/* evaluation of 4 sines at onces, using only SSE2.
The code is the exact rewriting of the cephes sinf function.
Precision is excellent as long as x < 8192 (I did not bother to
take into account the special handling they have for greater values
-- it does not return garbage for arguments over 8192, though, but
the extra precision is missing).
Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
surprising but correct result.
Performance is also surprisingly good, 1.33 times faster than the
macos vsinf SSE2 function, and 1.5 times faster than the
__vrs4_sinf of amd's ACML (which is only available in 64 bits). Not
too bad for an SSE1 function (with no special tuning) !
However the latter libraries probably have a much better handling of NaN,
Inf, denormalized and other special arguments..
On my core 1 duo, the execution of this function takes approximately 95 cycles.
From what I have observed on the experiments with Intel AMath lib, switching to an
SSE2 version would improve the perf by only 10%.
Since it is based on SSE intrinsics, it has to be compiled at -O2 to
deliver full speed.
*/
__m128 sin_ps(__m128 x) { // any x
typedef __m128 v4sf;
typedef __m128i v4si;
v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y;
v4si emm0, emm2;
sign_bit = x;
/* take the absolute value */
x = _mm_and_ps(x, constants::inv_mant_mask.ps);
/* extract the sign bit (upper one) */
sign_bit = _mm_and_ps(sign_bit, constants::sign_mask.ps);
/* scale by 4/Pi */
y = _mm_mul_ps(x, constants::cephes_FOPI.ps);
/* store the integer part of y in mm0 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, constants::pi32_1.pi);
emm2 = _mm_and_si128(emm2, constants::pi32_inv1.pi);
y = _mm_cvtepi32_ps(emm2);
/* get the swap sign flag */
emm0 = _mm_and_si128(emm2, constants::pi32_4.pi);
emm0 = _mm_slli_epi32(emm0, 29);
/* get the polynom selection mask
there is one polynom for 0 <= x <= Pi/4
and another one for Pi/4<x<=Pi/2
Both branches will be computed.
*/
emm2 = _mm_and_si128(emm2, constants::pi32_2.pi);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
v4sf swap_sign_bit = _mm_castsi128_ps(emm0);
v4sf poly_mask = _mm_castsi128_ps(emm2);
sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = constants::minus_cephes_DP1.ps;
xmm2 = constants::minus_cephes_DP2.ps;
xmm3 = constants::minus_cephes_DP3.ps;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = constants::coscof_p0.ps;
v4sf z = _mm_mul_ps(x,x);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, constants::coscof_p1.ps);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, constants::coscof_p2.ps);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
v4sf tmp = _mm_mul_ps(z, constants::ps_0p5.ps);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, constants::ps_1.ps);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v4sf y2 = constants::sincof_p0.ps;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, constants::sincof_p1.ps);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, constants::sincof_p2.ps);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
y = _mm_andnot_ps(xmm3, y);
y = _mm_add_ps(y,y2);
/* update the sign */
y = _mm_xor_ps(y, sign_bit);
return y;
}
inline GPR_t si_orhi( GPR_t RA, int64_t IMM )
{
return _mm_or_ps( RA, _mm_castsi128_ps( _mm_set1_epi16( (int16_t)IMM ) ) );
}
inline GPR_t si_orc( GPR_t RA, GPR_t RB )
{
const __m128 not_RB = _mm_andnot_ps( RB, _mm_castsi128_ps( _mm_set1_epi32(0xffffffff) ) );
return _mm_or_ps( RA, not_RB );
}
namespace embree {
const __m128 _mm_lookupmask_ps[16] = {
_mm_castsi128_ps(_mm_set_epi32( 0, 0, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1,-1,-1))
};
const __m128d _mm_lookupmask_pd[4] = {
_mm_castsi128_pd(_mm_set_epi32( 0, 0, 0, 0)),
_mm_castsi128_pd(_mm_set_epi32( 0, 0,-1,-1)),
_mm_castsi128_pd(_mm_set_epi32(-1,-1, 0, 0)),
_mm_castsi128_pd(_mm_set_epi32(-1,-1,-1,-1))
};
}
namespace q {
const __m128 _mm_lookupmask_ps[16] = {
_mm_castsi128_ps(_mm_set_epi32( 0, 0, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0, 0,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32( 0,-1,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1, 0,-1,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1, 0, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1, 0,-1)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1,-1, 0)),
_mm_castsi128_ps(_mm_set_epi32(-1,-1,-1,-1))
};
} /* namespace q */
inline GPR_t si_xorbi( GPR_t RA, int64_t IMM )
{
return _mm_xor_ps( RA, _mm_castsi128_ps( _mm_set1_epi8((uint8_t)IMM) ) );
}
GSVector4 GSVector4::cast(const GSVector4i& v)
{
return GSVector4(_mm_castsi128_ps(v.m));
}
/* Function: esl_sse_logf()
* Synopsis: <r[z] = log x[z]>
* Incept: SRE, Fri Dec 14 11:32:54 2007 [Janelia]
*
* Purpose: Given a vector <x> containing four floats, returns a
* vector <r> in which each element <r[z] = logf(x[z])>.
*
* Valid in the domain $x_z > 0$ for normalized IEEE754
* $x_z$.
*
* For <x> $< 0$, including -0, returns <NaN>. For <x> $==
* 0$ or subnormal <x>, returns <-inf>. For <x = inf>,
* returns <inf>. For <x = NaN>, returns <NaN>. For
* subnormal <x>, returns <-inf>.
*
* Xref: J2/71.
*
* Note: Derived from an SSE1 implementation by Julian
* Pommier. Converted to SSE2 and added handling
* of IEEE754 specials.
