[AMD Official Use Only]
Hi Ananyev,
The current memcpy implementation in Glibc is based out of assembly coding.
Although memcpy could have been implemented with intrinsic, but since our AMD
library developers are working on the Glibc functions, they have provided a
tailored implementation based out of inline assembly coding.
Thanks for your support,
Keesang
-----Original Message-----
From: Ananyev, Konstantin <konstantin.anan...@intel.com>
Sent: Thursday, October 21, 2021 10:40 AM
To: Song, Keesang <keesang.s...@amd.com>; Thomas Monjalon
<tho...@monjalon.net>; Aman Kumar <aman.ku...@vvdntech.in>
Cc: dev@dpdk.org; rasl...@nvidia.com; as...@nvidia.com; s...@nvidia.com;
viachesl...@nvidia.com; akozy...@nvidia.com; ma...@nvidia.com; Burakov, Anatoly
<anatoly.bura...@intel.com>; aman.ku...@vvdntech.in; jerinjac...@gmail.com;
Richardson, Bruce <bruce.richard...@intel.com>; david.march...@redhat.com
Subject: RE: [dpdk-dev] [PATCH v2 1/2] lib/eal: add amd epyc2 memcpy routine to
eal
[AMD Official Use Only]
[CAUTION: External Email]
>
> Hi Thomas,
>
> I hope this can make some explanation to your question.
> We(AMD Linux library support team) have implemented the custom
> tailored memcpy solution which is a close match with DPDK use case
> requirements like the below.
> 1) Min 64B length data packet with cache aligned Source and Destination.
> 2) Non-Temporal load and temporal store for cache aligned source for
> both RX and TX paths. Could not implement the non-temporal
> store for TX_PATH, as non-Temporal load/stores works only with 32B aligned
> addresses for AVX2
> 3) This solution works for all AVX2 supported AMD machines.
>
> Internally we have completed the integrity testing and benchmarking of
> the solution and found gains of 8.4% to 14.5% specifically on Milan
> CPU(3rd Gen of EPYC Processor)
It still not clear to me why it has to be written in assembler.
Why similar stuff can't be written in C with instincts, as rest of rte_memcpy.h
does?
>
> Thanks for your support,
> Keesang
>
> -----Original Message-----
> From: Thomas Monjalon <tho...@monjalon.net>
> Sent: Tuesday, October 19, 2021 5:31 AM
> To: Aman Kumar <aman.ku...@vvdntech.in>
> Cc: dev@dpdk.org; rasl...@nvidia.com; as...@nvidia.com;
> s...@nvidia.com; viachesl...@nvidia.com; akozy...@nvidia.com;
> ma...@nvidia.com; anatoly.bura...@intel.com; Song, Keesang
> <keesang.s...@amd.com>; aman.ku...@vvdntech.in; jerinjac...@gmail.com;
> bruce.richard...@intel.com; konstantin.anan...@intel.com;
> david.march...@redhat.com
> Subject: Re: [dpdk-dev] [PATCH v2 1/2] lib/eal: add amd epyc2 memcpy
> routine to eal
>
> [CAUTION: External Email]
>
> 19/10/2021 12:47, Aman Kumar:
> > This patch provides rte_memcpy* calls optimized for AMD EPYC
> > platforms. Use config/x86/x86_amd_epyc_linux_gcc as cross-file with
> > meson to build dpdk for AMD EPYC platforms.
>
> Please split in 2 patches: platform & memcpy.
>
> What optimization is specific to EPYC?
>
> I dislike the asm code below.
> What is AMD specific inside?
> Can it use compiler intrinsics as it is done elsewhere?
