Fix compilation on CPUs that don't support AVX and some white space i… (#561)

* Fix compilation on CPUs that don't support AVX and some white space issues
* Change from ifdef to ifndef

CMakeLists.txt

@@ -117,6 +117,8 @@ else(MSVC)
       message(STATUS "AVX support found")
       set(INTRINSICS "${INTRINSICS} -mavx")
       list(APPEND INTRINSICS_NVCC -Xcompiler\ -mavx)
+    else()
+      add_definitions(-DNO_AVX)
     endif(AVX_FOUND)
     if(AVX2_FOUND)
       message(STATUS "AVX2 support found")

src/common/types.h

@@ -31,7 +31,7 @@
 #include <cuda.h> // required to see CUDA_VERSION
 #if (CUDA_VERSION > 9000)
 #define COMPILE_FP16 1
-#else 
+#else
 #define COMPILE_FP16 0
 #endif
 #else
@@ -174,6 +174,7 @@ public:
 };
 
 // @TODO: consider how code can be shared via templating
+#ifndef NO_AVX
 struct float32x8 {
 private:
   __m256 f_;
@@ -199,6 +200,11 @@ public:
     return out;
   }
 };
+#else
+//Dummy version to get things to compile on older CPUs
+struct float32x8 {
+};
+#endif
 #endif
 
 // Internal to types.h, don't use. Use test functions below.
@@ -207,7 +213,7 @@ enum class TypeClass : size_t {
   unsigned_type = 0x200,
   float_type    = 0x400,
   packed_type   = 0x800, // special packed (CPU cache friendly) type class, used in FBGEMM, not meant to be used anywhere else
-            
+
   size_mask     = 0x0FF
 };
 

src/functional/operators.h

@@ -319,7 +319,7 @@ struct Ops<float32x4> {
 
 } // end namespace functional
 } // end namespace marian
-
+#ifndef NO_AVX
 #include "3rd_party/avx_mathfun.h"
 
 namespace marian {
@@ -438,7 +438,7 @@ struct Ops<float32x8> {
 
 } // end namespace functional
 } // end namespace marian
-
+#endif
 #endif // of "#ifndef __CUDACC__"
 
 #ifdef __CUDACC__
@@ -600,4 +600,4 @@ BINARY(sPReLU,     PReLU,     Ops<ElementType>::prelu(x, y));
 BINARY(sPReLUBack, PReLUback, Ops<ElementType>::preluBack(x, y));
 
 } // end namespace functional
-} // end namespace marian
+} // end namespace marian

src/functional/tensor.h

@@ -7,8 +7,8 @@
 namespace marian {
 namespace functional {
 
-// By default for single valued types like float do nothing. Usually the number of elements in a tensor 
-// is correctly mirrored in the shape object. Only special multi-element types like float32x4 (4 floats), 
+// By default for single valued types like float do nothing. Usually the number of elements in a tensor
+// is correctly mirrored in the shape object. Only special multi-element types like float32x4 (4 floats),
 // float32x8 (8 floats) and half2 (2 half) require special handling done by specializations below.
 // Similar for multi-element integer types to be added later.
 template <typename T>
@@ -31,7 +31,7 @@ inline marian::Shape adapt<float32x4>(const marian::Shape& shape) {
   x4Shape.set(-1, shape[-1] / 4);
   return x4Shape;
 }
-
+#ifndef NO_AVX
 template <>
 inline marian::Shape adapt<float32x8>(const marian::Shape& shape) {
   ABORT_IF(shape[-1] % 8 != 0,
@@ -42,7 +42,7 @@ inline marian::Shape adapt<float32x8>(const marian::Shape& shape) {
   x8Shape.set(-1, shape[-1] / 8);
   return x8Shape;
 }
-
+#endif
 #endif
 
 template <typename T, const int D>
@@ -211,4 +211,4 @@ template <typename T>
 using Tensor = View<T, CONST_SHAPE_DIMS>;
 
 }  // namespace functional
-}  // namespace marian
+}  // namespace marian

src/tensors/cpu/element.h

@@ -99,8 +99,10 @@ void elementFloat(const Functor& functor, marian::Tensor out, Tensors... tensors
 
   if(div8) {
     // std::cerr << "8: " << functor.to_string() << std::endl;
+#ifndef NO_AVX
     element<float32x8>(functor, out, tensors...);
     return;
+#endif
   }
 
   if(div4) {

src/tensors/cpu/tensor_operators.cpp

@@ -425,9 +425,13 @@ void Softmax(Tensor out, Tensor in) {
   matchOrAbort<float>(out->type());
   matchOrAbort<float>(in->type());
 
+#ifndef NO_AVX
   if(out->shape()[-1] % 8 == 0) {
     Softmax<float32x8>(out, in);
-  } else if(out->shape()[-1] % 4 == 0) {
+    return;
+  }
+#endif
+  if(out->shape()[-1] % 4 == 0) {
     Softmax<float32x4>(out, in);
   } else {
     Softmax<float>(out, in);
@@ -477,9 +481,13 @@ void LogSoftmax(Tensor out, Tensor in) {
   matchOrAbort<float>(out->type());
   matchOrAbort<float>(in->type());
 
+#ifndef NO_AVX
   if(out->shape()[-1] % 8 == 0) {
     LogSoftmax<float32x8>(out, in);
-  } else if(out->shape()[-1] % 4 == 0) {
+    return;
+  }
+#endif
+  if(out->shape()[-1] % 4 == 0) {
     LogSoftmax<float32x4>(out, in);
   } else {
     LogSoftmax<float>(out, in);