*/
__m128
esl_sse_logf(__m128 x)
{
static float cephes_p[9] = { 7.0376836292E-2f, -1.1514610310E-1f, 1.1676998740E-1f,
-1.2420140846E-1f, 1.4249322787E-1f, -1.6668057665E-1f,
2.0000714765E-1f, -2.4999993993E-1f, 3.3333331174E-1f };
__m128 onev = _mm_set1_ps(1.0f); /* all elem = 1.0 */
__m128 v0p5 = _mm_set1_ps(0.5f); /* all elem = 0.5 */
__m128i vneg = _mm_set1_epi32(0x80000000); /* all elem have IEEE sign bit up */
__m128i vexp = _mm_set1_epi32(0x7f800000); /* all elem have IEEE exponent bits up */
__m128i ei;
__m128 e;
__m128 invalid_mask, zero_mask, inf_mask; /* masks used to handle special IEEE754 inputs */
__m128 mask;
__m128 origx;
__m128 tmp;
__m128 y;
__m128 z;
/* first, split x apart: x = frexpf(x, &e); */
ei = _mm_srli_epi32( _mm_castps_si128(x), 23); /* shift right 23: IEEE754 floats: ei = biased exponents */
invalid_mask = _mm_castsi128_ps ( _mm_cmpeq_epi32( _mm_and_si128(_mm_castps_si128(x), vneg), vneg)); /* mask any elem that's negative; these become NaN */
zero_mask = _mm_castsi128_ps ( _mm_cmpeq_epi32(ei, _mm_setzero_si128())); /* mask any elem zero or subnormal; these become -inf */
inf_mask = _mm_castsi128_ps ( _mm_cmpeq_epi32( _mm_and_si128(_mm_castps_si128(x), vexp), vexp)); /* mask any elem inf or NaN; log(inf)=inf, log(NaN)=NaN */
origx = x; /* store original x, used for log(inf) = inf, log(NaN) = NaN */
x = _mm_and_ps(x, _mm_castsi128_ps(_mm_set1_epi32(~0x7f800000))); /* x now the stored 23 bits of the 24-bit significand */
x = _mm_or_ps (x, v0p5); /* sets hidden bit b[0] */
ei = _mm_sub_epi32(ei, _mm_set1_epi32(126)); /* -127 (ei now signed base-2 exponent); then +1 */
e = _mm_cvtepi32_ps(ei);
/* now, calculate the log */
mask = _mm_cmplt_ps(x, _mm_set1_ps(0.707106781186547524f)); /* avoid conditional branches. */
tmp = _mm_and_ps(x, mask); /* tmp contains x values < 0.707, else 0 */
x = _mm_sub_ps(x, onev);
e = _mm_sub_ps(e, _mm_and_ps(onev, mask));
x = _mm_add_ps(x, tmp);
z = _mm_mul_ps(x,x);
y = _mm_set1_ps(cephes_p[0]); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[1])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[2])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[3])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[4])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[5])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[6])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[7])); y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[8])); y = _mm_mul_ps(y, x);
y = _mm_mul_ps(y, z);
tmp = _mm_mul_ps(e, _mm_set1_ps(-2.12194440e-4f));
y = _mm_add_ps(y, tmp);
tmp = _mm_mul_ps(z, v0p5);
y = _mm_sub_ps(y, tmp);
tmp = _mm_mul_ps(e, _mm_set1_ps(0.693359375f));
x = _mm_add_ps(x, y);
x = _mm_add_ps(x, tmp);
/* IEEE754 cleanup: */
x = esl_sse_select_ps(x, origx, inf_mask); /* log(inf)=inf; log(NaN) = NaN */
x = _mm_or_ps(x, invalid_mask); /* log(x<0, including -0,-inf) = NaN */
x = esl_sse_select_ps(x, _mm_set1_ps(-eslINFINITY), zero_mask); /* x zero or subnormal = -inf */
return x;
}
static void SinCos(const float rad, float &sin, float &cos) // #include <emmintrin.h>, #include <xmmintrin.h>
{
const __m128 _ps_fopi = _mm_set1_ps(4.0f / pi);
const __m128 _ps_0p5 = _mm_set1_ps(0.5f);
const __m128 _ps_1 = _mm_set1_ps(1.0f);
const __m128 _ps_dp1 = _mm_set1_ps(-0.7851562f);
const __m128 _ps_dp2 = _mm_set1_ps(-2.4187564849853515625e-4f);
const __m128 _ps_dp3 = _mm_set1_ps(-3.77489497744594108e-8f);
const __m128 _ps_sincof_p0 = _mm_set1_ps(2.443315711809948e-5f);
const __m128 _ps_sincof_p1 = _mm_set1_ps(8.3321608736e-3f);
const __m128 _ps_sincof_p2 = _mm_set1_ps(-1.6666654611e-1f);
const __m128 _ps_coscof_p0 = _mm_set1_ps(2.443315711809948e-5f);
const __m128 _ps_coscof_p1 = _mm_set1_ps(-1.388731625493765e-3f);
const __m128 _ps_coscof_p2 = _mm_set1_ps(4.166664568298827e-2f);
const __m128i _pi32_1 = _mm_set1_epi32(1);
const __m128i _pi32_i1 = _mm_set1_epi32(~1);
const __m128i _pi32_2 = _mm_set1_epi32(2);
const __m128i _pi32_4 = _mm_set1_epi32(4);
const __m128 _mask_sign_raw = _mm_castsi128_ps(_mm_set1_epi32( 0x80000000));
const __m128 _mask_sign_inv = _mm_castsi128_ps(_mm_set1_epi32(~0x80000000));
__m128 mm1, mm2;
__m128i mmi0, mmi2, mmi4;
__m128 x, y, z;
__m128 y1, y2;
__m128 a = _mm_set1_ps(rad);
x = _mm_and_ps(a, _mask_sign_inv);
y = _mm_mul_ps(x, _ps_fopi);
mmi2 = _mm_cvtps_epi32(y);
mmi2 = _mm_add_epi32(mmi2, _pi32_1);
mmi2 = _mm_and_si128(mmi2, _pi32_i1);
y = _mm_cvtepi32_ps(mmi2);
mmi4 = mmi2;
mmi0 = _mm_and_si128(mmi2, _pi32_4);
mmi0 = _mm_slli_epi32(mmi0, 29);
__m128 swap_sign_bit_sin = _mm_castsi128_ps(mmi0);
mmi2 = _mm_and_si128(mmi2, _pi32_2);
mmi2 = _mm_cmpeq_epi32(mmi2, _mm_setzero_si128());
__m128 poly_mask = _mm_castsi128_ps(mmi2);
x = _mm_add_ps(x, _mm_mul_ps(y, _ps_dp1));
x = _mm_add_ps(x, _mm_mul_ps(y, _ps_dp2));
x = _mm_add_ps(x, _mm_mul_ps(y, _ps_dp3));
mmi4 = _mm_sub_epi32(mmi4, _pi32_2);
mmi4 = _mm_andnot_si128(mmi4, _pi32_4);
mmi4 = _mm_slli_epi32(mmi4, 29);
__m128 sign_bit_cos = _mm_castsi128_ps(mmi4);
__m128 sign_bit_sin = _mm_xor_ps(_mm_and_ps(a, _mask_sign_raw), swap_sign_bit_sin);
z = _mm_mul_ps(x, x);
