replaced discrete_uniform function by rand_int64 for consistency

include/mxnet/random_generator.h

@@ -65,6 +65,11 @@ class RandGenerator<cpu, DType> {
 
     MSHADOW_XINLINE int rand() { return engine_->operator()(); }
 
+    MSHADOW_XINLINE int64_t rand_int64() {
+      int64_t result = (engine_->operator()() << 32) + engine_->operator()();
+      return result;
+    }
+
     MSHADOW_XINLINE FType uniform() {
       typedef typename std::conditional<std::is_integral<DType>::value,
       std::uniform_int_distribution<DType>,
@@ -73,15 +78,6 @@ class RandGenerator<cpu, DType> {
       return dist_uniform(*engine_);
     }
 
-    MSHADOW_XINLINE IType discrete_uniform(const int64_t lower, const int64_t upper) {
-      typedef typename std::conditional<sizeof(IType) != sizeof(int32_t) ||
-      sizeof(IType) != sizeof(int64_t),
-      std::uniform_int_distribution<int>,
-      std::uniform_int_distribution<IType>>::type GType;
-      GType dist_discrete_uniform(lower, upper);
-      return dist_discrete_uniform(*engine_);
-    }
-
     MSHADOW_XINLINE FType normal() {
       std::normal_distribution<FType> dist_normal;
       return dist_normal(*engine_);
@@ -148,6 +144,10 @@ class RandGenerator<gpu, DType> {
       return curand(&state_);
     }
 
+    MSHADOW_FORCE_INLINE __device__ int64_t rand_int64() {
+      return static_cast<int64_t>(curand(&state_) << 32) + curand(&state_);
+    }
+
     MSHADOW_FORCE_INLINE __device__ float uniform() {
       return static_cast<float>(1.0) - curand_uniform(&state_);
     }
@@ -156,13 +156,6 @@ class RandGenerator<gpu, DType> {
       return curand_normal(&state_);
     }
 
-    MSHADOW_FORCE_INLINE __device__ int discrete_uniform(const int64_t lower, const int64_t upper) {
-      float randu_f = curand_uniform(&state_);
-      randu_f *= (upper-lower+0.999999);
-      randu_f += lower;
-      return static_cast<int>(trunc(randu_f));
-    }
-
    private:
     RandGenerator<gpu, DType> *global_gen_;
     int global_state_idx_;
@@ -207,6 +200,10 @@ class RandGenerator<gpu, double> {
       return curand(&state_);
     }
 
+    MSHADOW_FORCE_INLINE __device__ int64_t rand_int64() {
+      return static_cast<int64_t>(curand(&state_) << 32) + curand(&state_);
+    }
+
     MSHADOW_FORCE_INLINE __device__ double uniform() {
       return static_cast<float>(1.0) - curand_uniform_double(&state_);
     }
@@ -215,13 +212,6 @@ class RandGenerator<gpu, double> {
       return curand_normal_double(&state_);
     }
 
-    MSHADOW_FORCE_INLINE __device__ int discrete_uniform(const int64_t lower, const int64_t upper) {
-      float randu_f = curand_uniform(&state_);
-      randu_f *= (upper-lower+0.999999);
-      randu_f += lower;
-      return static_cast<int>(trunc(randu_f));
-    }
-
    private:
     RandGenerator<gpu, double> *global_gen_;
     int global_state_idx_;

src/operator/random/sampler.h

@@ -101,7 +101,12 @@ struct SampleRandIntKernel {
                                   const IType *lower, const IType *upper, OType *out) {
     RNG_KERNEL_LOOP(xpu, OType, id, gen, N, step, {
       index_t nBatch(1 + (nSample - 1) / nParm);
-      out[i] = OType(genImpl.discrete_uniform(lower[i / nBatch], upper[i / nBatch]));
+      if (sizeof(IType) == sizeof(int64_t))
+        out[i] = OType(lower[i / nBatch] +
+                     (upper[i / nBatch] - lower[i / nBatch]) * genImpl.rand_int64());
+      else
+        out[i] = OType(lower[i / nBatch] +
+                      (upper[i / nBatch] - lower[i / nBatch]) * genImpl.rand());
     });
   }
 };