CPU-Adam: Fix compile Issue (#1537)

* fixing the softmax masking when using triangular masking

* move the TILE declaration outside of the SIMD loop

* remove unrelated changes

* fix Adagrad compile issue

csrc/includes/cpu_adagrad.h

@@ -32,6 +32,7 @@ public:
         cudaFreeHost(_doubled_buffer[0]);
         cudaFreeHost(_doubled_buffer[1]);
     }
+#if defined(__AVX512__) or defined(__AVX256__)
     template <int span>
     void Step_AVX(size_t* rounded_size,
                   float* _params,
@@ -40,11 +41,10 @@ public:
                   size_t param_size,
                   __half* dev_param = nullptr,
                   bool half_precision = false);
-#if defined(__AVX512__) or defined(__AVX256__)
+#endif
     STEP(1)
     STEP(4)
     STEP(8)
-#endif
     inline void SynchronizeStreams()
     {
         for (int i = 0; i < 2; i++) cudaStreamSynchronize(_streams[i]);

csrc/includes/simd.h

@@ -5,12 +5,11 @@
 #include <x86intrin.h>
 #endif
 
+#define TILE (128 * 1024 * 1024)
 #if defined(__AVX512__) or defined(__AVX256__)
 
 #define ROUND_DOWN(size, step) ((size) & ~((step)-1))
 
-#define TILE (128 * 1024 * 1024)
-
 #if defined(__AVX512__)
 #define SIMD_STORE(a, d) _mm512_storeu_ps(a, d)
 #define SIMD_LOAD(x) _mm512_loadu_ps(x)