y1 = _mm_mul_ps(_ps_coscof_p0, z);
y1 = _mm_add_ps(y1, _ps_coscof_p1);
y1 = _mm_mul_ps(y1, z);
y1 = _mm_add_ps(y1, _ps_coscof_p2);
y1 = _mm_mul_ps(y1, z);
y1 = _mm_mul_ps(y1, z);
y1 = _mm_sub_ps(y1, _mm_mul_ps(z, _ps_0p5));
y1 = _mm_add_ps(y1, _ps_1);
y2 = _mm_mul_ps(_ps_sincof_p0, z);
y2 = _mm_add_ps(y2, _ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, _ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
__m128 sin1y = _mm_andnot_ps(poly_mask, y1);
__m128 sin2y = _mm_and_ps(poly_mask, y2);
mm1 = _mm_add_ps(sin1y, sin2y);
mm2 = _mm_add_ps(_mm_sub_ps(y1, sin1y), _mm_sub_ps(y2, sin2y));
sin = _mm_cvtss_f32(_mm_xor_ps(mm1, sign_bit_sin));
cos = _mm_cvtss_f32(_mm_xor_ps(mm2, sign_bit_cos));
}
int main()
{
//Transpose
vec4 mat[4] = {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}, {13, 14, 15, 16}};
__m128i xmm0 = _mm_unpacklo_epi32(_mm_castps_si128(mat[0]), _mm_castps_si128(mat[1]));
__m128i xmm1 = _mm_unpackhi_epi32(_mm_castps_si128(mat[0]), _mm_castps_si128(mat[1]));
__m128i xmm2 = _mm_unpacklo_epi32(_mm_castps_si128(mat[2]), _mm_castps_si128(mat[3]));
__m128i xmm3 = _mm_unpackhi_epi32(_mm_castps_si128(mat[2]), _mm_castps_si128(mat[3]));
vec4 trans[4];
trans[0] = _mm_castsi128_ps(_mm_unpacklo_epi64(xmm0, xmm2));
trans[1] = _mm_castsi128_ps(_mm_unpackhi_epi64(xmm0, xmm2));
trans[2] = _mm_castsi128_ps(_mm_unpacklo_epi64(xmm1, xmm3));
trans[3] = _mm_castsi128_ps(_mm_unpackhi_epi64(xmm1, xmm3));
vec4 trans2[4];
ml::transpose(trans2, mat);
FILE* file = fopen("..\\..\\AppData\\VT.swf", "rb");
fseek(file, 0, SEEK_END);
size_t size = ftell(file);
fseek(file, 0, SEEK_SET);
unsigned char* fileData = (unsigned char*)malloc(size);
fread(fileData, 1, size, file);
fclose(file);
MemReader data = {(const char*)fileData, (const char*)fileData+size, (const char*)fileData};
//Read SWF header
const u32 signatureAndVersion = data.read<u32>();
const u32 actualSize = data.read<u32>();
u32 signature = signatureAndVersion&0x00FFFFFF;
u8 version = signatureAndVersion>>24;
bool isCompressed = signature=='\0SWC';
bool isUncompressed = signature=='\0SWF';
//if !isCompressed && !isUncompressed return error;
MemReader data2 = {0, 0, 0};
char* uncompressed = 0;
if (isCompressed)
{
uncompressed = (char*)malloc(actualSize-8);
data2.cur = data2.start = uncompressed;
data2.end = uncompressed+actualSize-8;
uLongf uncompressedSize = actualSize-8;
uncompress((Bytef*)uncompressed, &uncompressedSize, data.as<Bytef>(), size-8);
}
else if (isCompressed)
{
data2.cur = data2.start = data.as<char>();
data2.end = data2.start+actualSize-8;
}
u8 bits = data2.read<u8>();
u8 numBits = bits>>3;
u32 rectSizeMinusOne = (numBits*4+5)>>3;
data2.move(rectSizeMinusOne);
const u16 frameRate = data2.read<u16>();
const u16 frameCount = data2.read<u16>();
std::set<u32> tagsUsed;
size_t tagCount = 0;
while (data2.cur!=data2.end)
{
u16 tagHeader = data2.read<u16>();
u32 tagLength = tagHeader&0x3F;
u32 tagType = tagHeader>>6;
tagsUsed.insert(tagType);
if (tagLength==0x3F)
tagLength = data2.read<u32>();
data2.move(tagLength);
parseTag(tagType);
++tagCount;
}
if (uncompressed) free(uncompressed);
printf("\nProcessed %d tags\n\n", tagCount);
printf(" Tags used \n");
printf("-------------------------\n");
std::set<u32>::iterator it = tagsUsed.begin(),
end = tagsUsed.end();
for (; it!=end; ++it)
{
parseTag(*it);
}
free(fileData);
}
/* Function: esl_sse_expf()
* Synopsis: <r[z] = exp x[z]>
* Incept: SRE, Fri Dec 14 14:46:27 2007 [Janelia]
*
* Purpose: Given a vector <x> containing four floats, returns a
* vector <r> in which each element <r[z] = expf(x[z])>.
*
* Valid for all IEEE754 floats $x_z$.
*
* Xref: J2/71
* J10/62: bugfix, minlogf/maxlogf range was too wide;
* (k+127) must be >=0 and <=255, so (k+127)<<23
* is a valid IEEE754 float, without touching
* the sign bit. Pommier had this right in the
* first place, and I didn't understand.
*
* Note: Derived from an SSE1 implementation by Julian
* Pommier. Converted to SSE2.
*
* Note on maxlogf/minlogf, which are close to but not
* exactly 127.5/log2 [J10/63]. We need -127<=k<=128, so
* k+127 is 0..255, a valid IEEE754 8-bit exponent
* (0..255), so the bit pattern (k+127)<<23 is IEEE754
* single-precision for 2^k. If k=-127, we get IEEE754 0.
* If k=128, we get IEEE754 +inf. If k<-127, k+127 is
* negative and we get screwed up. If k>128, k+127
* overflows the 8-bit exponent and sets the sign bit. So
* for x' (base 2) < -127.5 we must definitely return e^x ~
* 0; for x' < 126.5 we're going to calculate 0 anyway
* (because k=floor(-126.5-epsilon+0.5) = -127). So any
* minlogf between -126.5 log2 ... -127.5 log2 will suffice
* as the cutoff. Ditto for 126.5 log2 .. 127.5log2.
* That's 87.68312 .. 88.3762655. I think Pommier's
* thinking is, you don't want to get to close to the
* edges, lest fp roundoff error screw you (he may have
* consider 1 ulp carefully, I can't tell), but otherwise
* you may as well put your bounds close to the outer edge;
* so
* maxlogf = 127.5 log(2) - epsilon
* minlogf = -127.5 log(2) + epsilon
* for an epsilon that happen to be ~ 3e-6.