>
> > +static __rte_always_inline void *
> > +rte_memcpy_aligned_ntload_tstore16_amdepyc2(void *dst,
> > + const void *src,
> > + size_t size) {
> > + asm volatile goto("movq %0, %%rsi\n\t"
> > + "movq %1, %%rdi\n\t"
> > + "movq %2, %%rdx\n\t"
> > + "cmpq $(128), %%rdx\n\t"
> > + "jb 202f\n\t"
> > + "201:\n\t"
> > + "vmovntdqa (%%rsi), %%ymm0\n\t"
> > + "vmovntdqa 32(%%rsi), %%ymm1\n\t"
> > + "vmovntdqa 64(%%rsi), %%ymm2\n\t"
> > + "vmovntdqa 96(%%rsi), %%ymm3\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, 32(%%rdi)\n\t"
> > + "vmovdqu %%ymm2, 64(%%rdi)\n\t"
> > + "vmovdqu %%ymm3, 96(%%rdi)\n\t"
> > + "addq $128, %%rsi\n\t"
> > + "addq $128, %%rdi\n\t"
> > + "subq $128, %%rdx\n\t"
> > + "jz %l[done]\n\t"
> > + "cmpq $128, %%rdx\n\t" /*Vector Size 32B. */
> > + "jae 201b\n\t"
> > + "202:\n\t"
> > + "cmpq $64, %%rdx\n\t"
> > + "jb 203f\n\t"
> > + "vmovntdqa (%%rsi), %%ymm0\n\t"
> > + "vmovntdqa 32(%%rsi), %%ymm1\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, 32(%%rdi)\n\t"
> > + "addq $64, %%rsi\n\t"
> > + "addq $64, %%rdi\n\t"
> > + "subq $64, %%rdx\n\t"
> > + "jz %l[done]\n\t"
> > + "203:\n\t"
> > + "cmpq $32, %%rdx\n\t"
> > + "jb 204f\n\t"
> > + "vmovntdqa (%%rsi), %%ymm0\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "addq $32, %%rsi\n\t"
> > + "addq $32, %%rdi\n\t"
> > + "subq $32, %%rdx\n\t"
> > + "jz %l[done]\n\t"
> > + "204:\n\t"
> > + "cmpb $16, %%dl\n\t"
> > + "jb 205f\n\t"
> > + "vmovntdqa (%%rsi), %%xmm0\n\t"
> > + "vmovdqu %%xmm0, (%%rdi)\n\t"
> > + "addq $16, %%rsi\n\t"
> > + "addq $16, %%rdi\n\t"
> > + "subq $16, %%rdx\n\t"
> > + "jz %l[done]\n\t"
> > + "205:\n\t"
> > + "cmpb $2, %%dl\n\t"
> > + "jb 208f\n\t"
> > + "cmpb $4, %%dl\n\t"
> > + "jbe 207f\n\t"
> > + "cmpb $8, %%dl\n\t"
> > + "jbe 206f\n\t"
> > + "movq -8(%%rsi,%%rdx), %%rcx\n\t"
> > + "movq (%%rsi), %%rsi\n\t"
> > + "movq %%rcx, -8(%%rdi,%%rdx)\n\t"
> > + "movq %%rsi, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "206:\n\t"
> > + "movl -4(%%rsi,%%rdx), %%ecx\n\t"
> > + "movl (%%rsi), %%esi\n\t"
> > + "movl %%ecx, -4(%%rdi,%%rdx)\n\t"
> > + "movl %%esi, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "207:\n\t"
> > + "movzwl -2(%%rsi,%%rdx), %%ecx\n\t"
> > + "movzwl (%%rsi), %%esi\n\t"
> > + "movw %%cx, -2(%%rdi,%%rdx)\n\t"
> > + "movw %%si, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "208:\n\t"
> > + "movzbl (%%rsi), %%ecx\n\t"
> > + "movb %%cl, (%%rdi)"
> > + :
> > + : "r"(src), "r"(dst), "r"(size)
> > + : "rcx", "rdx", "rsi", "rdi", "ymm0", "ymm1", "ymm2", "ymm3", "memory"
> > + : done
> > + );
> > +done:
> > + return dst;
> > +}
> > +
> > +static __rte_always_inline void *
