3, 2), device=dev)
+        x.gaussian(0, 1)
+
 
 if __name__ == '__main__':
     unittest.main()