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

Fast Haswell ZGEMM kernel

tags/v0.3.8^2
wjc404 GitHub 6 years ago
parent
31d6c2eb7d
commit
d573d24de7
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 240 additions and 0 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +240
    -0
      kernel/x86_64/zgemm_kernel_4x2_haswell.c

+ 240
- 0
kernel/x86_64/zgemm_kernel_4x2_haswell.c View File

@@ -0,0 +1,240 @@
#include "common.h"
#include <stdint.h>

/* recommended settings: GEMM_P = 192, GEMM_Q = 192 */

#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
#define A_CONJ 0
#define B_CONJ 0
#endif
#if defined(RN) || defined(RT) || defined(CN) || defined(CT)
#define A_CONJ 1
#define B_CONJ 0
#endif
#if defined(NR) || defined(NC) || defined(TR) || defined(TC)
#define A_CONJ 0
#define B_CONJ 1
#endif
#if defined(RR) || defined(RC) || defined(CR) || defined(CC)
#define A_CONJ 1
#define B_CONJ 1
#endif

/* %0 = a_ptr, %1 = b_ptr, %2 = c_ptr, %3 = c_tmp, %4 = ldc(bytes), %5 = k_counter, %6 = &alpha, %7 = m_counter, %8 = b_pref */
/* r11 = m, r12 = k << 5, r13 = k, r14 = b_head, r15 = temp */

