diff options
| author | Suren A. Chilingaryan <csa@suren.me> | 2015-11-20 05:55:36 +0100 |
|---|---|---|
| committer | Suren A. Chilingaryan <csa@suren.me> | 2015-11-20 05:55:36 +0100 |
| commit | 2bda41263f2464c271509b0bd9ea9062c239d851 (patch) | |
| tree | a3f96dea1c86ecef4f822bff4f94faa133497e11 /pcilib/pagecpy.c | |
| parent | 8ee679f837aed09f2abe5d47186505f98ccb4b6d (diff) | |
| download | pcitool-2bda41263f2464c271509b0bd9ea9062c239d851.tar.gz pcitool-2bda41263f2464c271509b0bd9ea9062c239d851.tar.bz2 pcitool-2bda41263f2464c271509b0bd9ea9062c239d851.tar.xz pcitool-2bda41263f2464c271509b0bd9ea9062c239d851.zip | |
Fix AVX memory copy
Diffstat (limited to 'pcilib/pagecpy.c')
| -rw-r--r-- | pcilib/pagecpy.c | 78 |
1 files changed, 39 insertions, 39 deletions
diff --git a/pcilib/pagecpy.c b/pcilib/pagecpy.c index f474f9f..c1e7bbd 100644 --- a/pcilib/pagecpy.c +++ b/pcilib/pagecpy.c @@ -87,7 +87,7 @@ void *memcpy128(void * dst, void const * src, size_t len) { } */ -void pcilib_memcpy4k_avx(void *dst, void *src, size_t size) { +void pcilib_memcpy4k_avx(void *dst, const void *src, size_t size) { size_t sse_size = (size / 512); __asm__ __volatile__ ( @@ -96,57 +96,57 @@ void pcilib_memcpy4k_avx(void *dst, void *src, size_t size) { "1: \n\t" - "vmovdqa (%0,%%rax), %%ymm0 \n\t" - "vmovdqa 32(%0,%%rax), %%ymm1 \n\t" - "vmovdqa 64(%0,%%rax), %%ymm2 \n\t" - "vmovdqa 96(%0,%%rax), %%ymm3 \n\t" - "vmovdqa 128(%0,%%rax), %%ymm4 \n\t" - "vmovdqa 160(%0,%%rax), %%ymm5 \n\t" - "vmovdqa 192(%0,%%rax), %%ymm6 \n\t" - "vmovdqa 224(%0,%%rax), %%ymm7 \n\t" - - "vmovdqa 256(%0,%%rax), %%ymm8 \n\t" - "vmovdqa 288(%0,%%rax), %%ymm9 \n\t" - "vmovdqa 320(%0,%%rax), %%ymm10 \n\t" - "vmovdqa 352(%0,%%rax), %%ymm11 \n\t" - "vmovdqa 384(%0,%%rax), %%ymm12 \n\t" - "vmovdqa 416(%0,%%rax), %%ymm13 \n\t" - "vmovdqa 448(%0,%%rax), %%ymm14 \n\t" - "vmovdqa 480(%0,%%rax), %%ymm15 \n\t" - - "vmovntps %%ymm0, (%1,%%rax) \n\t" - "vmovntps %%ymm1, 32(%1,%%rax) \n\t" - "vmovntps %%ymm2, 64(%1,%%rax) \n\t" - "vmovntps %%ymm3, 96(%1,%%rax) \n\t" - "vmovntps %%ymm4, 128(%1,%%rax) \n\t" - "vmovntps %%ymm5, 160(%1,%%rax) \n\t" - "vmovntps %%ymm6, 192(%1,%%rax) \n\t" - "vmovntps %%ymm7, 224(%1,%%rax) \n\t" - - "vmovntps %%ymm8, 256(%1,%%rax) \n\t" - "vmovntps %%ymm9, 288(%1,%%rax) \n\t" - "vmovntps %%ymm10, 320(%1,%%rax) \n\t" - "vmovntps %%ymm11, 352(%1,%%rax) \n\t" - "vmovntps %%ymm12, 384(%1,%%rax) \n\t" - "vmovntps %%ymm13, 416(%1,%%rax) \n\t" - "vmovntps %%ymm14, 448(%1,%%rax) \n\t" - "vmovntps %%ymm15, 480(%1,%%rax) \n\t" + "vmovdqa (%1,%%rax), %%ymm0 \n\t" + "vmovdqa 32(%1,%%rax), %%ymm1 \n\t" + "vmovdqa 64(%1,%%rax), %%ymm2 \n\t" + "vmovdqa 96(%1,%%rax), %%ymm3 \n\t" + "vmovdqa 128(%1,%%rax), %%ymm4 \n\t" + "vmovdqa 160(%1,%%rax), %%ymm5 \n\t" + "vmovdqa 192(%1,%%rax), %%ymm6 \n\t" + "vmovdqa 224(%1,%%rax), %%ymm7 \n\t" + + "vmovdqa 256(%1,%%rax), %%ymm8 \n\t" + "vmovdqa 288(%1,%%rax), %%ymm9 \n\t" + "vmovdqa 320(%1,%%rax), %%ymm10 \n\t" + "vmovdqa 352(%1,%%rax), %%ymm11 \n\t" + "vmovdqa 384(%1,%%rax), %%ymm12 \n\t" + "vmovdqa 416(%1,%%rax), %%ymm13 \n\t" + "vmovdqa 448(%1,%%rax), %%ymm14 \n\t" + "vmovdqa 480(%1,%%rax), %%ymm15 \n\t" + + "vmovdqa %%ymm0, (%0,%%rax) \n\t" + "vmovdqa %%ymm1, 32(%0,%%rax) \n\t" + "vmovntps %%ymm2, 64(%0,%%rax) \n\t" + "vmovntps %%ymm3, 96(%0,%%rax) \n\t" + "vmovntps %%ymm4, 128(%0,%%rax) \n\t" + "vmovntps %%ymm5, 160(%0,%%rax) \n\t" + "vmovntps %%ymm6, 192(%0,%%rax) \n\t" + "vmovntps %%ymm7, 224(%0,%%rax) \n\t" + + "vmovntps %%ymm8, 256(%0,%%rax) \n\t" + "vmovntps %%ymm9, 288(%0,%%rax) \n\t" + "vmovntps %%ymm10, 320(%0,%%rax) \n\t" + "vmovntps %%ymm11, 352(%0,%%rax) \n\t" + "vmovntps %%ymm12, 384(%0,%%rax) \n\t" + "vmovntps %%ymm13, 416(%0,%%rax) \n\t" + "vmovntps %%ymm14, 448(%0,%%rax) \n\t" + "vmovntps %%ymm15, 480(%0,%%rax) \n\t" "add $512, %%rax \n\t" "dec %2 \n\t" "jnz 1b \n\t" "pop %2 \n\t" - "sfence" + "mfence" : : "p" (dst), "p" (src), "r" (sse_size) : "%rax" ); } -void pcilib_pagecpy(void *dst, void *src, size_t size) { +void pcilib_pagecpy(void *dst, const void *src, size_t size) { int gen = pcilib_get_cpu_gen(); - if ((gen > 3)&&(size%4096==0)&&((uintptr_t)dst%32==0)&&((uintptr_t)src%32==0)) { + if ((gen > 3)&&((size%4096)==0)&&(((uintptr_t)dst%32)==0)&&(((uintptr_t)src%32)==0)) { pcilib_memcpy4k_avx(dst, src, size); } else memcpy(dst, src, size); |