> > +rte_memcpy_generic(void *dst, const void *src, size_t len) {
> > + asm goto("movq %0, %%rsi\n\t"
> > + "movq %1, %%rdi\n\t"
> > + "movq %2, %%rdx\n\t"
> > + "movq %%rdi, %%rax\n\t"
> > + "cmp $32, %%rdx\n\t"
> > + "jb 101f\n\t"
> > + "cmp $(32 * 2), %%rdx\n\t"
> > + "ja 108f\n\t"
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu -32(%%rsi,%%rdx), %%ymm1\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, -32(%%rdi,%%rdx)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > + "101:\n\t"
> > + /* Less than 1 VEC. */
> > + "cmpb $32, %%dl\n\t"
> > + "jae 103f\n\t"
> > + "cmpb $16, %%dl\n\t"
> > + "jae 104f\n\t"
> > + "cmpb $8, %%dl\n\t"
> > + "jae 105f\n\t"
> > + "cmpb $4, %%dl\n\t"
> > + "jae 106f\n\t"
> > + "cmpb $1, %%dl\n\t"
> > + "ja 107f\n\t"
> > + "jb 102f\n\t"
> > + "movzbl (%%rsi), %%ecx\n\t"
> > + "movb %%cl, (%%rdi)\n\t"
> > + "102:\n\t"
> > + "jmp %l[done]\n\t"
> > + "103:\n\t"
> > + /* From 32 to 63. No branch when size == 32. */
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu -32(%%rsi,%%rdx), %%ymm1\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, -32(%%rdi,%%rdx)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > + /* From 16 to 31. No branch when size == 16. */
> > + "104:\n\t"
> > + "vmovdqu (%%rsi), %%xmm0\n\t"
> > + "vmovdqu -16(%%rsi,%%rdx), %%xmm1\n\t"
> > + "vmovdqu %%xmm0, (%%rdi)\n\t"
> > + "vmovdqu %%xmm1, -16(%%rdi,%%rdx)\n\t"
> > + "jmp %l[done]\n\t"
> > + "105:\n\t"
> > + /* From 8 to 15. No branch when size == 8. */
> > + "movq -8(%%rsi,%%rdx), %%rcx\n\t"
> > + "movq (%%rsi), %%rsi\n\t"
> > + "movq %%rcx, -8(%%rdi,%%rdx)\n\t"
> > + "movq %%rsi, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "106:\n\t"
> > + /* From 4 to 7. No branch when size == 4. */
> > + "movl -4(%%rsi,%%rdx), %%ecx\n\t"
> > + "movl (%%rsi), %%esi\n\t"
> > + "movl %%ecx, -4(%%rdi,%%rdx)\n\t"
> > + "movl %%esi, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "107:\n\t"
> > + /* From 2 to 3. No branch when size == 2. */
> > + "movzwl -2(%%rsi,%%rdx), %%ecx\n\t"
> > + "movzwl (%%rsi), %%esi\n\t"
> > + "movw %%cx, -2(%%rdi,%%rdx)\n\t"
> > + "movw %%si, (%%rdi)\n\t"
> > + "jmp %l[done]\n\t"
> > + "108:\n\t"
> > + /* More than 2 * VEC and there may be overlap between destination */
> > + /* and source. */
> > + "cmpq $(32 * 8), %%rdx\n\t"
> > + "ja 111f\n\t"
> > + "cmpq $(32 * 4), %%rdx\n\t"
> > + "jb 109f\n\t"
> > + /* Copy from 4 * VEC to 8 * VEC, inclusively. */
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu 32(%%rsi), %%ymm1\n\t"
> > + "vmovdqu (32 * 2)(%%rsi), %%ymm2\n\t"
> > + "vmovdqu (32 * 3)(%%rsi), %%ymm3\n\t"