/* m=4, ymm 0-3 temp, ymm 4-15 acc */
#if A_CONJ == B_CONJ
#define acc_m2n1_exp(ar,ai,b2,cl,cr) "vfmadd231pd %%ymm"#ar",%%ymm"#b2",%%ymm"#cl"; vfmadd231pd %%ymm"#ai",%%ymm"#b2",%%ymm"#cr";"
#define acc_m4n1_con(ua,la,b1,uc,lc) "vfmaddsub231pd %%ymm"#ua",%%ymm"#b1",%%ymm"#uc"; vfmaddsub231pd %%ymm"#la",%%ymm"#b1",%%ymm"#lc";"
#else
#define acc_m2n1_exp(ar,ai,b2,cl,cr) "vfmadd231pd %%ymm"#ar",%%ymm"#b2",%%ymm"#cl"; vfnmadd231pd %%ymm"#ai",%%ymm"#b2",%%ymm"#cr";"
#define acc_m4n1_con(ua,la,b1,uc,lc) "vfmsubadd231pd %%ymm"#ua",%%ymm"#b1",%%ymm"#uc"; vfmsubadd231pd %%ymm"#la",%%ymm"#b1",%%ymm"#lc";"
#endif
/* expanded accumulators for m4n1 and m4n2 */
#define KERNEL_k1m4n1 \
"vbroadcastf128 (%1),%%ymm0; addq $16,%1;"\
"vmovddup (%0),%%ymm1; vmovddup 8(%0),%%ymm2;" acc_m2n1_exp(1,2,0,4,5)\
"vmovddup 32(%0),%%ymm1; vmovddup 40(%0),%%ymm2;" acc_m2n1_exp(1,2,0,6,7)\
"addq $64,%0;"
#define KERNEL_k1m4n2 \
"vbroadcastf128 (%1),%%ymm0; vbroadcastf128 16(%1),%%ymm1; addq $32,%1;"\
"vmovddup (%0),%%ymm2; vmovddup 8(%0),%%ymm3;" acc_m2n1_exp(2,3,0,4,5) acc_m2n1_exp(2,3,1,8,9)\
"vmovddup 32(%0),%%ymm2; vmovddup 40(%0),%%ymm3;" acc_m2n1_exp(2,3,0,6,7) acc_m2n1_exp(2,3,1,10,11)\
"addq $64,%0;"
/* contracted accumulators for m4n4 and m4n6 */
#define acc_m4n2_con(ua,la,luc,llc,ruc,rlc,lboff,rboff,...) \
"vbroadcastsd "#lboff"("#__VA_ARGS__"),%%ymm2;" acc_m4n1_con(ua,la,2,luc,llc)\
"vbroadcastsd "#rboff"("#__VA_ARGS__"),%%ymm3;" acc_m4n1_con(ua,la,3,ruc,rlc)
#define KERNEL_1_k1m4n4 \
"vmovupd (%0),%%ymm0; vmovupd 32(%0),%%ymm1; prefetcht0 512(%0); addq $64,%0;"\
acc_m4n2_con(0,1,4,5,6,7,0,16,%1) acc_m4n2_con(0,1,8,9,10,11,0,16,%1,%%r12,1)
#define KERNEL_2_k1m4n4 \
"vpermilpd $5,%%ymm0,%%ymm0; vpermilpd $5,%%ymm1,%%ymm1;"\
acc_m4n2_con(0,1,4,5,6,7,8,24,%1) acc_m4n2_con(0,1,8,9,10,11,8,24,%1,%%r12,1)
#define KERNEL_1_k1m4n6 KERNEL_1_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,0,16,%1,%%r12,2)
#define KERNEL_2_k1m4n6 KERNEL_2_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,8,24,%1,%%r12,2)
#define KERNEL_k1m4n4 KERNEL_1_k1m4n4 KERNEL_2_k1m4n4 "addq $32,%1;"
#define KERNEL_k1m4n6 KERNEL_1_k1m4n6 KERNEL_2_k1m4n6 "addq $32,%1;"
#define zero_4ymm(no1,no2,no3,no4) \
"vpxor %%ymm"#no1",%%ymm"#no1",%%ymm"#no1"; vpxor %%ymm"#no2",%%ymm"#no2",%%ymm"#no2";"\
"vpxor %%ymm"#no3",%%ymm"#no3",%%ymm"#no3"; vpxor %%ymm"#no4",%%ymm"#no4",%%ymm"#no4";"
/* initialization and storage macros */
#define INIT_m4n1 zero_4ymm(4,5,6,7)
