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

sgemm pipeline improved, zgemm rewritten without inner packs, ABI lxvx v20 fixed with vs52

tags/v0.3.7
AbdelRauf 7 years ago
parent
8fe794f059
commit
a469b32cf4
10 changed files with 2067 additions and 2421 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +1
    -1
      benchmark/gemm.c
  2. +1
    -1
      kernel/power/KERNEL.POWER9
  3. +24
    -24
      kernel/power/dgemm_kernel_power9.S
  4. +63
    -77
      kernel/power/sgemm_kernel_power9.S
  5. +116
    -76
      kernel/power/sgemm_logic_power9.S
  6. +295
    -566
      kernel/power/sgemm_macros_power9.S
  7. +52
    -64
      kernel/power/zgemm_kernel_power9.S
  8. +463
    -323
      kernel/power/zgemm_logic_power9.S
  9. +1048
    -1285
      kernel/power/zgemm_macros_power9.S
  10. +4
    -4
      param.h

+ 1
- 1
benchmark/gemm.c View File

@@ -207,7 +207,7 @@ int main(int argc, char *argv[]){
for (i = 0; i < m * n * COMPSIZE; i++) { for (i = 0; i < m * n * COMPSIZE; i++) {
c[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; c[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
} }
fprintf(stderr, " SIZE Flops Time\n"); fprintf(stderr, " SIZE Flops Time\n");


for (i = from; i <= to; i += step) { for (i = from; i <= to; i += step) {


+ 1
- 1
kernel/power/KERNEL.POWER9 View File

@@ -42,7 +42,7 @@ ZGEMMKERNEL = zgemm_kernel_power9.S
ZGEMMONCOPY = ../generic/zgemm_ncopy_2.c ZGEMMONCOPY = ../generic/zgemm_ncopy_2.c
ZGEMMOTCOPY = ../generic/zgemm_tcopy_2.c ZGEMMOTCOPY = ../generic/zgemm_tcopy_2.c
ZGEMMINCOPY = ../generic/zgemm_ncopy_8.c ZGEMMINCOPY = ../generic/zgemm_ncopy_8.c
ZGEMMITCOPY = zgemm_tcopy_8_power8.S
ZGEMMITCOPY = ../generic/zgemm_tcopy_8.c
ZGEMMONCOPYOBJ = zgemm_oncopy.o ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o ZGEMMOTCOPYOBJ = zgemm_otcopy.o
ZGEMMINCOPYOBJ = zgemm_incopy.o ZGEMMINCOPYOBJ = zgemm_incopy.o


+ 24
- 24
kernel/power/dgemm_kernel_power9.S View File

@@ -135,18 +135,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
std r14, 280(SP) std r14, 280(SP)
stxv v20, 288(SP)
stxv v21, 304(SP)
stxv v22, 320(SP)
stxv v23, 336(SP)
stxv v24, 352(SP)
stxv v25, 368(SP)
stxv v26, 384(SP)
stxv v27, 400(SP)
stxv v28, 416(SP)
stxv v29, 432(SP)
stxv v30, 448(SP)
stxv v31, 464(SP)
stxv vs52, 288(SP)
stxv vs53, 304(SP)
stxv vs54, 320(SP)
stxv vs55, 336(SP)
stxv vs56, 352(SP)
stxv vs57, 368(SP)
stxv vs58, 384(SP)
stxv vs59, 400(SP)
stxv vs60, 416(SP)
stxv vs61, 432(SP)
stxv vs62, 448(SP)
stxv vs63, 464(SP)
stfd f1, ALPHA_SP stfd f1, ALPHA_SP
@@ -229,18 +229,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld r15, 272(SP) ld r15, 272(SP)
ld r14, 280(SP) ld r14, 280(SP)
lxv v20, 288(SP)
lxv v21, 304(SP)
lxv v22, 320(SP)
lxv v23, 336(SP)
lxv v24, 352(SP)
lxv v25, 368(SP)
lxv v26, 384(SP)
lxv v27, 400(SP)
lxv v28, 416(SP)
lxv v29, 432(SP)
lxv v30, 448(SP)
lxv v31, 464(SP)
lxv vs52, 288(SP)
lxv vs53, 304(SP)
lxv vs54, 320(SP)
lxv vs55, 336(SP)
lxv vs56, 352(SP)
lxv vs57, 368(SP)
lxv vs58, 384(SP)
lxv vs59, 400(SP)
lxv vs60, 416(SP)
lxv vs61, 432(SP)
lxv vs62, 448(SP)
lxv vs63, 464(SP)
addi SP, SP, STACKSIZE addi SP, SP, STACKSIZE
blr blr


