Michael S (already5chosen.delete@this.yahoo.com) on January 9, 2014 3:02 pm wrote:
>

> void Demo (const double * srcX, const double * srcY, double* __restrict dst, int size)

> {

>   // Avoid the code for unaligned

>   __assume_aligned(srcX, 64);

>   __assume_aligned(srcY, 64);

>   __assume_aligned(dst, 64);

> 

>   double a3 = 1.2;

>   double a2 = 3.4;

>   double a1 = 5.6;

>   double a0 = 7.8;

>   for(int i = 0; i      double x = srcX[i];

>     double y = srcY[i];

>     dst[i] = ((((x-y)*x+a2)*x)+a1)*x+a0;

>   }

> }

KNC version (well, at least I hope that icl /Qmic -O3 produces KNC-optimized code rather than KNF).
Inner loop only:
# LOE rdx rbx rbp rsi rdi r8 r9 r10 r11 r12 r13 r14 r15 eax ecx zmm0 zmm1 zmm2
..B1.4: # Preds ..B1.4 ..B1.3 Latency 101
vmovapd (%r15,%r9,8), %zmm3 #15.16 c1
vprefetch1 2048(%r15,%r9,8) #15.16 c1
vmovapd 64(%r15,%r9,8), %zmm5 #15.16 c5
vprefetch0 512(%r15,%r9,8) #15.16 c5
vmovapd 128(%r15,%r9,8), %zmm7 #15.16 c9
vprefetch1 2048(%r11,%r9,8) #16.16 c9
vmovapd 192(%r15,%r9,8), %zmm9 #15.16 c13
vprefetch0 512(%r11,%r9,8) #16.16 c13
vsubpd (%r11,%r9,8), %zmm3, %zmm4 #17.20 c17
vprefetche1 2048(%r10,%r9,8) #17.5 c17
vsubpd 64(%r11,%r9,8), %zmm5, %zmm6 #17.20 c21
vprefetche0 512(%r10,%r9,8) #17.5 c21
vsubpd 128(%r11,%r9,8), %zmm7, %zmm8 #17.20 c25
vprefetch1 2112(%r15,%r9,8) #15.16 c25
vsubpd 192(%r11,%r9,8), %zmm9, %zmm10 #17.20 c29
vprefetch0 576(%r15,%r9,8) #15.16 c29
vfmadd132pd (%r15,%r9,8), %zmm0, %zmm4 #17.25 c33
vprefetch1 2112(%r11,%r9,8) #16.16 c33
vfmadd132pd 64(%r15,%r9,8), %zmm0, %zmm6 #17.25 c37
vprefetch0 576(%r11,%r9,8) #16.16 c37
vfmadd132pd 128(%r15,%r9,8), %zmm0, %zmm8 #17.25 c41
vprefetche1 2112(%r10,%r9,8) #17.5 c41
vfmadd132pd 192(%r15,%r9,8), %zmm0, %zmm10 #17.25 c45
vprefetche0 576(%r10,%r9,8) #17.5 c45
vfmadd132pd (%r15,%r9,8), %zmm1, %zmm4 #17.32 c49
vprefetch1 2176(%r15,%r9,8) #15.16 c49
vfmadd132pd 64(%r15,%r9,8), %zmm1, %zmm6 #17.32 c53
vprefetch0 640(%r15,%r9,8) #15.16 c53
vfmadd132pd 128(%r15,%r9,8), %zmm1, %zmm8 #17.32 c57
vprefetch1 2176(%r11,%r9,8) #16.16 c57
vfmadd132pd 192(%r15,%r9,8), %zmm1, %zmm10 #17.32 c61
vprefetch0 640(%r11,%r9,8) #16.16 c61
vfmadd132pd (%r15,%r9,8), %zmm2, %zmm4 #17.38 c65
vprefetche1 2176(%r10,%r9,8) #17.5 c65
vfmadd132pd 64(%r15,%r9,8), %zmm2, %zmm6 #17.38 c69
vprefetche0 640(%r10,%r9,8) #17.5 c69
vfmadd132pd 128(%r15,%r9,8), %zmm2, %zmm8 #17.38 c73
vprefetch1 2240(%r15,%r9,8) #15.16 c73
vfmadd132pd 192(%r15,%r9,8), %zmm2, %zmm10 #17.38 c77
vprefetch0 704(%r15,%r9,8) #15.16 c77
vmovaps %zmm4, (%r10,%r9,8) #17.5 c81
vprefetch1 2240(%r11,%r9,8) #16.16 c81
vmovaps %zmm6, 64(%r10,%r9,8) #17.5 c85
vprefetch0 704(%r11,%r9,8) #16.16 c85
vmovaps %zmm8, 128(%r10,%r9,8) #17.5 c89
vprefetche1 2240(%r10,%r9,8) #17.5 c89
vmovaps %zmm10, 192(%r10,%r9,8) #17.5 c93
vprefetche0 704(%r10,%r9,8) #17.5 c93
addq $32, %r9 #14.3 c97
cmpq %r8, %r9 #14.3 c101
jb ..B1.4 # Prob 82% #14.3 c101
# LOE rdx rbx rbp rsi rdi r8 r9 r10 r11 r12 r13 r14 r15 eax ecx zmm0 zmm1 zmm2



Except for addition of prefetch instructions, generated code looks nearly identical to KNL.