#define INIT_m4n2 zero_4ymm(4,5,6,7) zero_4ymm(8,9,10,11)
#define INIT_m4n4 zero_4ymm(4,5,6,7) zero_4ymm(8,9,10,11)
#define INIT_m4n6 INIT_m4n4 zero_4ymm(12,13,14,15)
#if A_CONJ == B_CONJ
#define cont_expacc(cl,cr,dst) "vpermilpd $5,%%ymm"#cr",%%ymm"#cr"; vaddsubpd %%ymm"#cl",%%ymm"#cr",%%ymm"#dst";"
#else
#define cont_expacc(cl,cr,dst) "vpermilpd $5,%%ymm"#cr",%%ymm"#cr"; vaddsubpd %%ymm"#cr",%%ymm"#cl",%%ymm"#dst";"
#endif
#if A_CONJ == 0
#define save_1ymm(c,tmp,off,alpr,alpi,...) \
"vpermilpd $5,%%ymm"#c",%%ymm"#tmp"; vfmsubadd213pd "#off"("#__VA_ARGS__"),%%ymm"#alpr",%%ymm"#c";"\
"vfmsubadd231pd %%ymm"#tmp",%%ymm"#alpi",%%ymm"#c"; vmovupd %%ymm"#c","#off"("#__VA_ARGS__");"
#else
#define save_1ymm(c,tmp,off,alpr,alpi,...) \
"vpermilpd $5,%%ymm"#c",%%ymm"#tmp"; vfmaddsub213pd "#off"("#__VA_ARGS__"),%%ymm"#alpi",%%ymm"#tmp";"\
"vfmaddsub231pd %%ymm"#c",%%ymm"#alpr",%%ymm"#tmp"; vmovupd %%ymm"#tmp","#off"("#__VA_ARGS__");"
#endif
#define save_init_m4 "movq %2,%3; addq $64,%2; vbroadcastsd (%6),%%ymm0; vbroadcastsd 8(%6),%%ymm1;"
#define SAVE_m4n1 save_init_m4 cont_expacc(4,5,4) cont_expacc(6,7,6) save_1ymm(4,2,0,0,1,%3) save_1ymm(6,3,32,0,1,%3)
#define SAVE_m4n2 SAVE_m4n1\
cont_expacc(8,9,8) cont_expacc(10,11,10) save_1ymm(8,2,0,0,1,%3,%4,1) save_1ymm(10,3,32,0,1,%3,%4,1)
#define SAVE_m4n4 save_init_m4\
save_1ymm(4,2,0,0,1,%3) save_1ymm(5,3,32,0,1,%3) save_1ymm(6,2,0,0,1,%3,%4,1) save_1ymm(7,3,32,0,1,%3,%4,1) "leaq (%3,%4,2),%3;"\
save_1ymm(8,2,0,0,1,%3) save_1ymm(9,3,32,0,1,%3) save_1ymm(10,2,0,0,1,%3,%4,1) save_1ymm(11,3,32,0,1,%3,%4,1)
#define SAVE_m4n6 SAVE_m4n4 "leaq (%3,%4,2),%3;"\
save_1ymm(12,2,0,0,1,%3) save_1ymm(13,3,32,0,1,%3) save_1ymm(14,2,0,0,1,%3,%4,1) save_1ymm(15,3,32,0,1,%3,%4,1)
#define COMPUTE_m4(ndim) \
"movq %%r14,%1;" INIT_m4n##ndim "movq %2,%3; movq %%r13,%5;"\
"testq %5,%5; jz "#ndim"4443f; cmpq $10,%5; jb "#ndim"4442f;"\
"movq $10,%5; movq $84,%%r15;"\
#ndim"4441:\n\t"\
"prefetcht1 (%3); subq $63,%3; addq %%r15,%3;"\
"prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
"testq $12,%5; movq $84,%%r15; cmovz %4,%%r15; prefetcht1 (%8); addq $16,%8;"\
"prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
"addq $4,%5; cmpq %5,%%r13; jnb "#ndim"4441b;"\
"movq %2,%3; negq %5; leaq 10(%%r13,%5,1),%5; prefetcht0 (%6); prefetcht0 15(%6);"\
#ndim"4442:\n\t"\
"prefetcht0 (%3); prefetcht0 63(%3); addq %4,%3;"\
KERNEL_k1m4n##ndim "decq %5; jnz "#ndim"4442b;"\
#ndim"4443:\n\t"\
"prefetcht0 (%%r14); prefetcht0 64(%%r14);" SAVE_m4n##ndim

