[Performance] Improving speed by splitting kernels (pytorch#97)

With this patch, kernels can be splitted into odd and even parts in order to make computation faster.

If the processor is powerful enough, these two parts can be executed in parallel by out-of-order execution.
However, there is some overhead that is required to split the original kernel into two.
The degree of speed-up depends on the micro architecture, and the split kernels may be slower on some micro architectures.
This feature is enabled by SPLIT_KERNEL macro, and currently, it is enabled only for AVX2 and AVX512F.
This change affects computation speed of many functions, but for example,
the ratio of execution time between the previous version and the new version of asind4_u35 is 2.08.

src/arch/helperavx2.h

@@ -24,6 +24,7 @@
 #define ENABLE_FMA_SP
 
 #define FULL_FP_ROUNDING
+#define SPLIT_KERNEL
 
 #if defined(_MSC_VER)
 #include <intrin.h>

src/arch/helperavx2_128.h

@@ -24,6 +24,7 @@
 #define ENABLE_FMA_SP
 
 #define FULL_FP_ROUNDING
+#define SPLIT_KERNEL
 
 #if defined(_MSC_VER)
 #include <intrin.h>

src/arch/helperavx512f.h

@@ -24,6 +24,7 @@
 #define ENABLE_FMA_SP
 
 #define FULL_FP_ROUNDING
+#define SPLIT_KERNEL
 
 #if defined(_MSC_VER)
 #include <intrin.h>