Merge pull request #2568 from Sonicadvance1/aarch64_crc32_hash
[AArch64] Implement CRC32 texture hashing.
This commit is contained in:
commit
42d5b5b48c
|
@ -239,8 +239,14 @@ void CPUInfo::Detect()
|
||||||
if (GetCPUImplementer() == 0x51 && GetCPUPart() == 0x6F) // Krait(300) is 0x6F, Scorpion is 0x4D
|
if (GetCPUImplementer() == 0x51 && GetCPUPart() == 0x6F) // Krait(300) is 0x6F, Scorpion is 0x4D
|
||||||
bIDIVa = bIDIVt = true;
|
bIDIVa = bIDIVt = true;
|
||||||
// These two require ARMv8 or higher
|
// These two require ARMv8 or higher
|
||||||
|
#ifdef _M_ARM_64
|
||||||
bFP = CheckCPUFeature("fp");
|
bFP = CheckCPUFeature("fp");
|
||||||
bASIMD = CheckCPUFeature("asimd");
|
bASIMD = CheckCPUFeature("asimd");
|
||||||
|
bAES = CheckCPUFeature("aes");
|
||||||
|
bCRC32 = CheckCPUFeature("crc32");
|
||||||
|
bSHA1 = CheckCPUFeature("sha1");
|
||||||
|
bSHA2 = CheckCPUFeature("sha2");
|
||||||
|
#endif
|
||||||
#endif
|
#endif
|
||||||
// On android, we build a separate library for ARMv7 so this is fine.
|
// On android, we build a separate library for ARMv7 so this is fine.
|
||||||
// TODO: Check for ARMv7 on other platforms.
|
// TODO: Check for ARMv7 on other platforms.
|
||||||
|
|
|
@ -75,6 +75,9 @@ struct CPUInfo
|
||||||
// ARMv8 specific
|
// ARMv8 specific
|
||||||
bool bFP;
|
bool bFP;
|
||||||
bool bASIMD;
|
bool bASIMD;
|
||||||
|
bool bCRC32;
|
||||||
|
bool bSHA1;
|
||||||
|
bool bSHA2;
|
||||||
|
|
||||||
// Call Detect()
|
// Call Detect()
|
||||||
explicit CPUInfo();
|
explicit CPUInfo();
|
||||||
|
|
|
@ -228,7 +228,7 @@ u64 GetMurmurHash3(const u8 *src, u32 len, u32 samples)
|
||||||
// CRC32 hash using the SSE4.2 instruction
|
// CRC32 hash using the SSE4.2 instruction
|
||||||
u64 GetCRC32(const u8 *src, u32 len, u32 samples)
|
u64 GetCRC32(const u8 *src, u32 len, u32 samples)
|
||||||
{
|
{
|
||||||
#if _M_SSE >= 0x402
|
#if _M_SSE >= 0x402 || defined(_M_ARM_64)
|
||||||
u64 h[4] = { len, 0, 0, 0 };
|
u64 h[4] = { len, 0, 0, 0 };
|
||||||
u32 Step = (len / 8);
|
u32 Step = (len / 8);
|
||||||
const u64 *data = (const u64 *)src;
|
const u64 *data = (const u64 *)src;
|
||||||
|
@ -238,6 +238,9 @@ u64 GetCRC32(const u8 *src, u32 len, u32 samples)
|
||||||
Step = Step / samples;
|
Step = Step / samples;
|
||||||
if (Step < 1)
|
if (Step < 1)
|
||||||
Step = 1;
|
Step = 1;
|
||||||
|
#endif
|
||||||
|
|
||||||
|
#if _M_SSE >= 0x402
|
||||||
while (data < end - Step * 3)
|
while (data < end - Step * 3)
|
||||||
{
|
{
|
||||||
h[0] = _mm_crc32_u64(h[0], data[Step * 0]);
|
h[0] = _mm_crc32_u64(h[0], data[Step * 0]);
|
||||||
|
@ -259,12 +262,66 @@ u64 GetCRC32(const u8 *src, u32 len, u32 samples)
|
||||||
memcpy(&temp, end, len & 7);
|
memcpy(&temp, end, len & 7);
|
||||||
h[0] = _mm_crc32_u64(h[0], temp);
|
h[0] = _mm_crc32_u64(h[0], temp);
|
||||||
}
|
}
|
||||||
|
#elif defined(_M_ARM_64)
|
||||||
|
// We should be able to use intrinsics for this
|
||||||
|
// Too bad the intrinsics for this instruction was added in GCC 4.9.1
|
||||||
|
// The Android NDK (as of r10e) only has GCC 4.9
|
||||||
|
// Once the Android NDK has a newer GCC version, update these to use intrinsics
|
||||||
|
while (data < end - Step * 3)
|
||||||
|
{
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[0])
|
||||||
|
: [two] "r" (h[0]),
|
||||||
|
[three] "r" (data[Step * 0]));
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[1])
|
||||||
|
: [two] "r" (h[1]),
|
||||||
|
[three] "r" (data[Step * 1]));
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[2])
|
||||||
|
: [two] "r" (h[2]),
|
||||||
|
[three] "r" (data[Step * 2]));
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[3])
|
||||||
|
: [two] "r" (h[3]),
|
||||||
|
[three] "r" (data[Step * 3]));
|
||||||
|
|
||||||
|
data += Step * 4;
|
||||||
|
}
|
||||||
|
if (data < end - Step * 0)
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[0])
|
||||||
|
: [two] "r" (h[0]),
|
||||||
|
[three] "r" (data[Step * 0]));
|
||||||
|
if (data < end - Step * 1)
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[1])
|
||||||
|
: [two] "r" (h[1]),
|
||||||
|
[three] "r" (data[Step * 1]));
|
||||||
|
if (data < end - Step * 2)
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[2])
|
||||||
|
: [two] "r" (h[2]),
|
||||||
|
[three] "r" (data[Step * 2]));
|
||||||
|
|
||||||
|
if (len & 7)
|
||||||
|
{
|
||||||
|
u64 temp = 0;
|
||||||
|
memcpy(&temp, end, len & 7);
|
||||||
|
asm ("crc32x %w[res], %w[two], %x[three]"
|
||||||
|
: [res] "=r" (h[0])
|
||||||
|
: [two] "r" (h[0]),
|
||||||
|
[three] "r" (temp));
|
||||||
|
}
|
||||||
|
#endif
|
||||||
|
|
||||||
|
#if _M_SSE >= 0x402 || defined(_M_ARM_64)
|
||||||
// FIXME: is there a better way to combine these partial hashes?
|
// FIXME: is there a better way to combine these partial hashes?
|
||||||
return h[0] + (h[1] << 10) + (h[2] << 21) + (h[3] << 32);
|
return h[0] + (h[1] << 10) + (h[2] << 21) + (h[3] << 32);
|
||||||
#else
|
#else
|
||||||
return 0;
|
return 0;
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
|
@ -532,6 +589,12 @@ void SetHash64Function()
|
||||||
ptrHashFunction = &GetCRC32;
|
ptrHashFunction = &GetCRC32;
|
||||||
}
|
}
|
||||||
else
|
else
|
||||||
|
#elif defined(_M_ARM_64)
|
||||||
|
if (cpu_info.bCRC32)
|
||||||
|
{
|
||||||
|
ptrHashFunction = &GetCRC32;
|
||||||
|
}
|
||||||
|
else
|
||||||
#endif
|
#endif
|
||||||
{
|
{
|
||||||
ptrHashFunction = &GetMurmurHash3;
|
ptrHashFunction = &GetMurmurHash3;
|
||||||
|
|
Loading…
Reference in New Issue