remove redundant HPPL_TYPE_DOUBLE (#200)

CMakeLists.txt

@@ -104,7 +104,7 @@ else()
 endif(NOT WITH_GPU)
 
 if(WITH_DOUBLE)
-    add_definitions(-DPADDLE_TYPE_DOUBLE -DHPPL_TYPE_DOUBLE)
+    add_definitions(-DPADDLE_TYPE_DOUBLE)
     set(ACCURACY double)
 else(WITH_DOUBLE)
     set(ACCURACY float)

paddle/cuda/include/hl_base.h

@@ -185,7 +185,7 @@ typedef struct {
     size_t                  nnz;
 } _hl_sparse_matrix_s, *hl_sparse_matrix_s;
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 /**
  * HPPL data type: real (float or double)
  *

paddle/cuda/include/hl_cpu_gru.cuh

@@ -20,7 +20,7 @@ limitations under the License. */
 
 #include "paddle/math/MathFunctions.h"
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 #define     CBLAS_GEMM     paddle::gemm<float>
 #else
 #define     CBLAS_GEMM     paddle::gemm<double>

paddle/cuda/include/hl_gpu_functions.cuh

@@ -28,7 +28,7 @@ namespace hppl {
     const real min = SIGMOID_THRESHOLD_MIN;
     const real max = SIGMOID_THRESHOLD_MAX;
     real tmp = (a < min) ? min : ((a > max) ? max : a);
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     return __fdividef(1.0f, 1.0f + __expf(-tmp));
 #else
     return 1.0 / (1.0 + exp(-tmp));
@@ -36,7 +36,7 @@ namespace hppl {
   }
 
   __device__ static real tanh(const real a) {
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     return __fdividef(2.0f, (1.0f + __expf(-2.0f*a))) - 1.0f;
 #else
     return (2.0 / (1.0 + exp(-2.0*a))) - 1.0;

paddle/cuda/include/hl_matrix_base.cuh

@@ -30,7 +30,7 @@ limitations under the License. */
 #define INLINE   inline
 #endif
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 #define     DEVICE_FMAX     fmaxf
 #define     DEVICE_FMIN     fminf
 #else

paddle/cuda/include/hl_matrix_type.cuh

@@ -21,7 +21,7 @@ limitations under the License. */
 #ifdef __CUDA_ARCH__
 // typedef void*  vecType;
 #include <vector_types.h>
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 typedef float4 vecType;
 #else
 typedef double2 vecType;
@@ -30,7 +30,7 @@ typedef double2 vecType;
 #include <mmintrin.h>
 #include <xmmintrin.h>
 #include <emmintrin.h>
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 typedef __m128  vecType;
 #else
 typedef __m128d vecType;

paddle/cuda/include/hl_sse_matrix_kernel.cuh

@@ -20,7 +20,7 @@ limitations under the License. */
 
 #define VECTOR_SIZE     16
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 /* number of float in vector */
 #define     VECTOR_LEN      4
 #define     VECTOR_SET      _mm_set_ps1
@@ -41,7 +41,7 @@ inline bool hl_check_align(void *ptr) {
   return hl_check_align(reinterpret_cast<size_t>(ptr));
 }
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 template <class Agg>
 inline real hl_agg_op(Agg agg, vecType mm) {
   __m128 lo = _mm_unpacklo_ps(mm, mm);

paddle/cuda/src/hl_cuda_cublas.cc

@@ -84,7 +84,7 @@ CUBLAS_BLAS_ROUTINE_EACH(DYNAMIC_LOAD_CUBLAS_V2_WRAP)
 } /* namespace dynload */
 
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 #define     CUBLAS_GEAM     dynload::cublasSgeam
 #define     CUBLAS_GEMV     dynload::cublasSgemv
 #define     CUBLAS_GEMM     dynload::cublasSgemm

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -340,7 +340,7 @@ void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc,
         (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor));
     CHECK_NOTNULL(hl_desc);
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
 #else
     cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
@@ -373,7 +373,7 @@ void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc) {
         (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor));
     CHECK_NOTNULL(hl_desc);
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
 #else
     cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
@@ -611,7 +611,7 @@ void hl_create_filter_descriptor(hl_filter_descriptor* filter,
 
     CHECK_CUDNN(dynload::cudnnCreateFilterDescriptor(&hl_filter->desc));
 
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
 #else
     cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
@@ -921,7 +921,7 @@ void hl_softmax_forward(real *input,
                         int height,
                         int width)
 {
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
 #else
     cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
@@ -955,7 +955,7 @@ void hl_softmax_backward(real *output_value,
                          int height,
                          int width)
 {
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
 #else
     cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;

paddle/cuda/src/hl_cuda_device.cc

@@ -626,7 +626,7 @@ void hl_specify_devices_start(int* device, int number) {
 void hl_rand(real *dest_d, size_t num) {
   pthread_mutex_lock(t_resource.gen_mutex);
   CHECK_EQ(
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
   dynload::curandGenerateUniform(t_resource.gen, dest_d, num),
 #else
   dynload::curandGenerateUniformDouble(t_resource.gen, dest_d, num),

paddle/cuda/src/hl_cuda_matrix.cu

@@ -47,7 +47,7 @@ void hl_matrix_add(real *A_d,
   CHECK_SYNC("hl_matrix_add failed");
 }
 
-#ifdef HPPL_TYPE_DOUBLE
+#ifdef PADDLE_TYPE_DOUBLE
     #define THRESHOLD   128
 #else
     #define THRESHOLD   64
@@ -102,7 +102,7 @@ void subMaxAndExp(real* I,
       val = -THRESHOLD;
     }
     I[nextIdx] = val;
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
     O[nextIdx] = __expf(val);
 #else
     O[nextIdx] = exp(val);

paddle/cuda/src/hl_cuda_sparse.cuh

@@ -355,7 +355,7 @@ __global__ void KeSMatrixCscMulDense(real *C_d,
 }
 
 /* best perf */
-#ifndef HPPL_TYPE_DOUBLE
+#ifndef PADDLE_TYPE_DOUBLE
 #define CU_CSCMM_THREAD_M_BEST          9
 #else
 #define CU_CSCMM_THREAD_M_BEST          4