> > + "vmovdqu -32(%%rsi,%%rdx), %%ymm4\n\t"
> > + "vmovdqu -(32 * 2)(%%rsi,%%rdx), %%ymm5\n\t"
> > + "vmovdqu -(32 * 3)(%%rsi,%%rdx), %%ymm6\n\t"
> > + "vmovdqu -(32 * 4)(%%rsi,%%rdx), %%ymm7\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, 32(%%rdi)\n\t"
> > + "vmovdqu %%ymm2, (32 * 2)(%%rdi)\n\t"
> > + "vmovdqu %%ymm3, (32 * 3)(%%rdi)\n\t"
> > + "vmovdqu %%ymm4, -32(%%rdi,%%rdx)\n\t"
> > + "vmovdqu %%ymm5, -(32 * 2)(%%rdi,%%rdx)\n\t"
> > + "vmovdqu %%ymm6, -(32 * 3)(%%rdi,%%rdx)\n\t"
> > + "vmovdqu %%ymm7, -(32 * 4)(%%rdi,%%rdx)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > + "109:\n\t"
> > + /* Copy from 2 * VEC to 4 * VEC. */
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu 32(%%rsi), %%ymm1\n\t"
> > + "vmovdqu -32(%%rsi,%%rdx), %%ymm2\n\t"
> > + "vmovdqu -(32 * 2)(%%rsi,%%rdx), %%ymm3\n\t"
> > + "vmovdqu %%ymm0, (%%rdi)\n\t"
> > + "vmovdqu %%ymm1, 32(%%rdi)\n\t"
> > + "vmovdqu %%ymm2, -32(%%rdi,%%rdx)\n\t"
> > + "vmovdqu %%ymm3, -(32 * 2)(%%rdi,%%rdx)\n\t"
> > + "vzeroupper\n\t"
> > + "110:\n\t"
> > + "jmp %l[done]\n\t"
> > + "111:\n\t"
> > + "cmpq %%rsi, %%rdi\n\t"
> > + "ja 113f\n\t"
> > + /* Source == destination is less common. */
> > + "je 110b\n\t"
> > + /* Load the first VEC and last 4 * VEC to
> > + * support overlapping addresses.
> > + */
> > + "vmovdqu (%%rsi), %%ymm4\n\t"
> > + "vmovdqu -32(%%rsi, %%rdx), %%ymm5\n\t"
> > + "vmovdqu -(32 * 2)(%%rsi, %%rdx), %%ymm6\n\t"
> > + "vmovdqu -(32 * 3)(%%rsi, %%rdx), %%ymm7\n\t"
> > + "vmovdqu -(32 * 4)(%%rsi, %%rdx), %%ymm8\n\t"
> > + /* Save start and stop of the destination buffer. */
> > + "movq %%rdi, %%r11\n\t"
> > + "leaq -32(%%rdi, %%rdx), %%rcx\n\t"
> > + /* Align destination for aligned stores in the loop. Compute */
> > + /* how much destination is misaligned. */
> > + "movq %%rdi, %%r8\n\t"
> > + "andq $(32 - 1), %%r8\n\t"
> > + /* Get the negative of offset for alignment. */
> > + "subq $32, %%r8\n\t"
> > + /* Adjust source. */
> > + "subq %%r8, %%rsi\n\t"
> > + /* Adjust destination which should be aligned now. */
> > + "subq %%r8, %%rdi\n\t"
> > + /* Adjust length. */
> > + "addq %%r8, %%rdx\n\t"
> > + /* Check non-temporal store threshold. */
> > + "cmpq $(1024*1024), %%rdx\n\t"
> > + "ja 115f\n\t"
> > + "112:\n\t"
> > + /* Copy 4 * VEC a time forward. */
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu 32(%%rsi), %%ymm1\n\t"
> > + "vmovdqu (32 * 2)(%%rsi), %%ymm2\n\t"
> > + "vmovdqu (32 * 3)(%%rsi), %%ymm3\n\t"
> > + "addq $(32 * 4), %%rsi\n\t"
> > + "subq $(32 * 4), %%rdx\n\t"
> > + "vmovdqa %%ymm0, (%%rdi)\n\t"
> > + "vmovdqa %%ymm1, 32(%%rdi)\n\t"
> > + "vmovdqa %%ymm2, (32 * 2)(%%rdi)\n\t"
