Fix cmake error of failing to find UINT64_MAX.

CMakeLists.txt

@@ -25,6 +25,7 @@ find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
 
 include(system)
+include(simd)
 
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
@@ -64,7 +65,6 @@ include(external/openblas)  # download, build, install openblas
 include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
 
-include(simd)               # set simd flag
 include(package)            # set paddle packages
 include(cpplint)            # set paddle c++ style
 include(ccache)             # set ccache for compilation

cmake/configure.cmake

@@ -32,6 +32,16 @@ if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)
 
+if(NEON_FOUND)
+    set(SIMD_FLAG ${NEON_FLAG})
+else(NEON_FOUND)
+    if(WITH_AVX)
+        set(SIMD_FLAG ${AVX_FLAG})
+    else(WITH_AVX)
+        set(SIMD_FLAG ${SSE3_FLAG})
+    endif(WITH_AVX)
+endif(NEON_FOUND)
+
 if(NOT WITH_GPU)
     add_definitions(-DPADDLE_ONLY_CPU)
     add_definitions(-DHPPL_STUB_FUNC)
@@ -48,11 +58,7 @@ else()
         message(FATAL_ERROR "Paddle need cudnn to compile")
     endif()
 
-    if(WITH_AVX)
-        set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${AVX_FLAG}")
-    else(WITH_AVX)
-        set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SSE3_FLAG}")
-    endif(WITH_AVX)
+    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SIMD_FLAG}")
 
     # Include cuda and cudnn
     include_directories(${CUDNN_INCLUDE_DIR})

cmake/simd.cmake

@@ -88,14 +88,5 @@ int main()
     return 0;
 }" NEON_FOUND)
 
-if(NEON_FOUND)
-    set(SIMD_FLAG ${NEON_FLAG})
-else(NEON_FOUND)
-    if(WITH_AVX)
-        set(SIMD_FLAG ${AVX_FLAG})
-    else(WITH_AVX)
-        set(SIMD_FLAG ${SSE3_FLAG})
-    endif(WITH_AVX)
-endif(NEON_FOUND)
-
+set(CMAKE_REQUIRED_FLAGS "")
 mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND NEON_FOUND)

paddle/math/SIMDFunctions.cpp

@@ -13,122 +13,12 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "SIMDFunctions.h"
-#ifdef __SSE__
+#ifdef __SSE3__
 #include <immintrin.h>
 #endif
 #include <algorithm>
 