/* m=2, ymm 0-3 temp, ymm 4-15 acc, expanded accumulators */
#define KERNEL_k1m2n1 \
"vmovddup (%0),%%ymm1; vmovddup 8(%0),%%ymm2; addq $32,%0;"\
"vbroadcastf128 (%1),%%ymm0;" acc_m2n1_exp(1,2,0,4,5) "addq $16,%1;"
#define acc_m2n2_exp(c1l,c1r,c2l,c2r,...) \
"vbroadcastf128 ("#__VA_ARGS__"),%%ymm2;" acc_m2n1_exp(0,1,2,c1l,c1r)\
"vbroadcastf128 16("#__VA_ARGS__"),%%ymm3;" acc_m2n1_exp(0,1,3,c2l,c2r)
#define KERNEL_h_k1m2n2 \
"vmovddup (%0),%%ymm0; vmovddup 8(%0),%%ymm1; addq $32,%0;" acc_m2n2_exp(4,5,6,7,%1)
#define KERNEL_h_k1m2n4 KERNEL_h_k1m2n2 acc_m2n2_exp(8,9,10,11,%1,%%r12,1)
#define KERNEL_h_k1m2n6 KERNEL_h_k1m2n4 acc_m2n2_exp(12,13,14,15,%1,%%r12,2)
#define KERNEL_k1m2n2 KERNEL_h_k1m2n2 "addq $32,%1;"
#define KERNEL_k1m2n4 KERNEL_h_k1m2n4 "addq $32,%1;"
#define KERNEL_k1m2n6 KERNEL_h_k1m2n6 "addq $32,%1;"
#define INIT_m2n1 "vpxor %%ymm4,%%ymm4,%%ymm4; vpxor %%ymm5,%%ymm5,%%ymm5;"
#define INIT_m2n2 zero_4ymm(4,5,6,7)
#define INIT_m2n4 INIT_m2n2 zero_4ymm(8,9,10,11)
#define INIT_m2n6 INIT_m2n4 zero_4ymm(12,13,14,15)
#define save_init_m2 "movq %2,%3; addq $32,%2; vbroadcastsd (%6),%%ymm0; vbroadcastsd 8(%6),%%ymm1;"
#define SAVE_m2n1 save_init_m2 cont_expacc(4,5,4) save_1ymm(4,2,0,0,1,%3)
#define SAVE_m2n2 SAVE_m2n1 cont_expacc(6,7,6) save_1ymm(6,3,0,0,1,%3,%4,1)
#define SAVE_m2n4 SAVE_m2n2 "leaq (%3,%4,2),%3;"\
cont_expacc(8,9,8) cont_expacc(10,11,10) save_1ymm(8,2,0,0,1,%3) save_1ymm(10,3,0,0,1,%3,%4,1)
#define SAVE_m2n6 SAVE_m2n4 "leaq (%3,%4,2),%3;"\
cont_expacc(12,13,12) cont_expacc(14,15,14) save_1ymm(12,2,0,0,1,%3) save_1ymm(14,3,0,0,1,%3,%4,1)
#define COMPUTE_m2(ndim) \
"movq %%r14,%1;" INIT_m2n##ndim "movq %%r13,%5;"\
"testq %5,%5; jz "#ndim"2222f;"\
#ndim"2221:\n\t"\
KERNEL_k1m2n##ndim\
"decq %5; jnz "#ndim"2221b;"\
#ndim"2222:\n\t"\
SAVE_m2n##ndim

