Move float16 into fluid folder (#8394)

* move float16 into fluid

* fix include

* move to platform folder

paddle/fluid/platform/CMakeLists.txt

@@ -27,7 +27,7 @@ ELSE()
     set(MKLDNN_CTX_DEPS)
 ENDIF()
 
-# memcpy deoends on device_context, here add deps individually for
+# memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
     system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
@@ -39,3 +39,6 @@ nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
 
 cc_library(profiler SRCS profiler.cc DEPS device_context)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
+
+nv_test(float16_gpu_test SRCS float16_test.cu)
+cc_test(float16_test SRCS float16_test.cc)

paddle/math/float16.h → paddle/fluid/platform/float16.h

@@ -68,7 +68,7 @@ namespace paddle {
 // memory access of float16 struct and also makes float16 compatible
 // with CUDA half, ARM float16_t, and Eigen::half data types.
 struct PADDLE_ALIGN(2) float16 {
-public:
+ public:
   uint16_t x;
 
   // Constructors
@@ -319,7 +319,7 @@ public:
     return static_cast<double>(float(*this));
   }
 
-private:
+ private:
   union Bits {
     float f;
     int32_t si;
@@ -485,8 +485,7 @@ HOST inline float16 operator+(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&(res.x))
       :  // clobbers
       "memory", "v0", "v1");
@@ -502,8 +501,7 @@ HOST inline float16 operator-(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&(res.x))
       :  // clobbers
       "memory", "v0", "v1");
@@ -519,8 +517,7 @@ HOST inline float16 operator*(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&(res.x))
       :  // clobbers
       "memory", "v0", "v1");
@@ -536,8 +533,7 @@ HOST inline float16 operator/(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&(res.x))
       :  // clobbers
       "memory", "v0", "v1");
@@ -588,8 +584,7 @@ HOST inline bool operator==(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&res)
       :  // clobbers
       "memory", "v0", "v1");
@@ -609,8 +604,7 @@ HOST inline bool operator<(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&res)
       :  // clobbers
       "memory", "v0", "v1");
@@ -626,8 +620,7 @@ HOST inline bool operator<=(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&res)
       :  // clobbers
       "memory", "v0", "v1");
@@ -643,8 +636,7 @@ HOST inline bool operator>(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&res)
       :  // clobbers
       "memory", "v0", "v1");
@@ -660,8 +652,7 @@ HOST inline bool operator>=(const float16& a, const float16& b) {
       "st1 {v0.h}[0], [%[res_ptr]]\n"
       :  // outputs
       :  // inputs
-      [a_ptr] "r"(&(a.x)),
-      [b_ptr] "r"(&(b.x)),
+      [a_ptr] "r"(&(a.x)), [b_ptr] "r"(&(b.x)),
       [res_ptr] "r"(&res)
       :  // clobbers
       "memory", "v0", "v1");

paddle/math/tests/test_float16.cpp → paddle/fluid/platform/float16_test.cc

@@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/math/float16.h"
+#include "paddle/fluid/platform/float16.h"
 
 #include <gtest/gtest.h>
 

paddle/math/tests/test_float16.cu → paddle/fluid/platform/float16_test.cu

@@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/math/float16.h"
+#include "paddle/fluid/platform/float16.h"
 
 #include <gtest/gtest.h>
 

paddle/math/tests/CMakeLists.txt

@@ -22,7 +22,6 @@ if(WITH_GPU)
     link_paddle_test(test_Tensor)
     CUDA_ADD_EXECUTABLE(test_lazyAssign test_lazyAssign.cu)
     link_paddle_test(test_lazyAssign)
-    nv_test(test_float16_gpu SRCS test_float16.cu)
 else()
     compile_cu_as_cpp(test_Tensor.cu)
     add_unittest(test_Tensor test_Tensor.cu)
@@ -34,4 +33,3 @@ add_simple_unittest(test_FPException)
 add_simple_unittest(test_GpuProfiler)
 add_simple_unittest(test_BaseMatrix)
 add_simple_unittest(test_Matrix)
-add_simple_unittest(test_float16)