diff --git a/kernel/x86_64/caxpy.c b/kernel/x86_64/caxpy.c index be945a441..29342f46f 100644 --- a/kernel/x86_64/caxpy.c +++ b/kernel/x86_64/caxpy.c @@ -78,13 +78,13 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, if ( (inc_x == 1) && (inc_y == 1) ) { - int n1 = n & -8; + int n1 = n & -16; if ( n1 ) { da[0] = da_r; da[1] = da_i; - caxpy_kernel_8(n1, x, y , &da ); + caxpy_kernel_8(n1, x, y , da ); ix = 2 * n1; } i = n1; diff --git a/kernel/x86_64/caxpy_microk_bulldozer-2.c b/kernel/x86_64/caxpy_microk_bulldozer-2.c index 63575c374..33bda0943 100644 --- a/kernel/x86_64/caxpy_microk_bulldozer-2.c +++ b/kernel/x86_64/caxpy_microk_bulldozer-2.c @@ -31,89 +31,87 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __att static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) { +#if !defined(CONJ) + FLOAT mvec[4] = { -1.0, 1.0, -1.0, 1.0 }; +#else + FLOAT mvec[4] = { 1.0, -1.0, 1.0, -1.0 }; +#endif BLASLONG register i = 0; + if ( n < 640 ) + { + __asm__ __volatile__ ( + "vzeroupper \n\t" "vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha "vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulps (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulps (%5), %%xmm0 , %%xmm0 \n\t" +#endif ".align 16 \n\t" "1: \n\t" - "prefetcht0 768(%2,%0,4) \n\t" "vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x + ".align 2 \n\t" "vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x "vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x "vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x - "prefetcht0 768(%3,%0,4) \n\t" -#if !defined(CONJ) - "vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm12 \n\t" + "vmovups 64(%2,%0,4), %%xmm12 \n\t" // 2 complex values from x + "vmovups 80(%2,%0,4), %%xmm13 \n\t" // 2 complex values from x + "vmovups 96(%2,%0,4), %%xmm14 \n\t" // 2 complex values from x + "vmovups 112(%2,%0,4), %%xmm15 \n\t" // 2 complex values from x + "vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part - "vmulps %%xmm1, %%xmm4 , %%xmm4 \n\t" - - "vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm13 \n\t" "vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part - "vmulps %%xmm1, %%xmm6 , %%xmm6 \n\t" - - "vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm14 \n\t" "vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part - "vmulps %%xmm1, %%xmm8 , %%xmm8 \n\t" - - "vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm15 \n\t" "vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part - "vmulps %%xmm1, %%xmm10, %%xmm10 \n\t" - "vaddsubps %%xmm4, %%xmm12, %%xmm12 \n\t" - "vaddsubps %%xmm6, %%xmm13, %%xmm13 \n\t" - "vaddsubps %%xmm8, %%xmm14, %%xmm14 \n\t" - "vaddsubps %%xmm10,%%xmm15, %%xmm15 \n\t" + "vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm5 \n\t" + ".align 2 \n\t" + "vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm7 \n\t" + "vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm9 \n\t" + "vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm11 \n\t" -#else + "vfmaddps %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmaddps %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmaddps %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmaddps %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t" - "vmulps %%xmm0, %%xmm5, %%xmm4 \n\t" // a_r*x_r, a_r*x_i - "vmulps %%xmm1, %%xmm5, %%xmm5 \n\t" // a_i*x_r, a_i*x_i - "vmulps %%xmm0, %%xmm7, %%xmm6 \n\t" // a_r*x_r, a_r*x_i - "vmulps %%xmm1, %%xmm7, %%xmm7 \n\t" // a_i*x_r, a_i*x_i - "vmulps %%xmm0, %%xmm9, %%xmm8 \n\t" // a_r*x_r, a_r*x_i - "vmulps %%xmm1, %%xmm9, %%xmm9 \n\t" // a_i*x_r, a_i*x_i - "vmulps %%xmm0, %%xmm11, %%xmm10 \n\t" // a_r*x_r, a_r*x_i - "vmulps %%xmm1, %%xmm11, %%xmm11 \n\t" // a_i*x_r, a_i*x_i + "vpermilps $0xb1 , %%xmm12, %%xmm4 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm13, %%xmm6 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm14, %%xmm8 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm15, %%xmm10 \n\t" // exchange real and imag part - "vpermilps $0xb1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part - "vaddsubps %%xmm4 ,%%xmm5 , %%xmm4 \n\t" - "vpermilps $0xb1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part + "vfmaddps 64(%3,%0,4), %%xmm0 , %%xmm12, %%xmm12 \n\t" + "vfmaddps 80(%3,%0,4), %%xmm0 , %%xmm13, %%xmm13 \n\t" + "vfmaddps 96(%3,%0,4), %%xmm0 , %%xmm14, %%xmm14 \n\t" + "vfmaddps 112(%3,%0,4), %%xmm0 , %%xmm15, %%xmm15 \n\t" - "vpermilps $0xb1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part - "vaddsubps %%xmm6 ,%%xmm7 , %%xmm6 \n\t" - "vpermilps $0xb1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part + "vfmaddps %%xmm12, %%xmm1 , %%xmm4 , %%xmm12 \n\t" + "vfmaddps %%xmm13, %%xmm1 , %%xmm6 , %%xmm13 \n\t" + "vfmaddps %%xmm14, %%xmm1 , %%xmm8 , %%xmm14 \n\t" + "vfmaddps %%xmm15, %%xmm1 , %%xmm10, %%xmm15 \n\t" - "vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part - "vaddsubps %%xmm8 ,%%xmm9 , %%xmm8 \n\t" - "vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part + "vmovups %%xmm5 , (%3,%0,4) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,4) \n\t" + "vmovups %%xmm9 , 32(%3,%0,4) \n\t" + "vmovups %%xmm11, 48(%3,%0,4) \n\t" + "vmovups %%xmm12, 64(%3,%0,4) \n\t" + "vmovups %%xmm13, 80(%3,%0,4) \n\t" + "vmovups %%xmm14, 96(%3,%0,4) \n\t" + "vmovups %%xmm15,112(%3,%0,4) \n\t" - "vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part - "vaddsubps %%xmm10,%%xmm11, %%xmm10 \n\t" - "vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part - - "vaddps (%3,%0,4) ,%%xmm4 , %%xmm12 \n\t" - "vaddps 16(%3,%0,4) ,%%xmm6 , %%xmm13 \n\t" - "vaddps 32(%3,%0,4) ,%%xmm8 , %%xmm14 \n\t" - "vaddps 48(%3,%0,4) ,%%xmm10, %%xmm15 \n\t" - - -#endif - - "vmovups %%xmm12, (%3,%0,4) \n\t" - "vmovups %%xmm13, 16(%3,%0,4) \n\t" - "vmovups %%xmm14, 32(%3,%0,4) \n\t" - "vmovups %%xmm15, 48(%3,%0,4) \n\t" - - "addq $16, %0 \n\t" - "subq $8 , %1 \n\t" + "addq $32, %0 \n\t" + "subq $16, %1 \n\t" "jnz 1b \n\t" + "vzeroupper \n\t" : : @@ -121,7 +119,8 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) "r" (n), // 1 "r" (x), // 2 "r" (y), // 3 - "r" (alpha) // 4 + "r" (alpha), // 4 + "r" (mvec) // 5 : "cc", "%xmm0", "%xmm1", "%xmm4", "%xmm5", "%xmm6", "%xmm7", @@ -129,7 +128,73 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) "%xmm12", "%xmm13", "%xmm14", "%xmm15", "memory" ); + return; + } + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vbroadcastss (%4), %%xmm0 \n\t" // real part of alpha + "vbroadcastss 4(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulps (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulps (%5), %%xmm0 , %%xmm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "prefetcht0 512(%2,%0,4) \n\t" + "vmovups (%2,%0,4), %%xmm5 \n\t" // 2 complex values from x + ".align 2 \n\t" + "vmovups 16(%2,%0,4), %%xmm7 \n\t" // 2 complex values from x + "vmovups 32(%2,%0,4), %%xmm9 \n\t" // 2 complex values from x + "vmovups 48(%2,%0,4), %%xmm11 \n\t" // 2 complex values from x + + "vpermilps $0xb1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part + "vpermilps $0xb1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part + + "prefetcht0 512(%3,%0,4) \n\t" + "vfmaddps (%3,%0,4), %%xmm0 , %%xmm5, %%xmm5 \n\t" + ".align 2 \n\t" + "vfmaddps 16(%3,%0,4), %%xmm0 , %%xmm7, %%xmm7 \n\t" + "vfmaddps 32(%3,%0,4), %%xmm0 , %%xmm9, %%xmm9 \n\t" + "vfmaddps 48(%3,%0,4), %%xmm0 , %%xmm11,%%xmm11 \n\t" + + "vfmaddps %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmaddps %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmaddps %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmaddps %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t" + + "vmovups %%xmm5 , (%3,%0,4) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,4) \n\t" + "vmovups %%xmm9 , 32(%3,%0,4) \n\t" + "vmovups %%xmm11, 48(%3,%0,4) \n\t" + + "addq $16, %0 \n\t" + "subq $8, %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "memory" + ); + } -