*/
__m128
esl_sse_expf(__m128 x)
{
static float cephes_p[6] = { 1.9875691500E-4f, 1.3981999507E-3f, 8.3334519073E-3f,
4.1665795894E-2f, 1.6666665459E-1f, 5.0000001201E-1f
};
static float cephes_c[2] = { 0.693359375f, -2.12194440e-4f };
static float maxlogf = 88.3762626647949f; /* 127.5 log(2) - epsilon. above this, 0.5+x/log2 gives k>128 and breaks 2^k "float" construction, because (k+127)<<23 must be a valid IEEE754 exponent 0..255 */
static float minlogf = -88.3762626647949f; /*-127.5 log(2) + epsilon. below this, 0.5+x/log2 gives k<-127 and breaks 2^k, see above */
__m128i k;
__m128 mask, tmp, fx, z, y, minmask, maxmask;
/* handle out-of-range and special conditions */
maxmask = _mm_cmpgt_ps(x, _mm_set1_ps(maxlogf));
minmask = _mm_cmple_ps(x, _mm_set1_ps(minlogf));
/* range reduction: exp(x) = 2^k e^f = exp(f + k log 2); k = floorf(0.5 + x / log2): */
fx = _mm_mul_ps(x, _mm_set1_ps(eslCONST_LOG2R));
fx = _mm_add_ps(fx, _mm_set1_ps(0.5f));
/* floorf() with SSE: */
k = _mm_cvttps_epi32(fx); /* cast to int with truncation */
tmp = _mm_cvtepi32_ps(k); /* cast back to float */
mask = _mm_cmpgt_ps(tmp, fx); /* if it increased (i.e. if it was negative...) */
mask = _mm_and_ps(mask, _mm_set1_ps(1.0f)); /* ...without a conditional branch... */
fx = _mm_sub_ps(tmp, mask); /* then subtract one. */
k = _mm_cvttps_epi32(fx); /* k is now ready for the 2^k part. */
/* polynomial approx for e^f for f in range [-0.5, 0.5] */
tmp = _mm_mul_ps(fx, _mm_set1_ps(cephes_c[0]));
z = _mm_mul_ps(fx, _mm_set1_ps(cephes_c[1]));
x = _mm_sub_ps(x, tmp);
x = _mm_sub_ps(x, z);
z = _mm_mul_ps(x, x);
y = _mm_set1_ps(cephes_p[0]);
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[1]));
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[2]));
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[3]));
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[4]));
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(cephes_p[5]));
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, x);
y = _mm_add_ps(y, _mm_set1_ps(1.0f));
/* build 2^k by hand, by creating a IEEE754 float */
k = _mm_add_epi32(k, _mm_set1_epi32(127));
k = _mm_slli_epi32(k, 23);
fx = _mm_castsi128_ps(k);
/* put 2^k e^f together (fx = 2^k, y = e^f) and we're done */
y = _mm_mul_ps(y, fx);
/* special/range cleanup */
y = esl_sse_select_ps(y, _mm_set1_ps(eslINFINITY), maxmask); /* exp(x) = inf for x > log(2^128) */
y = esl_sse_select_ps(y, _mm_set1_ps(0.0f), minmask); /* exp(x) = 0 for x < log(2^-149) */
return y;
}
static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {
__m128i mask, hev, flat, flat2;
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi8(1);
__m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
__m128i abs_p1p0;
const __m128i thresh =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
const __m128i blimit =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
q4p4 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p)));
q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
q3p3 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p)));
q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
q2p2 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p)));
q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
q1p1 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p)));
p1q1 = _mm_shuffle_epi32(q1p1, 78);
q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
q0p0 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p)));
p0q0 = _mm_shuffle_epi32(q0p0, 78);
{
__m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
abs_p1p0 =
_mm_or_si128(_mm_subs_epu8(q1p1, q0p0), _mm_subs_epu8(q0p0, q1p1));
abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
fe = _mm_set1_epi8(0xfe);
ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
abs_p0q0 =
_mm_or_si128(_mm_subs_epu8(q0p0, p0q0), _mm_subs_epu8(p0q0, q0p0));
abs_p1q1 =
_mm_or_si128(_mm_subs_epu8(q1p1, p1q1), _mm_subs_epu8(p1q1, q1p1));
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
// mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
mask = _mm_max_epu8(abs_p1p0, mask);
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q2p2, q1p1), _mm_subs_epu8(q1p1, q2p2)),
_mm_or_si128(_mm_subs_epu8(q3p3, q2p2), _mm_subs_epu8(q2p2, q3p3)));
mask = _mm_max_epu8(work, mask);
mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
}
// lp filter
{
const __m128i t4 = _mm_set1_epi8(4);
const __m128i t3 = _mm_set1_epi8(3);
const __m128i t80 = _mm_set1_epi8(0x80);
const __m128i t1 = _mm_set1_epi16(0x1);
__m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
__m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
__m128i qs0 = _mm_xor_si128(p0q0, t80);
__m128i qs1 = _mm_xor_si128(p1q1, t80);
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
__m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
__m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
work_a = _mm_subs_epi8(qs0, qs0ps0);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
/* (vpx_filter + 3 * (qs0 - ps0)) & mask */
filt = _mm_and_si128(filt, mask);
filter1 = _mm_adds_epi8(filt, t4);
filter2 = _mm_adds_epi8(filt, t3);
filter1 = _mm_unpacklo_epi8(zero, filter1);
filter1 = _mm_srai_epi16(filter1, 0xB);
filter2 = _mm_unpacklo_epi8(zero, filter2);
filter2 = _mm_srai_epi16(filter2, 0xB);
/* Filter1 >> 3 */
filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
/* filt >> 1 */
filt = _mm_adds_epi16(filter1, t1);
filt = _mm_srai_epi16(filt, 1);
filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
filt);
filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
// loopfilter done
{
__m128i work;
flat = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q2p2, q0p0), _mm_subs_epu8(q0p0, q2p2)),
_mm_or_si128(_mm_subs_epu8(q3p3, q0p0), _mm_subs_epu8(q0p0, q3p3)));