+ 63
- 77
kernel/power/sgemm_kernel_power9.S View File

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define LOAD ld #define LOAD ld
#define STACKSIZE (512 ) #define STACKSIZE (512 )
#define FLINK_SAVE (STACKSIZE+16) /* 16($r12) */
#define M r3 #define M r3
#define N r4 #define N r4
#define K r5 #define K r5
@@ -91,7 +91,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
PROFCODE PROFCODE
addi SP, SP, -STACKSIZE addi SP, SP, -STACKSIZE
li r0, 0
mflr r0
stfd f14, 0(SP) stfd f14, 0(SP)
stfd f15, 8(SP) stfd f15, 8(SP)
@@ -137,19 +138,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
std r14, 280(SP) std r14, 280(SP)
stxv v20, 288(SP)
stxv v21, 304(SP)
stxv v22, 320(SP)
stxv v23, 336(SP)
stxv v24, 352(SP)
stxv v25, 368(SP)
stxv v26, 384(SP)
stxv v27, 400(SP)
stxv v28, 416(SP)
stxv v29, 432(SP)
stxv v30, 448(SP)
stxv v31, 464(SP)
stxv vs52, 288(SP)
stxv vs53, 304(SP)
stxv vs54, 320(SP)
stxv vs55, 336(SP)
stxv vs56, 352(SP)
stxv vs57, 368(SP)
stxv vs58, 384(SP)
stxv vs59, 400(SP)
stxv vs60, 416(SP)
stxv vs61, 432(SP)
stxv vs62, 448(SP)
stxv vs63, 464(SP)
std r0, FLINK_SAVE(SP)
#if defined(TRMMKERNEL) #if defined(TRMMKERNEL)
@@ -157,72 +158,54 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif #endif
slwi LDC, LDC, 2 slwi LDC, LDC, 2
/* cmpwi cr0, M, 0
ble .L999_H1
cmpwi cr0, N, 0
ble .L999_H1
cmpwi cr0, K, 0
ble .L999_H1
*/
/*alpha is stored in f1. convert to single and splat*/ /*alpha is stored in f1. convert to single and splat*/
xscvdpspn alpha_r,vs1
xxspltw alpha_r,alpha_r,0
xscvdpspn alpha_r,vs1
xxspltw alpha_r,alpha_r,0
/*load reverse permute mask for big endian /*load reverse permute mask for big endian
uint128 = 0xc0d0e0f08090a0b0405060700010203 uint128 = 0xc0d0e0f08090a0b0405060700010203
*/ */
lis T2, perm_const2@highest lis T2, perm_const2@highest
ori T2, T2, perm_const2@higher
rldicr T2, T2, 32, 31
oris T2, T2, perm_const2@h
ori T2, T2, perm_const2@l
lis T1, perm_const1@highest lis T1, perm_const1@highest
lis T3, save_permute_12@highest
lis T4, save_permute_11@highest
lis T5, save_permute_22@highest
lis T6, save_permute_21@highest
ori T2, T2, perm_const2@higher
ori T1, T1, perm_const1@higher ori T1, T1, perm_const1@higher
ori T3, T3, save_permute_12@higher
ori T4, T4, save_permute_11@higher
ori T5, T5, save_permute_22@higher