> > + "vmovdqa %%ymm3, (32 * 3)(%%rdi)\n\t"
> > + "addq $(32 * 4), %%rdi\n\t"
> > + "cmpq $(32 * 4), %%rdx\n\t"
> > + "ja 112b\n\t"
> > + /* Store the last 4 * VEC. */
> > + "vmovdqu %%ymm5, (%%rcx)\n\t"
> > + "vmovdqu %%ymm6, -32(%%rcx)\n\t"
> > + "vmovdqu %%ymm7, -(32 * 2)(%%rcx)\n\t"
> > + "vmovdqu %%ymm8, -(32 * 3)(%%rcx)\n\t"
> > + /* Store the first VEC. */
> > + "vmovdqu %%ymm4, (%%r11)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > + "113:\n\t"
> > + /* Load the first 4*VEC and last VEC to support overlapping
> > addresses.*/
> > + "vmovdqu (%%rsi), %%ymm4\n\t"
> > + "vmovdqu 32(%%rsi), %%ymm5\n\t"
> > + "vmovdqu (32 * 2)(%%rsi), %%ymm6\n\t"
> > + "vmovdqu (32 * 3)(%%rsi), %%ymm7\n\t"
> > + "vmovdqu -32(%%rsi,%%rdx), %%ymm8\n\t"
> > + /* Save stop of the destination buffer. */
> > + "leaq -32(%%rdi, %%rdx), %%r11\n\t"
> > + /* Align destination end for aligned stores in the loop. Compute */
> > + /* how much destination end is misaligned. */
> > + "leaq -32(%%rsi, %%rdx), %%rcx\n\t"
> > + "movq %%r11, %%r9\n\t"
> > + "movq %%r11, %%r8\n\t"
> > + "andq $(32 - 1), %%r8\n\t"
> > + /* Adjust source. */
> > + "subq %%r8, %%rcx\n\t"
> > + /* Adjust the end of destination which should be aligned now. */
> > + "subq %%r8, %%r9\n\t"
> > + /* Adjust length. */
> > + "subq %%r8, %%rdx\n\t"
> > + /* Check non-temporal store threshold. */
> > + "cmpq $(1024*1024), %%rdx\n\t"
> > + "ja 117f\n\t"
> > + "114:\n\t"
> > + /* Copy 4 * VEC a time backward. */
> > + "vmovdqu (%%rcx), %%ymm0\n\t"
> > + "vmovdqu -32(%%rcx), %%ymm1\n\t"
> > + "vmovdqu -(32 * 2)(%%rcx), %%ymm2\n\t"
> > + "vmovdqu -(32 * 3)(%%rcx), %%ymm3\n\t"
> > + "subq $(32 * 4), %%rcx\n\t"
> > + "subq $(32 * 4), %%rdx\n\t"
> > + "vmovdqa %%ymm0, (%%r9)\n\t"
> > + "vmovdqa %%ymm1, -32(%%r9)\n\t"
> > + "vmovdqa %%ymm2, -(32 * 2)(%%r9)\n\t"
> > + "vmovdqa %%ymm3, -(32 * 3)(%%r9)\n\t"
> > + "subq $(32 * 4), %%r9\n\t"
> > + "cmpq $(32 * 4), %%rdx\n\t"
> > + "ja 114b\n\t"
> > + /* Store the first 4 * VEC. */
> > + "vmovdqu %%ymm4, (%%rdi)\n\t"
> > + "vmovdqu %%ymm5, 32(%%rdi)\n\t"
> > + "vmovdqu %%ymm6, (32 * 2)(%%rdi)\n\t"
> > + "vmovdqu %%ymm7, (32 * 3)(%%rdi)\n\t"
> > + /* Store the last VEC. */
> > + "vmovdqu %%ymm8, (%%r11)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > +
> > + "115:\n\t"
> > + /* Don't use non-temporal store if there is overlap between */
> > + /* destination and source since destination may be in cache */
> > + /* when source is loaded. */
> > + "leaq (%%rdi, %%rdx), %%r10\n\t"
> > + "cmpq %%r10, %%rsi\n\t"
> > + "jb 112b\n\t"
> > + "116:\n\t"
> > + /* Copy 4 * VEC a time forward with non-temporal stores. */