flat = _mm_max_epu8(abs_p1p0, flat);
flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
flat = _mm_subs_epu8(flat, one);
flat = _mm_cmpeq_epi8(flat, zero);
flat = _mm_and_si128(flat, mask);
q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
q5p5 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p)));
q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
q6p6 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p)));
flat2 = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q4p4, q0p0), _mm_subs_epu8(q0p0, q4p4)),
_mm_or_si128(_mm_subs_epu8(q5p5, q0p0), _mm_subs_epu8(q0p0, q5p5)));
q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
q7p7 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p)));
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q6p6, q0p0), _mm_subs_epu8(q0p0, q6p6)),
_mm_or_si128(_mm_subs_epu8(q7p7, q0p0), _mm_subs_epu8(q0p0, q7p7)));
flat2 = _mm_max_epu8(work, flat2);
flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
flat2 = _mm_subs_epu8(flat2, one);
flat2 = _mm_cmpeq_epi8(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// flat and wide flat calculations
{
const __m128i eight = _mm_set1_epi16(8);
const __m128i four = _mm_set1_epi16(4);
__m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
__m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
__m128i pixelFilter_p, pixelFilter_q;
__m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
__m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
p7_16 = _mm_unpacklo_epi8(q7p7, zero);
p6_16 = _mm_unpacklo_epi8(q6p6, zero);
p5_16 = _mm_unpacklo_epi8(q5p5, zero);
p4_16 = _mm_unpacklo_epi8(q4p4, zero);
p3_16 = _mm_unpacklo_epi8(q3p3, zero);
p2_16 = _mm_unpacklo_epi8(q2p2, zero);
p1_16 = _mm_unpacklo_epi8(q1p1, zero);
p0_16 = _mm_unpacklo_epi8(q0p0, zero);
q0_16 = _mm_unpackhi_epi8(q0p0, zero);
q1_16 = _mm_unpackhi_epi8(q1p1, zero);
q2_16 = _mm_unpackhi_epi8(q2p2, zero);
q3_16 = _mm_unpackhi_epi8(q3p3, zero);
q4_16 = _mm_unpackhi_epi8(q4p4, zero);
q5_16 = _mm_unpackhi_epi8(q5p5, zero);
q6_16 = _mm_unpackhi_epi8(q6p6, zero);
q7_16 = _mm_unpackhi_epi8(q7p7, zero);
pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
_mm_add_epi16(p4_16, p3_16));
pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
_mm_add_epi16(q4_16, q3_16));
pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
pixelFilter_p =
_mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
pixetFilter_p2p1p0 = _mm_add_epi16(
four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4);
flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3);
flat_q0p0 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(p7_16, p7_16);
sum_q7 = _mm_add_epi16(q7_16, q7_16);
sum_p3 = _mm_add_epi16(p3_16, p3_16);
sum_q3 = _mm_add_epi16(q3_16, q3_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4);
flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
flat_q1p1 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
sum_p3 = _mm_add_epi16(sum_p3, p3_16);
sum_q3 = _mm_add_epi16(sum_q3, q3_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4);
flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
flat_q2p2 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4);
flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
res_p = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4);
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4);
flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
flat = _mm_shuffle_epi32(flat, 68);
flat2 = _mm_shuffle_epi32(flat2, 68);
q2p2 = _mm_andnot_si128(flat, q2p2);
flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
q2p2 = _mm_or_si128(q2p2, flat_q2p2);
qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
q6p6 = _mm_andnot_si128(flat2, q6p6);
flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
_mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
_mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
q5p5 = _mm_andnot_si128(flat2, q5p5);
flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
_mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
_mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
q4p4 = _mm_andnot_si128(flat2, q4p4);
flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
_mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
_mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
q3p3 = _mm_andnot_si128(flat2, q3p3);
flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
_mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
_mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
q2p2 = _mm_andnot_si128(flat2, q2p2);
flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
_mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
_mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
q1p1 = _mm_andnot_si128(flat2, q1p1);
flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
_mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
_mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
q0p0 = _mm_andnot_si128(flat2, q0p0);
flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
_mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
_mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
}
}
INLINE __m128 shade(BilinearSamplePos const& bsp, const SWR_TRIANGLE_DESC & work, WideVector<BilinearSamplePos::NUM_ATTRIBUTES, __m128> const& pAttrs, BYTE* pBuffer, BYTE*, UINT*)
{
TextureView *pTxv = (TextureView*)work.pTextureViews[KNOB_NUMBER_OF_TEXTURE_VIEWS + 0];
Sampler *pSmp = (Sampler*)work.pSamplers[0];
TexCoord tcidx;
tcidx.U = get<4>(pAttrs);
tcidx.V = get<5>(pAttrs);
UINT mips[] = {0,0,0,0};
WideColor color;
SampleSimplePointRGBAF32(*pTxv, *pSmp, tcidx, mips, color);
// modulate
color.R = _mm_mul_ps(color.R, get<0>(pAttrs));
color.G = _mm_mul_ps(color.G, get<1>(pAttrs));
color.B = _mm_mul_ps(color.B, get<2>(pAttrs));