ori T6, T6, save_permute_21@higher
rldicr T2, T2, 32, 31
rldicr T1, T1, 32, 31 rldicr T1, T1, 32, 31
rldicr T3, T3, 32, 31
rldicr T4, T4, 32, 31
rldicr T5, T5, 32, 31
rldicr T6, T6, 32, 31
oris T2, T2, perm_const2@h
oris T1, T1, perm_const1@h oris T1, T1, perm_const1@h
oris T3, T3, save_permute_12@h
oris T4, T4, save_permute_11@h
oris T5, T5, save_permute_22@h
oris T6, T6, save_permute_21@h
ori T2, T2, perm_const2@l
ori T1, T1, perm_const1@l ori T1, T1, perm_const1@l
ori T3, T3, save_permute_12@l
ori T4, T4, save_permute_11@l
ori T5, T5, save_permute_22@l
ori T6, T6, save_permute_21@l
li r0,0
mtvsrdd permute_mask,T2,T1 mtvsrdd permute_mask,T2,T1
lis T2, save_permute_12@highest
ori T2, T2, save_permute_12@higher
rldicr T2, T2, 32, 31
oris T2, T2, save_permute_12@h
ori T2, T2, save_permute_12@l
lis T1, save_permute_11@highest
ori T1, T1, save_permute_11@higher
rldicr T1, T1, 32, 31
oris T1, T1, save_permute_11@h
ori T1, T1, save_permute_11@l
mtvsrdd save_permute_1,T2,T1
lis T2, save_permute_22@highest
ori T2, T2, save_permute_22@higher
rldicr T2, T2, 32, 31
oris T2, T2, save_permute_22@h
ori T2, T2, save_permute_22@l
lis T1, save_permute_21@highest
ori T1, T1, save_permute_21@higher
rldicr T1, T1, 32, 31
oris T1, T1, save_permute_21@h
ori T1, T1, save_permute_21@l
mtvsrdd save_permute_2,T2,T1
mtvsrdd save_permute_1,T3,T4
mtvsrdd save_permute_2,T5,T6
#include "sgemm_logic_power9.S" #include "sgemm_logic_power9.S"
.L999:
addi r3, 0, 0
.L999:
lfd f14, 0(SP) lfd f14, 0(SP)
lfd f15, 8(SP) lfd f15, 8(SP)
lfd f16, 16(SP) lfd f16, 16(SP)
@@ -264,23 +247,26 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld r16, 264(SP) ld r16, 264(SP)
ld r15, 272(SP) ld r15, 272(SP)
ld r14, 280(SP) ld r14, 280(SP)
lxv v20, 288(SP)
lxv v21, 304(SP)
lxv v22, 320(SP)
lxv v23, 336(SP)
lxv v24, 352(SP)
lxv v25, 368(SP)
lxv v26, 384(SP)
lxv v27, 400(SP)
lxv v28, 416(SP)
lxv v29, 432(SP)
lxv v30, 448(SP)
lxv v31, 464(SP)
ld r0, FLINK_SAVE(SP)
addi SP, SP, STACKSIZE
lxv vs52, 288(SP)
lxv vs53, 304(SP)
lxv vs54, 320(SP)
lxv vs55, 336(SP)
lxv vs56, 352(SP)
lxv vs57, 368(SP)
lxv vs58, 384(SP)
lxv vs59, 400(SP)
mtlr r0
lxv vs60, 416(SP)
lxv vs61, 432(SP)
lxv vs62, 448(SP)
lxv vs63, 464(SP)
addi SP, SP, STACKSIZE
blr blr
EPILOGUE EPILOGUE
#endif #endif

