archive/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-07-14 18:58:36 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

SPU DMA: more tuning for mov_rdata_avx

Browse source 
Avoid unaligned stores.
Prefer asm path if __AVX2__ is not set.
Don't emit vzeroupper if __AVX__ is set.
This commit is contained in:
Nekotekina 2020-04-27 17:46:15 +03:00
parent 4f71c570bd
commit 3ec73b651e
2 changed files with 45 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

51
rpcs3/Emu/Cell/SPUThread.cpp
View file

@ -43,11 +43,12 @@ static FORCE_INLINE bool cmp_rdata(const decltype(spu_thread::rdata)& lhs, const
static FORCE_INLINE void mov_rdata_avx(__m256i* dst, const __m256i* src) static FORCE_INLINE void mov_rdata_avx(__m256i* dst, const __m256i* src)
{ {
#if defined(_MSC_VER) || defined(__AVX__) #if defined(_MSC_VER) || defined(__AVX2__)
_mm256_storeu_si256(dst + 0, _mm256_loadu_si256(src + 0)); // In AVX-only mode, for some older CPU models, GCC/Clang may emit 128-bit loads/stores instead.
_mm256_storeu_si256(dst + 1, _mm256_loadu_si256(src + 1)); _mm256_store_si256(dst + 0, _mm256_loadu_si256(src + 0));
_mm256_storeu_si256(dst + 2, _mm256_loadu_si256(src + 2)); _mm256_store_si256(dst + 1, _mm256_loadu_si256(src + 1));
_mm256_storeu_si256(dst + 3, _mm256_loadu_si256(src + 3)); _mm256_store_si256(dst + 2, _mm256_loadu_si256(src + 2));
_mm256_store_si256(dst + 3, _mm256_loadu_si256(src + 3));
#else #else
__asm__( __asm__(
"vmovdqu 0*32(%[src]), %%ymm0;" // load "vmovdqu 0*32(%[src]), %%ymm0;" // load
@ -58,11 +59,17 @@ static FORCE_INLINE void mov_rdata_avx(__m256i* dst, const __m256i* src)
"vmovdqu %%ymm0, 2*32(%[dst]);" "vmovdqu %%ymm0, 2*32(%[dst]);"
"vmovdqu 3*32(%[src]), %%ymm0;" "vmovdqu 3*32(%[src]), %%ymm0;"
"vmovdqu %%ymm0, 3*32(%[dst]);" "vmovdqu %%ymm0, 3*32(%[dst]);"
"vzeroupper" #ifndef __AVX__
"vzeroupper" // Don't need in AVX mode (should be emitted automatically)
#endif
: :
: [src] "r" (src) : [src] "r" (src)
, [dst] "r" (dst) , [dst] "r" (dst)
: "xmm0" #ifdef __AVX__
: "ymm0" // Clobber ymm0 register (acknowledge its modification)
#else
: "xmm0" // ymm0 is "unknown" if not compiled in AVX mode, so clobber xmm0 only
#endif
); );
#endif #endif
} }
@ -1512,6 +1519,16 @@ void spu_thread::do_dma_transfer(const spu_mfc_cmd& args)
// Split locking + transfer in two parts (before 64K border, and after it) // Split locking + transfer in two parts (before 64K border, and after it)
const auto lock = vm::range_lock(range_addr, nexta); const auto lock = vm::range_lock(range_addr, nexta);
// Avoid unaligned stores in mov_rdata_avx
if (reinterpret_cast<u64>(dst) & 0x10)
{
*reinterpret_cast<v128*>(dst) = *reinterpret_cast<const v128*>(src);
dst += 16;
src += 16;
size0 -= 16;
}
while (size0 >= 128) while (size0 >= 128)
{ {
mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src)); mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src));
@ -1536,6 +1553,16 @@ void spu_thread::do_dma_transfer(const spu_mfc_cmd& args)
const auto lock = vm::range_lock(range_addr, range_end); const auto lock = vm::range_lock(range_addr, range_end);
// Avoid unaligned stores in mov_rdata_avx
if (reinterpret_cast<u64>(dst) & 0x10)
{
*reinterpret_cast<v128*>(dst) = *reinterpret_cast<const v128*>(src);
dst += 16;
src += 16;
size -= 16;
}
while (size >= 128) while (size >= 128)
{ {
mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src)); mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src));
@ -1586,6 +1613,16 @@ void spu_thread::do_dma_transfer(const spu_mfc_cmd& args)
} }
default: default:
{ {
// Avoid unaligned stores in mov_rdata_avx
if (reinterpret_cast<u64>(dst) & 0x10)
{
*reinterpret_cast<v128*>(dst) = *reinterpret_cast<const v128*>(src);
dst += 16;
src += 16;
size -= 16;
}
while (size >= 128) while (size >= 128)
{ {
mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src)); mov_rdata(*reinterpret_cast<decltype(spu_thread::rdata)*>(dst), *reinterpret_cast<const decltype(spu_thread::rdata)*>(src));

2
rpcs3/Emu/Cell/SPUThread.h
View file

@ -556,7 +556,7 @@ public:
// Reservation Data // Reservation Data
u64 rtime = 0; u64 rtime = 0;
std::array<v128, 8> rdata{}; alignas(64) std::array<v128, 8> rdata{};
u32 raddr = 0; u32 raddr = 0;
u32 srr0; u32 srr0;