/* m=1, vmm 0-3 temp, vmm 4-15 acc, expanded accumulators */
#if A_CONJ == B_CONJ
#define acc_m1n1_exp(ar,ai,b2,cl,cr) "vfmadd231pd %%xmm"#ar",%%xmm"#b2",%%xmm"#cl"; vfmadd231pd %%xmm"#ai",%%xmm"#b2",%%xmm"#cr";"
#define acc_m1n2_exp(arb,aib,b4,cl,cr) "vfmadd231pd %%ymm"#arb",%%ymm"#b4",%%ymm"#cl"; vfmadd231pd %%ymm"#aib",%%ymm"#b4",%%ymm"#cr";"
#else
#define acc_m1n1_exp(ar,ai,b2,cl,cr) "vfmadd231pd %%xmm"#ar",%%xmm"#b2",%%xmm"#cl"; vfnmadd231pd %%xmm"#ai",%%xmm"#b2",%%xmm"#cr";"
#define acc_m1n2_exp(arb,aib,b4,cl,cr) "vfmadd231pd %%ymm"#arb",%%ymm"#b4",%%ymm"#cl"; vfnmadd231pd %%ymm"#aib",%%ymm"#b4",%%ymm"#cr";"
#endif
#define KERNEL_k1m1n1 \
"vmovddup (%0),%%xmm0; vmovddup 8(%0),%%xmm1; addq $16,%0;"\
"vmovupd (%1),%%xmm2; addq $16,%1;" acc_m1n1_exp(0,1,2,4,5)
#define KERNEL_h_k1m1n2 \
"vbroadcastsd (%0),%%ymm0; vbroadcastsd 8(%0),%%ymm1; addq $16,%0;"\
"vmovupd (%1),%%ymm2;" acc_m1n2_exp(0,1,2,4,5)
#define KERNEL_h_k1m1n4 KERNEL_h_k1m1n2 "vmovupd (%1,%%r12,1),%%ymm2;" acc_m1n2_exp(0,1,2,6,7)
#define KERNEL_h_k1m1n6 KERNEL_h_k1m1n4 "vmovupd (%1,%%r12,2),%%ymm2;" acc_m1n2_exp(0,1,2,8,9)
#define KERNEL_k1m1n2 KERNEL_h_k1m1n2 "addq $32,%1;"
#define KERNEL_k1m1n4 KERNEL_h_k1m1n4 "addq $32,%1;"
#define KERNEL_k1m1n6 KERNEL_h_k1m1n6 "addq $32,%1;"
#define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4; vpxor %%xmm5,%%xmm5,%%xmm5;"
#define INIT_m1n2 "vpxor %%ymm4,%%ymm4,%%ymm4; vpxor %%ymm5,%%ymm5,%%ymm5;"
#define INIT_m1n4 INIT_m1n2 "vpxor %%ymm6,%%ymm6,%%ymm6; vpxor %%ymm7,%%ymm7,%%ymm7;"
#define INIT_m1n6 INIT_m1n4 "vpxor %%ymm8,%%ymm8,%%ymm8; vpxor %%ymm9,%%ymm9,%%ymm9;"
#if A_CONJ == B_CONJ
#define cont_expxmmacc(cl,cr,dst) "vpermilpd $5,%%xmm"#cr",%%xmm"#cr"; vaddsubpd %%xmm"#cl",%%xmm"#cr",%%xmm"#dst";"
#else
#define cont_expxmmacc(cl,cr,dst) "vpermilpd $5,%%xmm"#cr",%%xmm"#cr"; vaddsubpd %%xmm"#cr",%%xmm"#cl",%%xmm"#dst";"
#endif
#if A_CONJ == 0
#define save_m1n1(c,tmp,alpr,alpi) \
"vpermilpd $5,%%xmm"#c",%%xmm"#tmp"; vfmsubadd213pd (%3),%%xmm"#alpr",%%xmm"#c";"\
"vfmsubadd231pd %%xmm"#tmp",%%xmm"#alpi",%%xmm"#c"; vmovupd %%xmm"#c",(%3);"
#define save_m1n2(c,tmp1,tmp2,alpr,alpi) \
"vpermilpd $5,%%ymm"#c",%%ymm"#tmp1"; vmovupd (%3),%%xmm"#tmp2"; vinsertf128 $1,(%3,%4,1),%%ymm"#tmp2",%%ymm"#tmp2";"\
"vfmsubadd213pd %%ymm"#tmp2",%%ymm"#alpr",%%ymm"#c"; vfmsubadd231pd %%ymm"#tmp1",%%ymm"#alpi",%%ymm"#c";"\
"vmovupd %%xmm"#c",(%3); vextractf128 $1,%%ymm"#c",(%3,%4,1); leaq (%3,%4,2),%3;"
#else
#define save_m1n1(c,tmp,alpr,alpi) \
"vpermilpd $5,%%xmm"#c",%%xmm"#tmp"; vfmaddsub213pd (%3),%%xmm"#alpi",%%xmm"#tmp";"\
"vfmaddsub231pd %%xmm"#c",%%xmm"#alpr",%%xmm"#tmp"; vmovupd %%xmm"#tmp",(%3);"
#define save_m1n2(c,tmp1,tmp2,alpr,alpi) \
"vpermilpd $5,%%ymm"#c",%%ymm"#tmp1"; vmovupd (%3),%%xmm"#tmp2"; vinsertf128 $1,(%3,%4,1),%%ymm"#tmp2",%%ymm"#tmp2";"\
"vfmaddsub213pd %%ymm"#tmp2",%%ymm"#alpi",%%ymm"#tmp1"; vfmaddsub231pd %%ymm"#c",%%ymm"#alpr",%%ymm"#tmp1";"\
"vmovupd %%xmm"#tmp1",(%3); vextractf128 $1,%%ymm"#tmp1",(%3,%4,1); leaq (%3,%4,2),%3;"
#endif
#define save_init_m1 "movq %2,%3; addq $16,%2; vbroadcastsd (%6),%%ymm0; vbroadcastsd 8(%6),%%ymm1;"
#define SAVE_m1n1 save_init_m1 cont_expxmmacc(4,5,4) save_m1n1(4,2,0,1)
#define SAVE_m1n2 save_init_m1 cont_expacc(4,5,4) save_m1n2(4,2,3,0,1)
#define SAVE_m1n4 SAVE_m1n2 cont_expacc(6,7,6) save_m1n2(6,2,3,0,1)
#define SAVE_m1n6 SAVE_m1n4 cont_expacc(8,9,8) save_m1n2(8,2,3,0,1)
#define COMPUTE_m1(ndim) \
"movq %%r14,%1;" INIT_m1n##ndim "movq %%r13,%5;"\
"testq %5,%5; jz "#ndim"1112f;"\
#ndim"1111:\n\t"\
KERNEL_k1m1n##ndim\
"decq %5; jnz "#ndim"1111b;"\
#ndim"1112:\n\t"\
SAVE_m1n##ndim