+ 116
- 76
kernel/power/sgemm_logic_power9.S View File

@@ -1,5 +1,94 @@
#define MY_ALIGN .align 3 #define MY_ALIGN .align 3
b L8
MY_ALIGN
LSGEMM_L8x16_LMAIN_SUB:
LOAD8x16_0
mtctr L
MY_ALIGN
LSGEMM_L8x16_LOOP:
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_2 64,32, 3,0
KERNEL8x16_I1_L4_2 64,32, 4,0
KERNEL8x16_I1_L4_2 64,32, 5,0
KERNEL8x16_I1_L4_2 64,32, 6,0
KERNEL8x16_I1_L4_2 64,32, 7,0
KERNEL8x16_I1_L4_2 64,32, 8,0
KERNEL8x16_I1_L4_2 64,32, 9,0
KERNEL8x16_I1_L4_2 64,32, 10,0
KERNEL8x16_I1_L4_2 64,32, 11,0
KERNEL8x16_I1_L4_2 64,32, 12,0
KERNEL8x16_I1_L4_2 64,32, 13,0
KERNEL8x16_I1_L4_2 64,32, 14,0
KERNEL8x16_I1_L4_2 64,32, 15,0
KERNEL8x16_I1_L4_2 64,32, 16,0
KERNEL8x16_I1_L4_2 64,32, 17,0
KERNEL8x16_I1_L4_2 64,32, 18,0
KERNEL8x16_I1_L4_2 64,32, 19,0
KERNEL8x16_I1_L4_2 64,32, 20,0
KERNEL8x16_I1_L4_2 64,32, 21,0
KERNEL8x16_I1_L4_2 64,32, 22,0
KERNEL8x16_I1_L4_2 64,32, 23,0
KERNEL8x16_I1_L4_2 64,32, 24,0
KERNEL8x16_I1_L4_2 64,32, 25,0
KERNEL8x16_I1_L4_2 64,32, 26,0
KERNEL8x16_I1_L4_2 64,32, 27,0
KERNEL8x16_I1_L4_2 64,32, 28,0
KERNEL8x16_I1_L4_2 64,32, 29,0
KERNEL8x16_I1_L4_2 64,32, 30,0
KERNEL8x16_I1_L4_2 64,32, 31,1
bdnz LSGEMM_L8x16_LOOP
MY_ALIGN
LSGEMM_L8x16_LOOP_END:
END8x16 0, AO, BO, 64, 32
blr
MY_ALIGN
LSGEMM_L8x16_L64_SUB:
LOAD8x16_0
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_2 64,32, 3,0
KERNEL8x16_I1_L4_2 64,32, 4,0
KERNEL8x16_I1_L4_2 64,32, 5,0
KERNEL8x16_I1_L4_2 64,32, 6,0
KERNEL8x16_I1_L4_2 64,32, 7,0
KERNEL8x16_I1_L4_2 64,32, 8,0
KERNEL8x16_I1_L4_2 64,32, 9,0
KERNEL8x16_I1_L4_2 64,32, 10,0
KERNEL8x16_I1_L4_2 64,32, 11,0
KERNEL8x16_I1_L4_2 64,32, 12,0
KERNEL8x16_I1_L4_2 64,32, 13,0
KERNEL8x16_I1_L4_2 64,32, 14,0
KERNEL8x16_I1_L4_3 64,32, 15,1
blr
LSGEMM_L8x16_L32_SUB:
LOAD8x16_0
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_2 64,32, 3,0
KERNEL8x16_I1_L4_2 64,32, 4,0
KERNEL8x16_I1_L4_2 64,32, 5,0
KERNEL8x16_I1_L4_2 64,32, 6,0
KERNEL8x16_I1_L4_3 64,32, 7,1
blr
LSGEMM_L8x16_L16_SUB:
LOAD8x16_0
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_3 64,32, 3,1
blr
L8:
#if defined(TRMMKERNEL) && !defined(LEFT) #if defined(TRMMKERNEL) && !defined(LEFT)
neg TEMP_REG, OFFSET neg TEMP_REG, OFFSET
#endif #endif
@@ -39,98 +128,50 @@ LSGEMM_L8x16_BEGIN:
REFRESH_TEMP_BK T11,K,TEMP_REG,16,8 REFRESH_TEMP_BK T11,K,TEMP_REG,16,8
mr T12, T11 mr T12, T11
addi T12,T12, -1 addi T12,T12, -1
srawi. L, T12, 6 /**(T11-1) % 64x */
srawi. L, T12, 7 /**(T11-1) % 128x */
#else #else
mr T12, K mr T12, K
addi T12,T12, -1 addi T12,T12, -1
srawi. L, T12, 6 /**(K-1) % 64x */
