Mirrored_Repos/LookingGlass
1
0
Fork 0
mirror of https://github.com/gnif/LookingGlass.git synced 2025-12-21 12:10:15 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

[host] use the new memcpySSE implementation

Browse Source
This commit is contained in:
Geoffrey McRae
2018-05-22 18:58:56 +10:00
parent 6f141fe393
commit 15a337fee8
6 changed files with 186 additions and 307 deletions
Download Patch File Download Diff File Expand all files Collapse all files

198
common/memcpySSE.h
View File

@@ -26,108 +26,110 @@ Place, Suite 330, Boston, MA 02111-1307 USA
#include "debug.h"
static inline void memcpySSE(void * dst, const void * src, size_t length)
{
// check if we can't perform an aligned copy
if (((uintptr_t)src & 0xF) != ((uintptr_t)dst & 0xF))
#if defined(NATIVE_MEMCPY)
#define memcpySSE memcpy
#elif defined(_MSC_VER)
extern "C" void * memcpySSE(void *dst, const void * src, size_t length);
#elif (defined(__GNUC__) || defined(__GNUG__)) && defined(__i386__)
inline static void * memcpySSE(void *dst, const void * src, size_t length)
{
if (length == 0 || dst == src)
return;
static bool unalignedDstWarn = false;
if (!unalignedDstWarn)
// copies under 1MB are faster with the inlined memcpy
// tell the dev to use that instead
if (length < 1048576)
{
DEBUG_WARN("Memcpy64 unable to perform aligned copy, performance will suffer");
unalignedDstWarn = true;
}
// fallback to system memcpy
memcpy(dst, src, length);
return;
}
// check if the source needs alignment
{
uint8_t * _src = (uint8_t *)src;
unsigned int count = (16 - ((uintptr_t)src & 0xF)) & 0xF;
static bool unalignedSrcWarn = false;
if (count > 0)
{
if (!unalignedSrcWarn)
static bool smallBufferWarn = false;
if (!smallBufferWarn)
{
DEBUG_WARN("Memcpy64 unaligned source, performance will suffer");
unalignedSrcWarn = true;
DEBUG_WARN("Do not use memcpySSE for copies under 1MB in size!");
smallBufferWarn = true;
}
uint8_t * _dst = (uint8_t *)dst;
for (unsigned int i = count; i > 0; --i)
*_dst++ = *_src++;
src = _src;
dst = _dst;
length -= count;
memcpy(dst, src, length);
return;
}
const void * end = dst + (length & ~0x7F);
const size_t off = (7 - ((length & 0x7F) >> 4)) * 10;
__asm__ __volatile__ (
"cmp %[dst],%[end] \n\t"
"je Remain_%= \n\t"
// perform SIMD block copy
"loop_%=: \n\t"
"vmovaps 0x00(%[src]),%%xmm0 \n\t"
"vmovaps 0x10(%[src]),%%xmm1 \n\t"
"vmovaps 0x20(%[src]),%%xmm2 \n\t"
"vmovaps 0x30(%[src]),%%xmm3 \n\t"
"vmovaps 0x40(%[src]),%%xmm4 \n\t"
"vmovaps 0x50(%[src]),%%xmm5 \n\t"
"vmovaps 0x60(%[src]),%%xmm6 \n\t"
"vmovaps 0x70(%[src]),%%xmm7 \n\t"
"vmovntdq %%xmm0 ,0x00(%[dst]) \n\t"
"vmovntdq %%xmm1 ,0x10(%[dst]) \n\t"
"vmovntdq %%xmm2 ,0x20(%[dst]) \n\t"
"vmovntdq %%xmm3 ,0x30(%[dst]) \n\t"
"vmovntdq %%xmm4 ,0x40(%[dst]) \n\t"
"vmovntdq %%xmm5 ,0x50(%[dst]) \n\t"
"vmovntdq %%xmm6 ,0x60(%[dst]) \n\t"
"vmovntdq %%xmm7 ,0x70(%[dst]) \n\t"
"add $0x80,%[dst] \n\t"