-#ifdef __SSE__
-static void addto_sse(float* a, const float* b, size_t len) {
-  int offset = len % 16;
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-
-  for (unsigned int k = 0; k < len / 16; k++, a += 16, b += 16) {
-    ma0 = _mm_load_ps(a);
-    ma1 = _mm_load_ps(a + 4);
-    ma2 = _mm_load_ps(a + 8);
-    ma3 = _mm_load_ps(a + 12);
-
-    mb0 = _mm_load_ps(b);
-    mb1 = _mm_load_ps(b + 4);
-    mb2 = _mm_load_ps(b + 8);
-    mb3 = _mm_load_ps(b + 12);
-
-    ma0 = _mm_add_ps(ma0, mb0);
-    ma1 = _mm_add_ps(ma1, mb1);
-    ma2 = _mm_add_ps(ma2, mb2);
-    ma3 = _mm_add_ps(ma3, mb3);
-
-    _mm_store_ps(a, ma0);
-    _mm_store_ps(a + 4, ma1);
-    _mm_store_ps(a + 8, ma2);
-    _mm_store_ps(a + 12, ma3);
-  }
-
-  for (int i = 0; i < offset; i++) a[i] += b[i];
-}
-
-static void batch_addto_sse(float* a, const float* b[], int batch, size_t len) {
-  int offset = len % 16;
-
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-
-  for (unsigned int k = 0; k < len / 16; k++, a += 16) {
-    ma0 = _mm_load_ps(a);
-    ma1 = _mm_load_ps(a + 4);
-    ma2 = _mm_load_ps(a + 8);
-    ma3 = _mm_load_ps(a + 12);
-
-    for (int i = 0; i < batch; i++) {
-      mb0 = _mm_load_ps(b[i]);
-      mb1 = _mm_load_ps(b[i] + 4);
-      mb2 = _mm_load_ps(b[i] + 8);
-      mb3 = _mm_load_ps(b[i] + 12);
-      ma0 = _mm_add_ps(ma0, mb0);
-      ma1 = _mm_add_ps(ma1, mb1);
-      ma2 = _mm_add_ps(ma2, mb2);
-      ma3 = _mm_add_ps(ma3, mb3);
-      b[i] += 16;
-    }
-
-    _mm_store_ps(a, ma0);
-    _mm_store_ps(a + 4, ma1);
-    _mm_store_ps(a + 8, ma2);
-    _mm_store_ps(a + 12, ma3);
-  }
-
-  for (int i = 0; i < offset; i++) {
-    for (int k = 0; k < batch; k++) a[i] += b[k][i];
-  }
-  return;
-}
-
-static void col_max_sse(float* result,
-                        const float* data,
-                        int dim,
-                        int numSamples) {
-  // first sample, direct copy
-  for (int d = 0; d < dim; ++d) {
-    result[d] = data[d];
-  }
-  int offset = dim % 16;
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-  // first 16n dims
-  for (int k = 0; k < dim / 16; k++, result += 16, data += 16) {
-    ma0 = _mm_load_ps(result);
-    ma1 = _mm_load_ps(result + 4);
-    ma2 = _mm_load_ps(result + 8);
-    ma3 = _mm_load_ps(result + 12);
-    for (int i = 1; i < numSamples; i++) {
-      mb0 = _mm_load_ps(data + i * dim);
-      mb1 = _mm_load_ps(data + i * dim + 4);
-      mb2 = _mm_load_ps(data + i * dim + 8);
-      mb3 = _mm_load_ps(data + i * dim + 12);
-      ma0 = _mm_max_ps(ma0, mb0);
-      ma1 = _mm_max_ps(ma1, mb1);
-      ma2 = _mm_max_ps(ma2, mb2);
-      ma3 = _mm_max_ps(ma3, mb3);
-    }
-    _mm_store_ps(result, ma0);
-    _mm_store_ps(result + 4, ma1);
-    _mm_store_ps(result + 8, ma2);
-    _mm_store_ps(result + 12, ma3);
-  }
-  // last dims
-  for (int d = 0; d < offset; ++d) {
-    float sm = data[d];
-    for (int i = 1; i < numSamples; ++i) {
-      sm = std::max(sm, data[i * dim + d]);
-    }
-    result[d] = sm;
-  }
-}
-
-#elif defined(__AVX__)
-
+#ifdef __AVX__
 static void addto_avx(float* a, const float* b, size_t len) {
   int offset = len % 32;
 
