OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

optimized cgemv_t kernel for haswell

tags/v0.2.12^2
wernsaar 11 years ago
parent
53b5726b04
commit
8f100a14f2
2 changed files with 287 additions and 1 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +1
    -1
      kernel/x86_64/KERNEL.HASWELL
  2. +286
    -0
      kernel/x86_64/cgemv_t_microk_haswell-4.c

+ 1
- 1
kernel/x86_64/KERNEL.HASWELL View File

@@ -8,7 +8,7 @@ ZGEMVNKERNEL = zgemv_n_4.c
ZGEMVTKERNEL = zgemv_t_4.c

CGEMVNKERNEL = cgemv_n_4.c
CGEMVTKERNEL = cgemv_t.c
CGEMVTKERNEL = cgemv_t_4.c

SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c


+ 286
- 0
kernel/x86_64/cgemv_t_microk_haswell-4.c View File

@@ -251,3 +251,289 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
}


#define HAVE_KERNEL_4x2 1
static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));

static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp
"vxorps %%ymm10, %%ymm10, %%ymm10 \n\t" // temp
"vxorps %%ymm11, %%ymm11, %%ymm11 \n\t" // temp

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1

"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts

"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"

".L08LABEL%=: \n\t"
"cmpq $0, %1 \n\t"
"je .L08END%= \n\t"

".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"prefetcht0 192(%5,%0,4) \n\t"
"vmovups (%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1

"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1

"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0
"vmovups 32(%5,%0,4), %%ymm5 \n\t" // 4 complex values from a1

"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts

"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"vfmadd231ps %%ymm5 , %%ymm0, %%ymm10 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm5 , %%ymm1, %%ymm11 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1

"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"

".L08END%=: \n\t"

"vbroadcastss (%6) , %%xmm0 \n\t" // value from alpha
"vbroadcastss 4(%6) , %%xmm1 \n\t" // value from alpha


#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vpermilps $0xb1 , %%ymm11, %%ymm11 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
"vaddsubps %%ymm11, %%ymm10, %%ymm10 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vaddsubps %%ymm10, %%ymm11, %%ymm10 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm10, %%ymm10 \n\t"
#endif

"vmovsd (%3), %%xmm4 \n\t" // read y
"vmovsd 8(%3), %%xmm5 \n\t"

"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"
"vextractf128 $1, %%ymm10, %%xmm11 \n\t"

"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"

"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"
"vshufpd $0x1, %%xmm10, %%xmm10, %%xmm11 \n\t"

"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm11, %%xmm10 \n\t"

"vmulps %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulps %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r

#if !defined(XCONJ)
"vpermilps $0xb1 , %%xmm9 , %%xmm9 \n\t"
"vpermilps $0xb1 , %%xmm11, %%xmm11 \n\t"
"vaddsubps %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubps %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
"vaddsubps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubps %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm10, %%xmm10 \n\t"
#endif


"vaddps %%xmm8 , %%xmm4 , %%xmm8 \n\t"
"vaddps %%xmm10, %%xmm5 , %%xmm10 \n\t"

"vmovsd %%xmm8 , (%3) \n\t"
"vmovsd %%xmm10, 8(%3) \n\t"

"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (alpha) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}


#define HAVE_KERNEL_4x1 1
static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));

static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"

"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // temp
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // temp

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0

"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts

"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1
"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"

".L08LABEL%=: \n\t"
"cmpq $0, %1 \n\t"
"je .L08END%= \n\t"

".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmovups (%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0

"prefetcht0 192(%2,%0,4) \n\t"
"vmovups (%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts
"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1

"vmovups 32(%4,%0,4), %%ymm4 \n\t" // 4 complex values from a0

"vmovups 32(%2,%0,4) , %%ymm6 \n\t" // 4 complex values from x
"vpermilps $0xb1, %%ymm6, %%ymm7 \n\t" // exchange real and imap parts
"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0 \n\t" // only the real parts
"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1 \n\t" // only the imag parts

"vfmadd231ps %%ymm4 , %%ymm0, %%ymm8 \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1
"vfmadd231ps %%ymm4 , %%ymm1, %%ymm9 \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1

"addq $16 , %0 \n\t"
"subq $8 , %1 \n\t"
"jnz .L01LOOP%= \n\t"

".L08END%=: \n\t"

"vbroadcastss (%5) , %%xmm0 \n\t" // value from alpha
"vbroadcastss 4(%5) , %%xmm1 \n\t" // value from alpha


#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilps $0xb1 , %%ymm9 , %%ymm9 \n\t"
"vaddsubps %%ymm9 , %%ymm8, %%ymm8 \n\t"
#else
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
"vaddsubps %%ymm8 , %%ymm9 , %%ymm8 \n\t"
"vpermilps $0xb1 , %%ymm8 , %%ymm8 \n\t"
#endif

"vmovsd (%3), %%xmm4 \n\t" // read y

"vextractf128 $1, %%ymm8 , %%xmm9 \n\t"

"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"

"vshufpd $0x1, %%xmm8 , %%xmm8 , %%xmm9 \n\t"

"vaddps %%xmm8 , %%xmm9 , %%xmm8 \n\t"

"vmulps %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulps %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r

#if !defined(XCONJ)
"vpermilps $0xb1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubps %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubps %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilps $0xb1 , %%xmm8 , %%xmm8 \n\t"
#endif


"vaddps %%xmm8 , %%xmm4 , %%xmm8 \n\t"

"vmovsd %%xmm8 , (%3) \n\t"

"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



Write Preview
Loading…
Cancel
Save