"add $0x80,%[src] \n\t"
"cmp %[dst],%[end] \n\t"
"jne loop_%= \n\t"
"Remain_%=: \n\t"
// copy any remaining 16 byte blocks
"call GetPC_%=\n\t"
"Offset_%=:\n\t"
"add $(BlockTable_%= - Offset_%=), %%eax \n\t"
"add %[off],%%eax \n\t"
"jmp *%%eax \n\t"
"GetPC_%=:\n\t"
"mov (%%esp), %%eax \n\t"
"ret \n\t"
"BlockTable_%=:\n\t"
"vmovaps 0x60(%[src]),%%xmm6 \n\t"
"vmovntdq %%xmm6 ,0x60(%[dst]) \n\t"
"vmovaps 0x50(%[src]),%%xmm5 \n\t"
"vmovntdq %%xmm5 ,0x50(%[dst]) \n\t"
"vmovaps 0x40(%[src]),%%xmm4 \n\t"
"vmovntdq %%xmm4 ,0x40(%[dst]) \n\t"
"vmovaps 0x30(%[src]),%%xmm3 \n\t"
"vmovntdq %%xmm3 ,0x30(%[dst]) \n\t"
"vmovaps 0x20(%[src]),%%xmm2 \n\t"
"vmovntdq %%xmm2 ,0x20(%[dst]) \n\t"
"vmovaps 0x10(%[src]),%%xmm1 \n\t"
"vmovntdq %%xmm1 ,0x10(%[dst]) \n\t"
"vmovaps 0x00(%[src]),%%xmm0 \n\t"
"vmovntdq %%xmm0 ,0x00(%[dst]) \n\t"
"nop\n\t"
"nop\n\t"
: [dst]"+r" (dst),
[src]"+r" (src)
: [off]"r" (off),
[end]"r" (end)
: "eax",
"xmm0",
"xmm1",
"xmm2",
"xmm3",
"xmm4",
"xmm5",
"xmm6",
"xmm7",
"memory"
);
//copy any remaining bytes
memcpy(dst, src, length & 0xF);
}
__m128i * _src = (__m128i *)src;
__m128i * _dst = (__m128i *)dst;
__m128i v0, v1, v2, v3, v4, v5, v6, v7;
const size_t sselen = length & ~0x7F;
const __m128i * _end = (__m128i *)((uintptr_t)src + sselen);
for (; _src != _end; _src += 8, _dst += 8)
{
_mm_prefetch(((char *)(_src + 8 )), _MM_HINT_NTA);
_mm_prefetch(((char *)(_src + 9 )), _MM_HINT_NTA);
_mm_prefetch(((char *)(_src + 10)), _MM_HINT_NTA);
_mm_prefetch(((char *)(_src + 11)), _MM_HINT_NTA);
v0 = _mm_load_si128(_src + 0);
v1 = _mm_load_si128(_src + 1);
v2 = _mm_load_si128(_src + 2);
v3 = _mm_load_si128(_src + 3);
v4 = _mm_load_si128(_src + 4);
v5 = _mm_load_si128(_src + 5);
v6 = _mm_load_si128(_src + 6);
v7 = _mm_load_si128(_src + 7);
_mm_stream_si128(_dst + 0, v0);
_mm_stream_si128(_dst + 1, v1);
_mm_stream_si128(_dst + 2, v2);
_mm_stream_si128(_dst + 3, v3);
_mm_stream_si128(_dst + 4, v4);
_mm_stream_si128(_dst + 5, v5);
_mm_stream_si128(_dst + 6, v6);
_mm_stream_si128(_dst + 7, v7);
}
const size_t remain = length - sselen;
switch (remain & ~0xF)
{
case 112: v0 = _mm_load_si128(_src++);
case 96: v1 = _mm_load_si128(_src++);
case 80: v2 = _mm_load_si128(_src++);
case 64: v3 = _mm_load_si128(_src++);
case 48: v4 = _mm_load_si128(_src++);
case 32: v5 = _mm_load_si128(_src++);
case 16: v6 = _mm_load_si128(_src++);
}
switch (remain & ~0xF)
{
case 112: _mm_stream_si128(_dst++, v0);
case 96: _mm_stream_si128(_dst++, v1);
case 80: _mm_stream_si128(_dst++, v2);
case 64: _mm_stream_si128(_dst++, v3);
case 48: _mm_stream_si128(_dst++, v4);
case 32: _mm_stream_si128(_dst++, v5);
case 16: _mm_stream_si128(_dst++, v6);
}
// copy any remaining data
if (remain & 0xF)
{
uint8_t * rsrc = (uint8_t *)_src;
uint8_t * rdst = (uint8_t *)_dst;
for (size_t i = remain & 0xF; i > 0; --i)
*rdst++ = *rsrc++;
}
}
#else
#define memcpySSE memcpy
#endif