color.A = _mm_mul_ps(color.A, get<3>(pAttrs));
// convert float to unorm
__m128i r = vFloatToUnorm( color.R );
__m128i g = vFloatToUnorm( color.G );
__m128i b = vFloatToUnorm( color.B );
__m128i a = vFloatToUnorm( color.A );
// pack
__m128i vPixel = b;
vPixel = _mm_or_si128(vPixel, _mm_slli_epi32(g, 8));
vPixel = _mm_or_si128(vPixel, _mm_slli_epi32(r, 16));
vPixel = _mm_or_si128(vPixel, _mm_slli_epi32(a, 24));
// blend with GL_ONE and GL_ONE
if (bsp.sFactor == GL_ONE && bsp.dFactor == GL_ONE)
{
__m128i vColorBuffer = _mm_load_si128((const __m128i*)pBuffer);
vPixel = _mm_adds_epu8(vPixel, vColorBuffer);
}
if (bsp.sFactor == GL_SRC_ALPHA && bsp.dFactor == GL_ONE_MINUS_SRC_ALPHA)
{
const __m128i SHUF_ALPHA = _mm_set_epi32(0x8080800f, 0x8080800b, 0x80808007, 0x80808003);
const __m128i SHUF_RED = _mm_set_epi32(0x8080800e, 0x8080800a, 0x80808006, 0x80808002);
const __m128i SHUF_GREEN = _mm_set_epi32(0x8080800d, 0x80808009, 0x80808005, 0x80808001);
const __m128i SHUF_BLUE = _mm_set_epi32(0x8080800c, 0x80808008, 0x80808004, 0x80808000);
// mul by src_alpha
__m128 vSrcRedF = _mm_mul_ps(color.R, color.A);
__m128 vSrcGreenF = _mm_mul_ps(color.G, color.A);
__m128 vSrcBlueF = _mm_mul_ps(color.B, color.A);
// convert to int
__m128i vSrcRed = vFloatToUnorm(vSrcRedF);
__m128i vSrcGreen = vFloatToUnorm(vSrcGreenF);
__m128i vSrcBlue = vFloatToUnorm(vSrcBlueF);
__m128i vSrcAlpha = vFloatToUnorm(color.A);
// pack
__m128i vSrcPixel = vSrcBlue;
vSrcPixel = _mm_or_si128(vSrcPixel, _mm_slli_epi32(vSrcGreen, 8));
vSrcPixel = _mm_or_si128(vSrcPixel, _mm_slli_epi32(vSrcRed, 16));
vSrcPixel = _mm_or_si128(vSrcPixel, _mm_slli_epi32(vSrcAlpha, 24));
// shuffle dst R,G,B,A
__m128i vColorBuffer = _mm_load_si128((const __m128i*)pBuffer);
__m128i vDstAlpha = _mm_shuffle_epi8(vColorBuffer, SHUF_ALPHA);
__m128i vDstRed = _mm_shuffle_epi8(vColorBuffer, SHUF_RED);
__m128i vDstGreen = _mm_shuffle_epi8(vColorBuffer, SHUF_GREEN);
__m128i vDstBlue = _mm_shuffle_epi8(vColorBuffer, SHUF_BLUE);
// convert to float
__m128 vDstAlphaF = _mm_cvtepi32_ps(vDstAlpha);
__m128 vDstRedF = _mm_cvtepi32_ps(vDstRed);
__m128 vDstGreenF = _mm_cvtepi32_ps(vDstGreen);
__m128 vDstBlueF = _mm_cvtepi32_ps(vDstBlue);
// mul by 1-src_alpha
__m128 vOneMinusSrcAlphaF = _mm_sub_ps(_mm_set1_ps(1.0f), color.A);
vDstAlphaF = _mm_mul_ps(vDstAlphaF, vOneMinusSrcAlphaF);
vDstRedF = _mm_mul_ps(vDstRedF, vOneMinusSrcAlphaF);
vDstGreenF = _mm_mul_ps(vDstGreenF, vOneMinusSrcAlphaF);
vDstBlueF = _mm_mul_ps(vDstBlueF, vOneMinusSrcAlphaF);
// convert to int
vDstAlpha = _mm_cvtps_epi32(vDstAlphaF);
vDstRed = _mm_cvtps_epi32(vDstRedF);
vDstGreen = _mm_cvtps_epi32(vDstGreenF);
vDstBlue = _mm_cvtps_epi32(vDstBlueF);
// pack
__m128i vDstPixel = vDstBlue;
vDstPixel = _mm_or_si128(vDstPixel, _mm_slli_epi32(vDstGreen, 8));
vDstPixel = _mm_or_si128(vDstPixel, _mm_slli_epi32(vDstRed, 16));
vDstPixel = _mm_or_si128(vDstPixel, _mm_slli_epi32(vDstAlpha, 24));
// final rgba = min(src + dst,255)
vPixel = _mm_adds_epu8(vSrcPixel, vDstPixel);
}
return _mm_castsi128_ps(vPixel);
}
/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos_ps(__m128 x, __m128* s, __m128* c) {
typedef __m128 v4sf;
typedef __m128i v4si;
v4sf xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
v4si emm0, emm2, emm4;
sign_bit_sin = x;
/* take the absolute value */
x = _mm_and_ps(x, constants::inv_sign_mask.ps);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm_and_ps(sign_bit_sin, constants::sign_mask.ps);
/* scale by 4/Pi */
y = _mm_mul_ps(x, constants::cephes_FOPI.ps);
/* store the integer part of y in emm2 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, constants::pi32_1.pi);
emm2 = _mm_and_si128(emm2, constants::pi32_inv1.pi);
y = _mm_cvtepi32_ps(emm2);
emm4 = emm2;
/* get the swap sign flag for the sine */
emm0 = _mm_and_si128(emm2, constants::pi32_4.pi);
emm0 = _mm_slli_epi32(emm0, 29);
v4sf swap_sign_bit_sin = _mm_castsi128_ps(emm0);
/* get the polynom selection mask for the sine*/
emm2 = _mm_and_si128(emm2, constants::pi32_2.pi);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
v4sf poly_mask = _mm_castsi128_ps(emm2);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = constants::minus_cephes_DP1.ps;
xmm2 = constants::minus_cephes_DP2.ps;
xmm3 = constants::minus_cephes_DP3.ps;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
emm4 = _mm_sub_epi32(emm4, constants::pi32_2.pi);
emm4 = _mm_andnot_si128(emm4, constants::pi32_4.pi);
emm4 = _mm_slli_epi32(emm4, 29);
v4sf sign_bit_cos = _mm_castsi128_ps(emm4);
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
v4sf z = _mm_mul_ps(x,x);
y = constants::coscof_p0.ps;
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, constants::coscof_p1.ps);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, constants::coscof_p2.ps);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
v4sf tmp = _mm_mul_ps(z, constants::ps_0p5.ps);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, constants::ps_1.ps);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v4sf y2 = constants::sincof_p0.ps;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, constants::sincof_p1.ps);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, constants::sincof_p2.ps);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
v4sf ysin2 = _mm_and_ps(xmm3, y2);
v4sf ysin1 = _mm_andnot_ps(xmm3, y);
y2 = _mm_sub_ps(y2,ysin2);
y = _mm_sub_ps(y, ysin1);
xmm1 = _mm_add_ps(ysin1,ysin2);
xmm2 = _mm_add_ps(y,y2);
/* update the sign */
*s = _mm_xor_ps(xmm1, sign_bit_sin);
*c = _mm_xor_ps(xmm2, sign_bit_cos);
}
void sincos_ps(__m128 x, __m128 *s, __m128 *c) {
__m128 xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
__m128i emm0, emm2, emm4;
sign_bit_sin = x;
x = _mm_and_ps(x, *reinterpret_cast<const __m128*>(_pi_inv_sign_mask));
sign_bit_sin = _mm_and_ps(sign_bit_sin,
*reinterpret_cast<const __m128*>(_pi_sign_mask));
y = _mm_mul_ps(x, *_ps_cephes_FOPI);
emm2 = _mm_cvttps_epi32(y);
emm2 = _mm_add_epi32(emm2, *_pi_1);
emm2 = _mm_and_si128(emm2, *_pi_inv1);
y = _mm_cvtepi32_ps(emm2);
emm4 = emm2;
emm0 = _mm_and_si128(emm2, *_pi_4);
emm0 = _mm_slli_epi32(emm0, 29);
__m128 swap_sign_bit_sin = _mm_castsi128_ps(emm0);
emm2 = _mm_and_si128(emm2, *_pi_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 poly_mask = _mm_castsi128_ps(emm2);
xmm1 = *_ps_minus_cephes_DP1;
xmm2 = *_ps_minus_cephes_DP2;
xmm3 = *_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
emm4 = _mm_sub_epi32(emm4, *_pi_2);
emm4 = _mm_andnot_si128(emm4, *_pi_4);
emm4 = _mm_slli_epi32(emm4, 29);
__m128 sign_bit_cos = _mm_castsi128_ps(emm4);
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
__m128 z = _mm_mul_ps(x, x);
y = *_ps_coscof_p0;
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *_ps_1);
__m128 y2 = *_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
xmm3 = poly_mask;
__m128 ysin2 = _mm_and_ps(xmm3, y2);
__m128 ysin1 = _mm_andnot_ps(xmm3, y);
y2 = _mm_sub_ps(y2, ysin2);
y = _mm_sub_ps(y, ysin1);
xmm1 = _mm_add_ps(ysin1, ysin2);
xmm2 = _mm_add_ps(y, y2);
*s = _mm_xor_ps(xmm1, sign_bit_sin);
*c = _mm_xor_ps(xmm2, sign_bit_cos);
}
__m128 exp_ps(v4sfu *xPtr) {
__m128 x=*((__m128 *)xPtr);
__m128 tmp = _mm_setzero_ps(), fx;
#ifdef USE_SSE2
__m128i emm0;
#else
__m64 mm0, mm1;
#endif
__m128 one = *(__m128*)_ps_1;
x = _mm_min_ps(x, *(__m128*)_ps_exp_hi);
x = _mm_max_ps(x, *(__m128*)_ps_exp_lo);
/* express exp(x) as exp(g + n*log(2)) */
fx = _mm_mul_ps(x, *(__m128*)_ps_cephes_LOG2EF);
fx = _mm_add_ps(fx, *(__m128*)_ps_0p5);
/* how to perform a floorf with SSE: just below */
#ifndef USE_SSE2
/* step 1 : cast to int */
tmp = _mm_movehl_ps(tmp, fx);
mm0 = _mm_cvttps_pi32(fx);
mm1 = _mm_cvttps_pi32(tmp);
/* step 2 : cast back to float */
tmp = _mm_cvtpi32x2_ps(mm0, mm1);
#else
emm0 = _mm_cvttps_epi32(fx);
tmp = _mm_cvtepi32_ps(emm0);
#endif
/* if greater, substract 1 */
__m128 mask = _mm_cmpgt_ps(tmp, fx);
mask = _mm_and_ps(mask, one);
fx = _mm_sub_ps(tmp, mask);
tmp = _mm_mul_ps(fx, *(__m128*)_ps_cephes_exp_C1);
__m128 z = _mm_mul_ps(fx, *(__m128*)_ps_cephes_exp_C2);
x = _mm_sub_ps(x, tmp);
x = _mm_sub_ps(x, z);
z = _mm_mul_ps(x,x);
__m128 y = *(__m128*)_ps_cephes_exp_p0;
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, *(__m128*)_ps_cephes_exp_p1);
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, *(__m128*)_ps_cephes_exp_p2);
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, *(__m128*)_ps_cephes_exp_p3);
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, *(__m128*)_ps_cephes_exp_p4);
y = _mm_mul_ps(y, x);
y = _mm_add_ps(y, *(__m128*)_ps_cephes_exp_p5);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, x);
y = _mm_add_ps(y, one);
/* build 2^n */
#ifndef USE_SSE2
z = _mm_movehl_ps(z, fx);
mm0 = _mm_cvttps_pi32(fx);
mm1 = _mm_cvttps_pi32(z);
mm0 = _mm_add_pi32(mm0, *(__m64*)_pi32_0x7f);
mm1 = _mm_add_pi32(mm1, *(__m64*)_pi32_0x7f);
mm0 = _mm_slli_pi32(mm0, 23);
mm1 = _mm_slli_pi32(mm1, 23);
__m128 pow2n;
COPY_MM_TO_XMM(mm0, mm1, pow2n);
_mm_empty();
#else
emm0 = _mm_cvttps_epi32(fx);
emm0 = _mm_add_epi32(emm0, *(__m128i*)_pi32_0x7f);
emm0 = _mm_slli_epi32(emm0, 23);
__m128 pow2n = _mm_castsi128_ps(emm0);
#endif
y = _mm_mul_ps(y, pow2n);
return y;
}
void spu_rotmahi( SPU_t* SPU, SPU_INSTRUCTION op )
{
const int s = 0x1f&( 0 - SignExtend(op.RI7.I7, 7) );
SPU->GPR[op.RI7.RT] = _mm_castsi128_ps( _mm_srai_epi16( _mm_castps_si128( SPU->GPR[op.RI7.RA] ), s ) );
}
/* almost the same as sin_ps */
__m128 cos_ps(v4sfu *xPtr) { // any x
__m128 x=*((__m128 *)xPtr);
__m128 xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
#ifdef USE_SSE2
__m128i emm0, emm2;
#else
__m64 mm0, mm1, mm2, mm3;
#endif
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in mm0 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm2 = _mm_sub_epi32(emm2, *(__m128i*)_pi32_2);
/* get the swap sign flag */
emm0 = _mm_andnot_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
/* get the polynom selection mask */
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 sign_bit = _mm_castsi128_ps(emm0);
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm0:mm1 */
xmm2 = _mm_movehl_ps(xmm2, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm2);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm2 = _mm_sub_pi32(mm2, *(__m64*)_pi32_2);
mm3 = _mm_sub_pi32(mm3, *(__m64*)_pi32_2);
/* get the swap sign flag in mm0:mm1 and the
polynom selection mask in mm2:mm3 */
mm0 = _mm_andnot_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_andnot_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 sign_bit, poly_mask;
COPY_MM_TO_XMM(mm0, mm1, sign_bit);
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
_mm_empty(); /* good-bye mmx */
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(__m128*)_ps_coscof_p0;
__m128 z = _mm_mul_ps(x,x);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
y = _mm_andnot_ps(xmm3, y);
y = _mm_add_ps(y,y2);
/* update the sign */
y = _mm_xor_ps(y, sign_bit);
return y;
}
static void cftmdl_128_SSE2(float* a) {
const int l = 8;
const __m128 mm_swap_sign = _mm_load_ps(k_swap_sign);
int j0;
__m128 wk1rv = _mm_load_ps(cftmdl_wk1r);
for (j0 = 0; j0 < l; j0 += 2) {
const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]);
const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]);
const __m128i a_32 = _mm_loadl_epi64((__m128i*)&a[j0 + 32]);
const __m128i a_40 = _mm_loadl_epi64((__m128i*)&a[j0 + 40]);
const __m128 a_00_32 = _mm_shuffle_ps(_mm_castsi128_ps(a_00),
_mm_castsi128_ps(a_32),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_08_40 = _mm_shuffle_ps(_mm_castsi128_ps(a_08),