srawi. L, T12, 7 /**(K-1) % 128x */
#endif #endif
ZERO8x16 ZERO8x16
ble LSGEMM_L8x16_SUB0 ble LSGEMM_L8x16_SUB0
MY_ALIGN
LSGEMM_L8x16_LOOP_START:
LOAD8x16_0 /*we already zeroed */
/*##OffsetA=64 OffsetB=32
#addi AO,AO,2112
#addi BO,BO,32 */
mtctr L
MY_ALIGN
LSGEMM_L8x16_LOOP:
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_2 64,32, 3,0
KERNEL8x16_I1_L4_2 64,32, 4,0
KERNEL8x16_I1_L4_2 64,32, 5,0
KERNEL8x16_I1_L4_2 64,32, 6,0
KERNEL8x16_I1_L4_2 64,32, 7,0
KERNEL8x16_I1_L4_2 64,32, 8,0
KERNEL8x16_I1_L4_2 64,32, 9,0
KERNEL8x16_I1_L4_2 64,32, 10,0
KERNEL8x16_I1_L4_2 64,32, 11,0
KERNEL8x16_I1_L4_2 64,32, 12,0
KERNEL8x16_I1_L4_2 64,32, 13,0
KERNEL8x16_I1_L4_2 64,32, 14,0
KERNEL8x16_I1_L4_2 64,32, 15,1
bdnz LSGEMM_L8x16_LOOP
MY_ALIGN
LSGEMM_L8x16_LOOP_END:
END8x16 0, AO, BO, 64, 32
b LSGEMM_L8x16_SUB1
bl LSGEMM_L8x16_LMAIN_SUB
andi. L, T12, 127
ble LSGEMM_L8x16_SAVE
b LSGEMM_L8x16_SUB2
MY_ALIGN MY_ALIGN
LSGEMM_L8x16_SUB0: LSGEMM_L8x16_SUB0:
#if defined(TRMMKERNEL) #if defined(TRMMKERNEL)
andi. L, T11, 127
andi. L, T11, 255
cmpwi T11,128
#else #else
andi. L, K, 127
andi. L, K, 255
cmpwi K,128
#endif #endif
b LSGEMM_L8x16_SUB2
MY_ALIGN
LSGEMM_L8x16_SUB1:
#if defined(TRMMKERNEL)
andi. L, T12, 63
#else
andi. L, T12, 63
#endif
ble LSGEMM_L8x16_SAVE
bne LSGEMM_L8x16_SUB2
MY_ALIGN
LSGEMM_L8x16_SUB2_128:
bl LSGEMM_L8x16_L64_SUB
bl LSGEMM_L8x16_L64_SUB
b LSGEMM_L8x16_SAVE
MY_ALIGN MY_ALIGN
LSGEMM_L8x16_SUB2: LSGEMM_L8x16_SUB2:
srawi. T10,L, 5
andi. T10,L,64
ble LSGEMM_L8x16_SUB2_32
bl LSGEMM_L8x16_L64_SUB
MY_ALIGN
LSGEMM_L8x16_SUB2_32:
andi. T10,L, 32
ble LSGEMM_L8x16_SUB2_16 ble LSGEMM_L8x16_SUB2_16
mtctr T10
MY_ALIGN
LSGEMM_L8x16_SUB2_LOOP:
LOAD8x16_0
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_2 64,32, 3,0
KERNEL8x16_I1_L4_2 64,32, 4,0
KERNEL8x16_I1_L4_2 64,32, 5,0
KERNEL8x16_I1_L4_2 64,32, 6,0
KERNEL8x16_I1_L4_3 64,32, 7,1
bdnz LSGEMM_L8x16_SUB2_LOOP
MY_ALIGN
bl LSGEMM_L8x16_L32_SUB
MY_ALIGN
LSGEMM_L8x16_SUB2_16: LSGEMM_L8x16_SUB2_16:
andi. T10,L, 16 andi. T10,L, 16
ble LSGEMM_L8x16_SUB2_8 ble LSGEMM_L8x16_SUB2_8
LOAD8x16_0
KERNEL8x16_I1_L4_2 64,32, 0,0
KERNEL8x16_I1_L4_2 64,32, 1,0
KERNEL8x16_I1_L4_2 64,32, 2,0
KERNEL8x16_I1_L4_3 64,32, 3,1
bl LSGEMM_L8x16_L16_SUB
MY_ALIGN MY_ALIGN
LSGEMM_L8x16_SUB2_8: LSGEMM_L8x16_SUB2_8:
andi. T10,L, 8 andi. T10,L, 8
@@ -155,8 +196,7 @@ LSGEMM_L8x16_SUB2_1:
andi. T10,L, 1 andi. T10,L, 1
ble LSGEMM_L8x16_SAVE ble LSGEMM_L8x16_SAVE
KERNEL8x16 0 KERNEL8x16 0
# addic. L, L, -1
# bgt LSGEMM_L8x16_SUB2
MY_ALIGN MY_ALIGN
LSGEMM_L8x16_SAVE: LSGEMM_L8x16_SAVE:


