RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2026-10-11 22:49:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

BufferUtils: Address reviews

Browse Source 
-drop iota16, remove avx2 check, better reduction
This commit is contained in:
Malcolm
2026-10-10 12:40:20 +03:00
committed by kd-11
parent cd4d4e02ec
commit f9b8c209a3
3 changed files with 21 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rpcs3/Emu/RSX/Common/BufferUtils.cpp
+2 -5
View File
@@ -470,7 +470,7 @@ namespace
#if defined(ARCH_X64)
const upload_untouched_skip_restart_dispatch s_avx512_upload_untouched_skip_restart_dispatch =
{
s_use_avx2 ? upload_swapped_avx2_skip_restart<u16> : upload_untouched_skip_restart<u16>,
upload_swapped_avx2_skip_restart<u16>,
upload_u32_swapped_avx3_skip_restart,
};
@@ -748,13 +748,10 @@ void iota16(u16* dst, u32 count)
iota16_avx2(dst, count);
return;
}
#elif defined(ARCH_ARM64)
iota16_neon(dst, count);
return;
#endif
u32 i = 0;
#if defined(ARCH_X64)
#if defined(ARCH_X64) || defined(ARCH_ARM64)
// Force 16-byte alignment
const uptr mem_addr = reinterpret_cast<uptr>(dst);
const u32 head = std::min<u32>(count, ((0 - mem_addr) & 15) / sizeof(u16));
rpcs3/Emu/RSX/Common/BufferUtils_compact.h
+19 -6
View File
@@ -165,12 +165,25 @@ namespace
}
}
alignas(32) T min_lanes[lanes];
alignas(32) T max_lanes[lanes];
_mm256_store_si256(reinterpret_cast<__m256i*>(min_lanes), min);
_mm256_store_si256(reinterpret_cast<__m256i*>(max_lanes), max);
T min_index = *std::min_element(std::begin(min_lanes), std::end(min_lanes));
T max_index = *std::max_element(std::begin(max_lanes), std::end(max_lanes));
T min_index, max_index;
if constexpr (sizeof(T) == 2)
{
const auto min128 = _mm_min_epu16(_mm256_castsi256_si128(min), _mm256_extracti128_si256(min, 1));
const auto max128 = _mm_max_epu16(_mm256_castsi256_si128(max), _mm256_extracti128_si256(max, 1));
min_index = static_cast<T>(_mm_cvtsi128_si32(_mm_minpos_epu16(min128)));
max_index = static_cast<T>(~_mm_cvtsi128_si32(_mm_minpos_epu16(_mm_xor_si128(max128, _mm_set1_epi32(-1)))));
}
else
{
auto min128 = _mm_min_epu32(_mm256_castsi256_si128(min), _mm256_extracti128_si256(min, 1));
auto max128 = _mm_max_epu32(_mm256_castsi256_si128(max), _mm256_extracti128_si256(max, 1));
min128 = _mm_min_epu32(min128, _mm_shuffle_epi32(min128, _MM_SHUFFLE(1, 0, 3, 2)));
max128 = _mm_max_epu32(max128, _mm_shuffle_epi32(max128, _MM_SHUFFLE(1, 0, 3, 2)));
min128 = _mm_min_epu32(min128, _mm_shuffle_epi32(min128, _MM_SHUFFLE(2, 3, 0, 1)));
max128 = _mm_max_epu32(max128, _mm_shuffle_epi32(max128, _MM_SHUFFLE(2, 3, 0, 1)));
min_index = static_cast<T>(_mm_cvtsi128_si32(min128));
max_index = static_cast<T>(_mm_cvtsi128_si32(max128));
}
for (; i < count; ++i)
{
rpcs3/Emu/RSX/Common/BufferUtils_neon.h
-18
View File
@@ -122,22 +122,4 @@ namespace
return (u64{max_index} << 32) | min_index;
}
void iota16_neon(u16* dst, u32 count)
{
const u16 initial[] = {0, 1, 2, 3, 4, 5, 6, 7};
auto value = vld1q_u16(initial);
u32 i = 0;
for (; count - i >= 8; i += 8)
{
vst1q_u16(dst + i, value);
value = vaddq_u16(value, vdupq_n_u16(8));
}
for (; i < count; ++i)
{
dst[i] = static_cast<u16>(i);
}
}
} // namespace