Skip to content

Instantly share code, notes, and snippets.

@castano
Created March 21, 2012 18:27
Show Gist options
  • Save castano/2150795 to your computer and use it in GitHub Desktop.
Save castano/2150795 to your computer and use it in GitHub Desktop.
ISPC half to float conversion
#include <intrin.h> // __rdtsc
#include <stdio.h>
typedef unsigned int uint;
typedef unsigned short uint16;
typedef unsigned long long uint64;
// ISPC functions:
extern "C" void half_to_float_test(const uint16 * vin, float * vout, int count);
extern "C" void half_to_float_test_fast(const uint16 * vin, float * vout, int count);
extern "C" void half_to_float_test_ryg(const uint16 * vin, float * vout, int count);
extern "C" void half_to_float_test_fast_ryg(const uint16 * vin, float * vout, int count);
union FP32
{
uint u;
float f;
struct
{
uint Mantissa : 23;
uint Exponent : 8;
uint Sign : 1;
};
};
union FP16
{
unsigned short u;
struct
{
uint Mantissa : 10;
uint Exponent : 5;
uint Sign : 1;
};
};
static FP32 half_to_float_full(FP16 h)
{
FP32 o = { 0 };
// From ISPC ref code
if (h.Exponent == 0 && h.Mantissa == 0) // (Signed) zero
o.Sign = h.Sign;
else
{
if (h.Exponent == 0) // Denormal (will convert to normalized)
{
// Adjust mantissa so it's normalized (and keep track of exp adjust)
int e = -1;
uint m = h.Mantissa;
do
{
e++;
m <<= 1;
} while ((m & 0x400) == 0);
o.Mantissa = (m & 0x3ff) << 13;
o.Exponent = 127 - 15 - e;
o.Sign = h.Sign;
}
else if (h.Exponent == 0x1f) // Inf/NaN
{
// NOTE: It's safe to treat both with the same code path by just truncating
// lower Mantissa bits in NaNs (this is valid).
o.Mantissa = h.Mantissa << 13;
o.Exponent = 255;
o.Sign = h.Sign;
}
else // Normalized number
{
o.Mantissa = h.Mantissa << 13;
o.Exponent = 127 - 15 + h.Exponent;
o.Sign = h.Sign;
}
}
return o;
}
static __m128 half_to_float4_SSE2(__m128i h)
{
#define SSE_CONST4(name, val) static const __declspec(align(16)) uint name[4] = { (val), (val), (val), (val) }
#define CONST(name) *(const __m128i *)&name
SSE_CONST4(mask_nosign, 0x7fff);
SSE_CONST4(mask_justsign, 0x8000);
SSE_CONST4(mask_shifted_exp, 0x7c00 << 13);
SSE_CONST4(expadjust_normal, (127 - 15) << 23);
SSE_CONST4(expadjust_infnan, (128 - 16) << 23);
SSE_CONST4(expadjust_denorm, 1 << 23);
SSE_CONST4(magic_denorm, 113 << 23);
__m128i mnosign = CONST(mask_nosign);
__m128i expmant = _mm_and_si128(mnosign, h);
__m128i justsign = _mm_and_si128(h, CONST(mask_justsign));
__m128i mshiftexp = CONST(mask_shifted_exp);
__m128i eadjust = CONST(expadjust_normal);
__m128i shifted = _mm_slli_epi32(expmant, 13);
__m128i adjusted = _mm_add_epi32(eadjust, shifted);
__m128i justexp = _mm_and_si128(shifted, mshiftexp);
__m128i zero = _mm_setzero_si128();
__m128i b_isinfnan = _mm_cmpeq_epi32(mshiftexp, justexp);
__m128i b_isdenorm = _mm_cmpeq_epi32(zero, justexp);
__m128i adj_infnan = _mm_and_si128(b_isinfnan, CONST(expadjust_infnan));
__m128i adjusted2 = _mm_add_epi32(adjusted, adj_infnan);
__m128i adj_den = CONST(expadjust_denorm);
__m128i den1 = _mm_add_epi32(adj_den, adjusted2);
__m128 den2 = _mm_sub_ps(_mm_castsi128_ps(den1), *(const __m128 *)&magic_denorm);
__m128 adjusted3 = _mm_and_ps(den2, _mm_castsi128_ps(b_isdenorm));
__m128 adjusted4 = _mm_andnot_ps(_mm_castsi128_ps(b_isdenorm), _mm_castsi128_ps(adjusted2));
__m128 adjusted5 = _mm_or_ps(adjusted3, adjusted4);
__m128i sign = _mm_slli_epi32(justsign, 16);
__m128 final = _mm_or_ps(adjusted5, _mm_castsi128_ps(sign));
// ~21 SSE2 ops.
return final;
#undef SSE_CONST4
#undef CONST
}
void half_to_float_test_sse2(const uint16 * vin, float * vout, int count) {
__m128i zero = _mm_setzero_si128();
for (int i = 0; i < count; i += 8)
{
__m128i in = _mm_loadu_si128((const __m128i *)(vin + i));
__m128i a = _mm_unpacklo_epi16(in, zero);
__m128i b = _mm_unpackhi_epi16(in, zero);
__m128 outa = half_to_float4_SSE2(a);
_mm_storeu_ps((float *)(vout + i), outa);
__m128 outb = half_to_float4_SSE2(b);
_mm_storeu_ps((float *)(vout + i + 4), outb);
}
}
int test_results(const uint16 * vin, const float * vout, int count) {
int error_count = 0;
for (int i = 0; i < count; i++)
{
FP16 in;
in.u = vin[i];
FP32 full = half_to_float_full(in);
FP32 out;
out.f = vout[i];
if (full.u != out.u) error_count++;
}
return error_count;
}
int main(void) {
uint16 vin[0x10000];
float vout[0x10000];
// Init input.