@@ -358,18 +248,128 @@ static void decayL1_avx(
   }
 }
 
+#elif defined(__SSE3__)
+
+static void addto_sse(float* a, const float* b, size_t len) {
+  int offset = len % 16;
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+
+  for (unsigned int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = _mm_load_ps(a);
+    ma1 = _mm_load_ps(a + 4);
+    ma2 = _mm_load_ps(a + 8);
+    ma3 = _mm_load_ps(a + 12);
+
+    mb0 = _mm_load_ps(b);
+    mb1 = _mm_load_ps(b + 4);
+    mb2 = _mm_load_ps(b + 8);
+    mb3 = _mm_load_ps(b + 12);
+
+    ma0 = _mm_add_ps(ma0, mb0);
+    ma1 = _mm_add_ps(ma1, mb1);
+    ma2 = _mm_add_ps(ma2, mb2);
+    ma3 = _mm_add_ps(ma3, mb3);
+
+    _mm_store_ps(a, ma0);
+    _mm_store_ps(a + 4, ma1);
+    _mm_store_ps(a + 8, ma2);
+    _mm_store_ps(a + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) a[i] += b[i];
+}
+
+static void batch_addto_sse(float* a, const float* b[], int batch, size_t len) {
+  int offset = len % 16;
+
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+
+  for (unsigned int k = 0; k < len / 16; k++, a += 16) {
+    ma0 = _mm_load_ps(a);
+    ma1 = _mm_load_ps(a + 4);
+    ma2 = _mm_load_ps(a + 8);
+    ma3 = _mm_load_ps(a + 12);
+
+    for (int i = 0; i < batch; i++) {
+      mb0 = _mm_load_ps(b[i]);
+      mb1 = _mm_load_ps(b[i] + 4);
+      mb2 = _mm_load_ps(b[i] + 8);
+      mb3 = _mm_load_ps(b[i] + 12);
+      ma0 = _mm_add_ps(ma0, mb0);
+      ma1 = _mm_add_ps(ma1, mb1);
+      ma2 = _mm_add_ps(ma2, mb2);
+      ma3 = _mm_add_ps(ma3, mb3);
+      b[i] += 16;
+    }
+
+    _mm_store_ps(a, ma0);
+    _mm_store_ps(a + 4, ma1);
+    _mm_store_ps(a + 8, ma2);
+    _mm_store_ps(a + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    for (int k = 0; k < batch; k++) a[i] += b[k][i];
+  }
+  return;
+}
+
+static void col_max_sse(float* result,
+                        const float* data,
+                        int dim,
+                        int numSamples) {
+  // first sample, direct copy
+  for (int d = 0; d < dim; ++d) {
+    result[d] = data[d];
+  }
+  int offset = dim % 16;
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+  // first 16n dims
+  for (int k = 0; k < dim / 16; k++, result += 16, data += 16) {
+    ma0 = _mm_load_ps(result);
+    ma1 = _mm_load_ps(result + 4);
+    ma2 = _mm_load_ps(result + 8);
+    ma3 = _mm_load_ps(result + 12);
+    for (int i = 1; i < numSamples; i++) {
+      mb0 = _mm_load_ps(data + i * dim);
+      mb1 = _mm_load_ps(data + i * dim + 4);
+      mb2 = _mm_load_ps(data + i * dim + 8);
+      mb3 = _mm_load_ps(data + i * dim + 12);
+      ma0 = _mm_max_ps(ma0, mb0);
+      ma1 = _mm_max_ps(ma1, mb1);
+      ma2 = _mm_max_ps(ma2, mb2);
+      ma3 = _mm_max_ps(ma3, mb3);
+    }
+    _mm_store_ps(result, ma0);
+    _mm_store_ps(result + 4, ma1);
+    _mm_store_ps(result + 8, ma2);
+    _mm_store_ps(result + 12, ma3);
+  }
+  // last dims
+  for (int d = 0; d < offset; ++d) {
+    float sm = data[d];
+    for (int i = 1; i < numSamples; ++i) {
+      sm = std::max(sm, data[i * dim + d]);
+    }
+    result[d] = sm;
+  }
+}
+
 #endif
 
-#ifdef __SSE__
-#define SIMD_INVOKE(func, ...) func##_sse(__VA_ARGS__)
-#elif __AVX__
+#if defined(__AVX__)
 #define SIMD_INVOKE(func, ...) func##_avx(__VA_ARGS__)
+#elif defined(__SSE3__)
+#define SIMD_INVOKE(func, ...) func##_sse(__VA_ARGS__)
 #endif
 
 namespace paddle {
 namespace simd {
 namespace internal {
-#ifdef __SSE__
+#ifdef __SSE3__
 void addToImpl(float* a, const float* b, size_t len) {
   SIMD_INVOKE(addto, a, b, len);
 }
@@ -390,8 +390,8 @@ void decayL1AvxImpl(
     float* dst, float* src, float* lr, float lambda, size_t len) {
   decayL1_avx(dst, src, lr, lambda, len);
 }
-
 #endif
+
 }  // namespace internal
 }  // namespace simd
 }  // namespace paddle

paddle/math/SIMDFunctions.h

@@ -128,7 +128,7 @@ void decayL1AvxImpl(
 
 template <>
 inline void addTo(float* a, const float* b, size_t len) {
-#ifdef __SSE__
+#ifdef __SSE3__
   internal::addToImpl(a, b, len);
 #else
   naive::addTo(a, b, len);
@@ -137,7 +137,7 @@ inline void addTo(float* a, const float* b, size_t len) {
 
 template <>
 inline void batchAddTo(float* a, const float* b[], int batch, size_t len) {
-#ifdef __SSE__
+#ifdef __SSE3__
   internal::batchAddToImpl(a, b, batch, len);
 #else
   naive::batchAddTo(a, b, batch, len);
@@ -146,7 +146,7 @@ inline void batchAddTo(float* a, const float* b[], int batch, size_t len) {
 
 template <>
 inline void colMax(float* result, const float* data, int dim, int numSamples) {
-#ifdef __SSE__
+#ifdef __SSE3__
   internal::colMaxImpl(result, data, dim, numSamples);
 #else
   naive::colMax(result, data, dim, numSamples);