+ 295
- 566
kernel/power/sgemm_macros_power9.S
File diff suppressed because it is too large
View File


+ 52
- 64
kernel/power/zgemm_kernel_power9.S View File

@@ -30,10 +30,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define LOAD ld #define LOAD ld
#define STACKSIZE 32192
#define STACKSIZE 512
#define FZERO 312+192(SP) #define FZERO 312+192(SP)
#define FLINK_SAVE (STACKSIZE+16) /* 16($r12) */
#define M r3 #define M r3
#define N r4 #define N r4
@@ -56,20 +57,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define FRAMEPOINTER r12 #define FRAMEPOINTER r12
#define BBUFFER r14
#define T10 r14
#define L r15 #define L r15
#define ALPHA r16
#define T8 r16
#define T5 r17 #define T5 r17
#define T2 r19 #define T2 r19
#define BBO r20
#define o8 r21
#define T9 r20
#define T6 r21
#define I r22 #define I r22
#define J r23 #define J r23
#define AO r24 #define AO r24
#define BO r25 #define BO r25
#define CO r26 #define CO r26
#define o16 r27
#define T7 r27
#define T3 r28 #define T3 r28
#define T4 r29 #define T4 r29
@@ -82,12 +83,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
PROFCODE PROFCODE
mr FRAMEPOINTER, SP mr FRAMEPOINTER, SP
addi SP, SP, -STACKSIZE
addi SP, SP, -STACKSIZE
addi SP, SP, -STACKSIZE
addi SP, SP, -STACKSIZE
li r0, 0
addi SP, SP, -STACKSIZE
mflr r0
stfd f14, 0(SP) stfd f14, 0(SP)
stfd f15, 8(SP) stfd f15, 8(SP)
stfd f16, 16(SP) stfd f16, 16(SP)
@@ -111,6 +108,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
stfd f30, 128(SP) stfd f30, 128(SP)
stfd f31, 136(SP) stfd f31, 136(SP)
xxspltd alpha_r,vs1,0 /*copy from register f1 */
xxspltd alpha_i,vs2,0 /*copy from register f2 */
std r31, 144(SP) std r31, 144(SP)
std r30, 152(SP) std r30, 152(SP)
@@ -132,21 +131,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
std r14, 280(SP) std r14, 280(SP)
stxv v20, 288(SP)
stxv v21, 304(SP)
stxv v22, 320(SP)
stxv v23, 336(SP)
stxv v24, 352(SP)
stxv v25, 368(SP)
stxv v26, 384(SP)
stxv v27, 400(SP)
stxv v28, 416(SP)
stxv v29, 432(SP)
stxv v30, 448(SP)
stxv v31, 464(SP)
stxv vs52, 288(SP)
stxv vs53, 304(SP)
stxv vs54, 320(SP)
stxv vs55, 336(SP)
stxv vs56, 352(SP)
stxv vs57, 368(SP)
stxv vs58, 384(SP)
stxv vs59, 400(SP)
stxv vs60, 416(SP)
stxv vs61, 432(SP)
stxv vs62, 448(SP)
stxv vs63, 464(SP)
std r0, FLINK_SAVE(SP)
stw r0, FZERO
#ifdef linux #ifdef linux
ld LDC, FRAMESLOT(0) + 0(FRAMEPOINTER) ld LDC, FRAMESLOT(0) + 0(FRAMEPOINTER)
@@ -162,35 +161,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "zgemm_macros_power9.S" #include "zgemm_macros_power9.S"
cmpwi cr0, M, 0
ble L999
cmpwi cr0, N, 0
ble L999
cmpwi cr0, K, 0
ble L999
slwi LDC, LDC, ZBASE_SHIFT slwi LDC, LDC, ZBASE_SHIFT
li PRE, 512
li o8 , 8
li o16 , 16
addi BBUFFER, SP, 512+4096
li T1, -4096
and BBUFFER, BBUFFER, T1
addi ALPHA, SP, 296+192
li PRE, 512
li r0, 0
xxlor alpha_r,vs1,vs1 /*copy from register f1 */
xxlor alpha_i,vs2,vs2 /*copy from register f2 */
#if defined(CC) || defined(CR) || defined(RC) || defined(RR)
/*negate for this case as we will use addition -1*(a+b) */
xvnegdp alpha_r,alpha_r
xvnegdp alpha_i,alpha_i
#endif
.align 4 .align 4
#include "zgemm_logic_power9.S" #include "zgemm_logic_power9.S"
L999: L999:
addi r3, 0, 0
lfd f14, 0(SP) lfd f14, 0(SP)
lfd f15, 8(SP) lfd f15, 8(SP)
lfd f16, 16(SP) lfd f16, 16(SP)
@@ -233,24 +221,24 @@ L999:
ld r16, 264(SP) ld r16, 264(SP)
ld r15, 272(SP) ld r15, 272(SP)
ld r14, 280(SP) ld r14, 280(SP)
ld r0, FLINK_SAVE(SP)
lxv v20, 288(SP)
lxv v21, 304(SP)
lxv v22, 320(SP)
lxv v23, 336(SP)
lxv v24, 352(SP)
lxv v25, 368(SP)
lxv v26, 384(SP)
lxv v27, 400(SP)
lxv v28, 416(SP)
lxv v29, 432(SP)
lxv v30, 448(SP)
lxv v31, 464(SP)
addi SP, SP, STACKSIZE
addi SP, SP, STACKSIZE
addi SP, SP, STACKSIZE
addi SP, SP, STACKSIZE
lxv vs52, 288(SP)
lxv vs53, 304(SP)
lxv vs54, 320(SP)
lxv vs55, 336(SP)
lxv vs56, 352(SP)
lxv vs57, 368(SP)
lxv vs58, 384(SP)
lxv vs59, 400(SP)
mtlr r0
lxv vs60, 416(SP)
lxv vs61, 432(SP)
lxv vs62, 448(SP)
lxv vs63, 464(SP)
addi SP, SP, STACKSIZE
blr blr
EPILOGUE EPILOGUE