for (int i = 0; i < 0x10000; i++)
{
vin[i] = i;
}
// Run tests.
uint64 start, end;
half_to_float_test(vin, vout, 0x10000);
int error_count = test_results(vin, vout, 0x10000);
start = __rdtsc();
for (int i = 0; i < 64; i++) half_to_float_test(vin, vout, 0x10000);
end = __rdtsc();
printf("half_to_float: %.3f (%d)\n", double(end-start) / (1000*1000*4), error_count);
half_to_float_test_fast(vin, vout, 0x10000);
error_count = test_results(vin, vout, 0x10000);
start = __rdtsc();
for (int i = 0; i < 64; i++) half_to_float_test_fast(vin, vout, 0x10000);
end = __rdtsc();
printf("half_to_float_fast: %.3f (%d)\n", double(end-start) / (1000*1000*4), error_count);
half_to_float_test_ryg(vin, vout, 0x10000);
error_count = test_results(vin, vout, 0x10000);
start = __rdtsc();
for (int i = 0; i < 64; i++) half_to_float_test_ryg(vin, vout, 0x10000);
end = __rdtsc();
printf("half_to_float_ryg: %.3f (%d)\n", double(end-start) / (1000*1000*4), error_count);
half_to_float_test_fast_ryg(vin, vout, 0x10000);
error_count = test_results(vin, vout, 0x10000);
start = __rdtsc();
for (int i = 0; i < 64; i++) half_to_float_test_fast_ryg(vin, vout, 0x10000);
end = __rdtsc();
printf("half_to_float_fast_ryg: %.3f (%d)\n", double(end-start) / (1000*1000*4), error_count);
half_to_float_test_sse2(vin, vout, 0x10000);
error_count = test_results(vin, vout, 0x10000);
start = __rdtsc();
for (int i = 0; i < 64; i++) half_to_float_test_sse2(vin, vout, 0x10000);
end = __rdtsc();
printf("half_to_float_sse2: %.3f (%d)\n", double(end-start) / (1000*1000*4), error_count);
return 0;
}
inline float half_to_float_ryg(unsigned int16 h)
{
if (__have_native_half) {
return __half_to_float_varying(h);
}
else {
unsigned int32 hs = h & (int32)0x8000u; // Pick off sign bit
unsigned int32 hem = h & (int32)0x7fffu; // Pick off exponent-mantissa bits
unsigned int32 xs = ((unsigned int32) hs) << 16;
unsigned int32 xem = ((unsigned int32) hem) << 13;
unsigned int32 xe = xem & 0xF800000; // & (0x7c00 << 13)
xem += 0x38000000; // (127 - 15) << 23
// handle exponent special cases
if (xe == 0xF800000) // Inf/NaN?
{
xem += 0x38000000; // extra exp adjust
}
else if (xe == 0) // Zero/Denormal?
{
xem += 0x800000; // extra exp adjust
xem = intbits(floatbits(xem) - floatbits(0x38800000)); // renormalize
}
return floatbits(xs | xem);
}
}
inline float half_to_float_fast_ryg(unsigned int16 h)
{
if (__have_native_half) {
return __half_to_float_varying(h);
}
else {
unsigned int32 hs = h & (int32)0x8000u; // Pick off sign bit
unsigned int32 hem = h & (int32)0x7fffu; // Pick off exponent-mantissa bits
unsigned int32 xs = ((unsigned int32) hs) << 16;
unsigned int32 xem = ((unsigned int32) hem) << 13;
xem += 0x38000000; // (127 - 15) << 23
return floatbits(xs | xem);
}
}
export void half_to_float_test(uniform const unsigned int16 vin[], uniform float vout[], uniform int count)
{
foreach(i = 0 ... count) {
vout[i] = half_to_float(vin[i]);
}
}
export void half_to_float_test_fast(uniform const unsigned int16 vin[], uniform float vout[], uniform int count)
{
foreach(i = 0 ... count) {
vout[i] = half_to_float_fast(vin[i]);
}
}
export void half_to_float_test_ryg(uniform const unsigned int16 vin[], uniform float vout[], uniform int count)
{
foreach(i = 0 ... count) {
vout[i] = half_to_float_ryg(vin[i]);
}
}
export void half_to_float_test_fast_ryg(uniform const unsigned int16 vin[], uniform float vout[], uniform int count)
{
foreach(i = 0 ... count) {
vout[i] = half_to_float_fast_ryg(vin[i]);
}
}
Reference:
half_to_float_sse2: 2.306
ISPC targeting SSE2:
half_to_float: 6.668
half_to_float_fast: 2.554
half_to_float_ryg: 4.390
half_to_float_fast_ryg: 2.005
ISPC targeting SSE4:
half_to_float: 4.776
half_to_float_fast: 1.939
half_to_float_ryg: 3.121
half_to_float_fast_ryg: 1.448
ISPC targeting AVX:
half_to_float: 4.368
half_to_float_fast: 1.857
half_to_float_ryg: 2.791
half_to_float_fast_ryg: 1.275
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment