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

minor improvements and code cleanup

tags/v0.2.7
wernsaar 13 years ago
parent
7641f6e253
commit
6821677489
1 changed files with 87 additions and 172 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +87
    -172
      kernel/x86_64/dgemm_kernel_4x4_bulldozer.S

+ 87
- 172
kernel/x86_64/dgemm_kernel_4x4_bulldozer.S View File

@@ -56,7 +56,7 @@
* 53 GFLOPS with 4 threads on 2 modules
* 46 GFLOPS with 2 threads on 2 modules
* 28 GFLOPS with 2 threads on 1 module
* 23,6 GFLOPS with 1 thread on 1 module
* 23,1 GFLOPS with 1 thread on 1 module
*********************************************************************/
#define ASSEMBLER
@@ -114,132 +114,132 @@
#define B_PR1 512
#define KERNEL1(xx) \
vmovddup -15 * SIZE(BO, %rax, 4), %xmm3 ;\
vmovups -14 * SIZE(AO, %rax, 4),%xmm2 ;\
vfmaddpd %xmm8,%xmm6,%xmm7,%xmm8 ;\
vfmaddpd %xmm12,%xmm2,%xmm7,%xmm12 ;\
vfmaddpd %xmm8,%xmm6,%xmm7,%xmm8 ;\
vmovddup -14 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm12,%xmm2,%xmm7,%xmm12 ;\
vmovddup -13 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm9,%xmm6,%xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2,%xmm3,%xmm13 ;\
vmovddup -8 * SIZE(BO, %rax, 4), %xmm7 ;\
vfmaddpd %xmm10,%xmm6,%xmm5,%xmm10 ;\
vfmaddpd %xmm11,%xmm6,%xmm4,%xmm11 ;\
vfmaddpd %xmm9,%xmm6,%xmm3,%xmm9 ;\
vmovddup -12 * SIZE(BO, %rax, 4), %xmm1 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovups -8 * SIZE(AO, %rax, 4),%xmm6 ;\
vfmaddpd %xmm13,%xmm2,%xmm3,%xmm13 ;\
vmovddup -8 * SIZE(BO, %rax, 4), %xmm7 ;\
vfmaddpd %xmm10,%xmm6,%xmm5,%xmm10 ;\
vmovddup -11 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm6,%xmm4,%xmm11 ;\
vmovups -12 * SIZE(AO, %rax, 4),%xmm0 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
vmovups -8 * SIZE(AO, %rax, 4),%xmm6 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
#define KERNEL2(xx) \
vmovups -12 * SIZE(AO, %rax, 4), %xmm2 ;\
vmovups -10 * SIZE(AO, %rax, 4),%xmm0 ;\
vmovddup -11 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8, %xmm2, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\
vmovups -10 * SIZE(AO, %rax, 4), %xmm2 ;\
vfmaddpd %xmm8, %xmm0, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\
vmovddup -10 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm9, %xmm2, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\
vfmaddpd %xmm9, %xmm0, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup -9 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\
vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovddup -7 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm0, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL3(xx) \
vmovups -6 * SIZE(AO, %rax, 4),%xmm2 ;\
vmovddup -7 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8, %xmm6, %xmm7, %xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vmovups -6 * SIZE(AO, %rax, 4),%xmm2 ;\
vfmaddpd %xmm8, %xmm6, %xmm7, %xmm8 ;\
vmovddup -6 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vmovddup -5 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm9, %xmm6, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vfmaddpd %xmm9, %xmm6, %xmm3,%xmm9 ;\
vmovddup (BO, %rax, 4), %xmm7 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup -4 * SIZE(BO, %rax, 4), %xmm1 ;\
vfmaddpd %xmm11,%xmm6, %xmm4, %xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vmovddup -3 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm6, %xmm4, %xmm11 ;\
vmovups -4 * SIZE(AO, %rax, 4),%xmm0 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovups (AO, %rax, 4), %xmm6 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL4(xx) \
vmovups -4 * SIZE(AO, %rax, 4), %xmm2 ;\
vmovups -2 * SIZE(AO, %rax, 4),%xmm0 ;\
vmovddup -3 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm0, %xmm1 ,%xmm12;\
vmovups -2 * SIZE(AO, %rax, 4), %xmm2 ;\
vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm1 ,%xmm12;\
vmovddup -2 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\
vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup -1 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\
vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovddup 1 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm0, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL5(xx) \
vmovups 2 * SIZE(AO, %rax, 4),%xmm2 ;\
vmovddup 1 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\
vmovddup 2 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vmovddup 3 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\
vmovddup 8 * SIZE(BO, %rax, 4), %xmm7 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovddup 4 * SIZE(BO, %rax, 4), %xmm1 ;\
vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup 4 * SIZE(BO, %rax, 4), %xmm1 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vmovddup 5 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\
vmovups 4 * SIZE(AO, %rax, 4),%xmm0 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovups 8 * SIZE(AO, %rax, 4), %xmm6 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL6(xx) \
vmovups 4 * SIZE(AO, %rax, 4), %xmm2 ;\
vmovups 6 * SIZE(AO, %rax, 4),%xmm0 ;\
vmovddup 5 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\
vmovups 6 * SIZE(AO, %rax, 4), %xmm2 ;\
vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\
vmovddup 6 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\
vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup 7 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\
vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovddup 9 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm0, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL7(xx) \
vmovups 10 * SIZE(AO, %rax, 4),%xmm2 ;\
vmovddup 9 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\
vmovddup 10 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\
vmovddup 11 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\
vmovddup 16 * SIZE(BO, %rax, 4), %xmm7 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\
vmovddup 12 * SIZE(BO, %rax, 4), %xmm1 ;\
vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\
vmovddup 13 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\
vmovups 12 * SIZE(AO, %rax, 4), %xmm0 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovups 16 * SIZE(AO, %rax, 4),%xmm6 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL8(xx) \
vmovups 12 * SIZE(AO, %rax, 4), %xmm2 ;\
vmovups 14 * SIZE(AO, %rax, 4), %xmm0 ;\
vmovddup 13 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\
vmovups 14 * SIZE(AO, %rax, 4), %xmm2 ;\
vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\
vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\
vmovddup 14 * SIZE(BO, %rax, 4), %xmm5 ;\
vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\
vfmaddpd %xmm13, %xmm0, %xmm3,%xmm13 ;\
vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\
vfmaddpd %xmm13, %xmm2, %xmm3,%xmm13 ;\
vmovddup 15 * SIZE(BO, %rax, 4), %xmm4 ;\
vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\
vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\
vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\
vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\
vmovddup 17 * SIZE(BO, %rax, 4), %xmm3 ;\
vfmaddpd %xmm11,%xmm0, %xmm4,%xmm11 ;\
vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\
#define KERNEL_SUB1(xx) \
vmovups -16 * SIZE(AO),%xmm0 ;\
@@ -456,97 +456,13 @@
vmovups -16 * SIZE(AO, %rax, 4),%xmm6
vmovddup -16 * SIZE(BO, %rax, 4), %xmm7
vmovddup -15 * SIZE(BO, %rax, 4), %xmm3
.align 16
.L12:
#ifndef SMP
prefetcht0 A_PR1(AO,%rax,4)
prefetcht0 B_PR1(BO,%rax,4)
KERNEL1(16 * 0)
KERNEL2(16 * 0)
prefetcht0 A_PR1+64(AO,%rax,4)
prefetcht0 B_PR1+64(BO,%rax,4)
KERNEL3(16 * 0)
KERNEL4(16 * 0)
prefetcht0 A_PR1+128(AO,%rax,4)
prefetcht0 B_PR1+128(BO,%rax,4)
KERNEL5(16 * 0)
KERNEL6(16 * 0)
prefetcht0 A_PR1+192(AO,%rax,4)
prefetcht0 B_PR1+192(BO,%rax,4)
KERNEL7(16 * 0)
KERNEL8(16 * 0)
addq $8 * SIZE, %rax
je .L15
prefetcht0 A_PR1(AO,%rax,4)
prefetcht0 B_PR1(BO,%rax,4)
KERNEL1(16 * 0)
KERNEL2(16 * 0)
prefetcht0 A_PR1+64(AO,%rax,4)
prefetcht0 B_PR1+64(BO,%rax,4)
KERNEL3(16 * 0)
KERNEL4(16 * 0)
prefetcht0 A_PR1+128(AO,%rax,4)
prefetcht0 B_PR1+128(BO,%rax,4)
KERNEL5(16 * 0)
KERNEL6(16 * 0)
prefetcht0 A_PR1+192(AO,%rax,4)
prefetcht0 B_PR1+192(BO,%rax,4)
KERNEL7(16 * 0)
KERNEL8(16 * 0)
addq $8 * SIZE, %rax
je .L15
prefetcht0 A_PR1(AO,%rax,4)
prefetcht0 B_PR1(BO,%rax,4)
KERNEL1(16 * 0)
KERNEL2(16 * 0)
prefetcht0 A_PR1+64(AO,%rax,4)
prefetcht0 B_PR1+64(BO,%rax,4)
KERNEL3(16 * 0)
KERNEL4(16 * 0)
prefetcht0 A_PR1+128(AO,%rax,4)
prefetcht0 B_PR1+128(BO,%rax,4)
KERNEL5(16 * 0)
KERNEL6(16 * 0)
prefetcht0 A_PR1+192(AO,%rax,4)
prefetcht0 B_PR1+192(BO,%rax,4)
KERNEL7(16 * 0)
KERNEL8(16 * 0)
addq $8 * SIZE, %rax
je .L15
prefetcht0 A_PR1(AO,%rax,4)
prefetcht0 B_PR1(BO,%rax,4)
KERNEL1(16 * 0)
KERNEL2(16 * 0)
prefetcht0 A_PR1+64(AO,%rax,4)
prefetcht0 B_PR1+64(BO,%rax,4)
KERNEL3(16 * 0)
KERNEL4(16 * 0)
prefetcht0 A_PR1+128(AO,%rax,4)
prefetcht0 B_PR1+128(BO,%rax,4)
KERNEL5(16 * 0)
KERNEL6(16 * 0)
prefetcht0 A_PR1+192(AO,%rax,4)
prefetcht0 B_PR1+192(BO,%rax,4)
KERNEL7(16 * 0)
KERNEL8(16 * 0)
addq $8 * SIZE, %rax
jnz .L12
.align 16
#else
#ifdef OPTMODULE
#if defined(OPTBYMODULE) || !defined(SMP)
prefetcht0 A_PR1(AO,%rax,4)
prefetcht0 B_PR1(BO,%rax,4)
@@ -586,7 +502,6 @@
.align 16
#endif
#endif
.L15:


Write Preview
Loading…
Cancel
Save