> > + "prefetcht0 (32*4*2)(%%rsi)\n\t"
> > + "prefetcht0 (32*4*2 + 64)(%%rsi)\n\t"
> > + "prefetcht0 (32*4*3)(%%rsi)\n\t"
> > + "prefetcht0 (32*4*3 + 64)(%%rsi)\n\t"
> > + "vmovdqu (%%rsi), %%ymm0\n\t"
> > + "vmovdqu 32(%%rsi), %%ymm1\n\t"
> > + "vmovdqu (32 * 2)(%%rsi), %%ymm2\n\t"
> > + "vmovdqu (32 * 3)(%%rsi), %%ymm3\n\t"
> > + "addq $(32*4), %%rsi\n\t"
> > + "subq $(32*4), %%rdx\n\t"
> > + "vmovntdq %%ymm0, (%%rdi)\n\t"
> > + "vmovntdq %%ymm1, 32(%%rdi)\n\t"
> > + "vmovntdq %%ymm2, (32 * 2)(%%rdi)\n\t"
> > + "vmovntdq %%ymm3, (32 * 3)(%%rdi)\n\t"
> > + "addq $(32*4), %%rdi\n\t"
> > + "cmpq $(32*4), %%rdx\n\t"
> > + "ja 116b\n\t"
> > + "sfence\n\t"
> > + /* Store the last 4 * VEC. */
> > + "vmovdqu %%ymm5, (%%rcx)\n\t"
> > + "vmovdqu %%ymm6, -32(%%rcx)\n\t"
> > + "vmovdqu %%ymm7, -(32 * 2)(%%rcx)\n\t"
> > + "vmovdqu %%ymm8, -(32 * 3)(%%rcx)\n\t"
> > + /* Store the first VEC. */
> > + "vmovdqu %%ymm4, (%%r11)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]\n\t"
> > + "117:\n\t"
> > + /* Don't use non-temporal store if there is overlap between */
> > + /* destination and source since destination may be in cache */
> > + /* when source is loaded. */
> > + "leaq (%%rcx, %%rdx), %%r10\n\t"
> > + "cmpq %%r10, %%r9\n\t"
> > + "jb 114b\n\t"
> > + "118:\n\t"
> > + /* Copy 4 * VEC a time backward with non-temporal stores. */
> > + "prefetcht0 (-32 * 4 * 2)(%%rcx)\n\t"
> > + "prefetcht0 (-32 * 4 * 2 - 64)(%%rcx)\n\t"
> > + "prefetcht0 (-32 * 4 * 3)(%%rcx)\n\t"
> > + "prefetcht0 (-32 * 4 * 3 - 64)(%%rcx)\n\t"
> > + "vmovdqu (%%rcx), %%ymm0\n\t"
> > + "vmovdqu -32(%%rcx), %%ymm1\n\t"
> > + "vmovdqu -(32 * 2)(%%rcx), %%ymm2\n\t"
> > + "vmovdqu -(32 * 3)(%%rcx), %%ymm3\n\t"
> > + "subq $(32*4), %%rcx\n\t"
> > + "subq $(32*4), %%rdx\n\t"
> > + "vmovntdq %%ymm0, (%%r9)\n\t"
> > + "vmovntdq %%ymm1, -32(%%r9)\n\t"
> > + "vmovntdq %%ymm2, -(32 * 2)(%%r9)\n\t"
> > + "vmovntdq %%ymm3, -(32 * 3)(%%r9)\n\t"
> > + "subq $(32 * 4), %%r9\n\t"
> > + "cmpq $(32 * 4), %%rdx\n\t"
> > + "ja 118b\n\t"
> > + "sfence\n\t"
> > + /* Store the first 4 * VEC. */
> > + "vmovdqu %%ymm4, (%%rdi)\n\t"
> > + "vmovdqu %%ymm5, 32(%%rdi)\n\t"
> > + "vmovdqu %%ymm6, (32 * 2)(%%rdi)\n\t"
> > + "vmovdqu %%ymm7, (32 * 3)(%%rdi)\n\t"
> > + /* Store the last VEC. */
> > + "vmovdqu %%ymm8, (%%r11)\n\t"
> > + "vzeroupper\n\t"
> > + "jmp %l[done]"
> > + :
> > + : "r"(src), "r"(dst), "r"(len)
> > + : "rax", "rcx", "rdx", "rdi", "rsi", "r8", "r9", "r10", "r11", "r12",
> > "ymm0",
> > + "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8",
> > "memory"
> > + : done
> > + );
> > +done:
> > + return dst;
> > +}
>
>
