The present patch verifies that, on machines declaring an Athlon CPU model and
family, the 3dnow and 3dnowext feature flags are indeed present. If they are
not, it fallbacks on the most generic x86 kernel. This prevents crashes due to
illegal instruction on qemu guests with a weird configuration.

Closes #272

Conflicts:
	kernel/x86_64/KERNEL.BULLDOZER

Conflicts:
	Makefile.system

parallelization model is used by OpenBLAS.

I used a smaller threshold since the stack size is 1MB on windows.

For example, make CC=lsbcc NO_LAPACK=1.

When NUM_THREADS(MAX_CPU_NUNBERS) is very large ,e.g. 256.

typedef struct {
  volatile BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE];
} job_t;

job_t          job[MAX_CPU_NUMBER];

The job array is equal 8MB.

Thus, We use malloc instead of stack allocation.

Refs #230. New dgemm and sgemm Kernel for BULLDOZER

Instead of using thread 0 buffer, each thread uses its own sb buffer.
Thus, it can avoid overflowing thread 0 buffer.

if we have 64 bit pointers we can't have a 32 bit cpu, so fall back to
the 64bit cpu fallback (prescott)
E.g. the cpu detection fails in amd qemu64 emulation (family 6 model 2)
causing it to use the uninitialized gotoblas_ATHLON

openblas_get_config function returns the configure string.
So far, it supports USE64BITINT, NO_CBLAS, NO_LAPACK, NO_LAPACKE,
DYNAMIC_ARCH, NO_AFFINITY.

Example:
 #include <stdio.h>
extern char * openblas_get_config();
void main()
{
  printf("%s\n",openblas_get_config());
  return;
}

Please read the issue on github.com for the detail.

For example, make NO_AVX=1 or make DYNAMIC_ARCH=1 NO_AVX=1