#define COMPUTE(ndim) {\
b_pref = b_ptr + ndim * K *2;\
__asm__ __volatile__ (\
"movq %1,%%r14; movq %5,%%r13; movq %5,%%r12; salq $5,%%r12; movq %7,%%r11;"\
"cmpq $4,%7; jb "#ndim"9992f;"\
#ndim"9991:\n\t"\
COMPUTE_m4(ndim)\
"subq $4,%7; cmpq $4,%7; jnb "#ndim"9991b;"\
#ndim"9992:\n\t"\
"cmpq $2,%7; jb "#ndim"9993f;"\
COMPUTE_m2(ndim) "subq $2,%7;"\
#ndim"9993:\n\t"\
"testq %7,%7; jz "#ndim"9994f;"\
COMPUTE_m1(ndim)\
#ndim"9994:\n\t"\
"movq %%r14,%1; movq %%r13,%5; movq %%r11,%7; vzeroupper;"\
:"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_in_bytes),"+r"(K),"+r"(alp),"+r"(M),"+r"(b_pref)\
::"cc","memory","r11","r12","r13","r14","r15","xmm0","xmm1","xmm2","xmm3","xmm4","xmm5",\
"xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15");\
a_ptr -= M * K *2; b_ptr += ndim * K *2; c_ptr += (ndim * LDC - M) * 2;\
}

int __attribute__ ((noinline))
CNAME(BLASLONG m, BLASLONG n, BLASLONG k, double alphar, double alphai, double * __restrict__ A, double * __restrict__ B, double * __restrict__ C, BLASLONG LDC)
{
if(m==0||n==0||k==0||(alphar==0.0 && alphai==0.0)) return 0;
int64_t ldc_in_bytes = (int64_t)LDC * sizeof(double) * 2;
#if A_CONJ == B_CONJ
double const_val[2] = {-alphar, -alphai};
#else
double const_val[2] = {alphar, alphai};
#endif
int64_t M = (int64_t)m, K = (int64_t)k;
BLASLONG n_count = n;
double *a_ptr = A,*b_ptr = B,*c_ptr = C,*c_tmp = C,*alp = const_val,*b_pref = B;
for(;n_count>5;n_count-=6) COMPUTE(6)
for(;n_count>3;n_count-=4) COMPUTE(4)
for(;n_count>1;n_count-=2) COMPUTE(2)
if(n_count>0) COMPUTE(1)
return 0;
}

Write Preview
Loading…
Cancel
Save