+ 463
- 323
kernel/power/zgemm_logic_power9.S
File diff suppressed because it is too large
View File


+ 1048
- 1285
kernel/power/zgemm_macros_power9.S
File diff suppressed because it is too large
View File


+ 4
- 4
param.h View File

@@ -2248,15 +2248,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ZGEMM_DEFAULT_UNROLL_M 8 #define ZGEMM_DEFAULT_UNROLL_M 8
#define ZGEMM_DEFAULT_UNROLL_N 2 #define ZGEMM_DEFAULT_UNROLL_N 2


#define SGEMM_DEFAULT_P 640
#define SGEMM_DEFAULT_P 832
#define DGEMM_DEFAULT_P 128 #define DGEMM_DEFAULT_P 128
#define CGEMM_DEFAULT_P 640 #define CGEMM_DEFAULT_P 640
#define ZGEMM_DEFAULT_P 512
#define ZGEMM_DEFAULT_P 256


#define SGEMM_DEFAULT_Q 1408
#define SGEMM_DEFAULT_Q 1025
#define DGEMM_DEFAULT_Q 384 #define DGEMM_DEFAULT_Q 384
#define CGEMM_DEFAULT_Q 640 #define CGEMM_DEFAULT_Q 640
#define ZGEMM_DEFAULT_Q 1152
#define ZGEMM_DEFAULT_Q 1025


#define SYMV_P 8 #define SYMV_P 8




Write Preview
Loading…
Cancel
Save