_mm_castsi128_ps(a_40),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 x0r0_0i0_0r1_x0i1 = _mm_add_ps(a_00_32, a_08_40);
const __m128 x1r0_1i0_1r1_x1i1 = _mm_sub_ps(a_00_32, a_08_40);
const __m128i a_16 = _mm_loadl_epi64((__m128i*)&a[j0 + 16]);
const __m128i a_24 = _mm_loadl_epi64((__m128i*)&a[j0 + 24]);
const __m128i a_48 = _mm_loadl_epi64((__m128i*)&a[j0 + 48]);
const __m128i a_56 = _mm_loadl_epi64((__m128i*)&a[j0 + 56]);
const __m128 a_16_48 = _mm_shuffle_ps(_mm_castsi128_ps(a_16),
_mm_castsi128_ps(a_48),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_24_56 = _mm_shuffle_ps(_mm_castsi128_ps(a_24),
_mm_castsi128_ps(a_56),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 x2r0_2i0_2r1_x2i1 = _mm_add_ps(a_16_48, a_24_56);
const __m128 x3r0_3i0_3r1_x3i1 = _mm_sub_ps(a_16_48, a_24_56);
const __m128 xx0 = _mm_add_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx1 = _mm_sub_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 x3i0_3r0_3i1_x3r1 = _mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(x3r0_3i0_3r1_x3i1), _MM_SHUFFLE(2, 3, 0, 1)));
const __m128 x3_swapped = _mm_mul_ps(mm_swap_sign, x3i0_3r0_3i1_x3r1);
const __m128 x1_x3_add = _mm_add_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 x1_x3_sub = _mm_sub_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 yy0 =
_mm_shuffle_ps(x1_x3_add, x1_x3_sub, _MM_SHUFFLE(2, 2, 2, 2));
const __m128 yy1 =
_mm_shuffle_ps(x1_x3_add, x1_x3_sub, _MM_SHUFFLE(3, 3, 3, 3));
const __m128 yy2 = _mm_mul_ps(mm_swap_sign, yy1);
const __m128 yy3 = _mm_add_ps(yy0, yy2);
const __m128 yy4 = _mm_mul_ps(wk1rv, yy3);
_mm_storel_epi64((__m128i*)&a[j0 + 0], _mm_castps_si128(xx0));
_mm_storel_epi64(
(__m128i*)&a[j0 + 32],
_mm_shuffle_epi32(_mm_castps_si128(xx0), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 16], _mm_castps_si128(xx1));
_mm_storel_epi64(
(__m128i*)&a[j0 + 48],
_mm_shuffle_epi32(_mm_castps_si128(xx1), _MM_SHUFFLE(2, 3, 2, 3)));
a[j0 + 48] = -a[j0 + 48];
_mm_storel_epi64((__m128i*)&a[j0 + 8], _mm_castps_si128(x1_x3_add));
_mm_storel_epi64((__m128i*)&a[j0 + 24], _mm_castps_si128(x1_x3_sub));
_mm_storel_epi64((__m128i*)&a[j0 + 40], _mm_castps_si128(yy4));
_mm_storel_epi64(
(__m128i*)&a[j0 + 56],
_mm_shuffle_epi32(_mm_castps_si128(yy4), _MM_SHUFFLE(2, 3, 2, 3)));
}
{
int k = 64;
int k1 = 2;
int k2 = 2 * k1;
const __m128 wk2rv = _mm_load_ps(&rdft_wk2r[k2 + 0]);
const __m128 wk2iv = _mm_load_ps(&rdft_wk2i[k2 + 0]);
const __m128 wk1iv = _mm_load_ps(&rdft_wk1i[k2 + 0]);
const __m128 wk3rv = _mm_load_ps(&rdft_wk3r[k2 + 0]);
const __m128 wk3iv = _mm_load_ps(&rdft_wk3i[k2 + 0]);
wk1rv = _mm_load_ps(&rdft_wk1r[k2 + 0]);
for (j0 = k; j0 < l + k; j0 += 2) {
const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]);
const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]);
const __m128i a_32 = _mm_loadl_epi64((__m128i*)&a[j0 + 32]);
const __m128i a_40 = _mm_loadl_epi64((__m128i*)&a[j0 + 40]);
const __m128 a_00_32 = _mm_shuffle_ps(_mm_castsi128_ps(a_00),
_mm_castsi128_ps(a_32),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_08_40 = _mm_shuffle_ps(_mm_castsi128_ps(a_08),
_mm_castsi128_ps(a_40),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 x0r0_0i0_0r1_x0i1 = _mm_add_ps(a_00_32, a_08_40);
const __m128 x1r0_1i0_1r1_x1i1 = _mm_sub_ps(a_00_32, a_08_40);
const __m128i a_16 = _mm_loadl_epi64((__m128i*)&a[j0 + 16]);
const __m128i a_24 = _mm_loadl_epi64((__m128i*)&a[j0 + 24]);
const __m128i a_48 = _mm_loadl_epi64((__m128i*)&a[j0 + 48]);
const __m128i a_56 = _mm_loadl_epi64((__m128i*)&a[j0 + 56]);
const __m128 a_16_48 = _mm_shuffle_ps(_mm_castsi128_ps(a_16),
_mm_castsi128_ps(a_48),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_24_56 = _mm_shuffle_ps(_mm_castsi128_ps(a_24),
_mm_castsi128_ps(a_56),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 x2r0_2i0_2r1_x2i1 = _mm_add_ps(a_16_48, a_24_56);
const __m128 x3r0_3i0_3r1_x3i1 = _mm_sub_ps(a_16_48, a_24_56);
const __m128 xx = _mm_add_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx1 = _mm_sub_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx2 = _mm_mul_ps(xx1, wk2rv);
const __m128 xx3 =
_mm_mul_ps(wk2iv,
_mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(xx1), _MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx4 = _mm_add_ps(xx2, xx3);
const __m128 x3i0_3r0_3i1_x3r1 = _mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(x3r0_3i0_3r1_x3i1), _MM_SHUFFLE(2, 3, 0, 1)));
const __m128 x3_swapped = _mm_mul_ps(mm_swap_sign, x3i0_3r0_3i1_x3r1);
const __m128 x1_x3_add = _mm_add_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 x1_x3_sub = _mm_sub_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 xx10 = _mm_mul_ps(x1_x3_add, wk1rv);
const __m128 xx11 = _mm_mul_ps(
wk1iv,
_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(x1_x3_add),
_MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx12 = _mm_add_ps(xx10, xx11);
const __m128 xx20 = _mm_mul_ps(x1_x3_sub, wk3rv);
const __m128 xx21 = _mm_mul_ps(
wk3iv,
_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(x1_x3_sub),
_MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx22 = _mm_add_ps(xx20, xx21);
_mm_storel_epi64((__m128i*)&a[j0 + 0], _mm_castps_si128(xx));
_mm_storel_epi64(
(__m128i*)&a[j0 + 32],
_mm_shuffle_epi32(_mm_castps_si128(xx), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 16], _mm_castps_si128(xx4));
_mm_storel_epi64(
(__m128i*)&a[j0 + 48],
_mm_shuffle_epi32(_mm_castps_si128(xx4), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 8], _mm_castps_si128(xx12));
_mm_storel_epi64(
(__m128i*)&a[j0 + 40],
_mm_shuffle_epi32(_mm_castps_si128(xx12), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 24], _mm_castps_si128(xx22));
_mm_storel_epi64(
(__m128i*)&a[j0 + 56],
_mm_shuffle_epi32(_mm_castps_si128(xx22), _MM_SHUFFLE(3, 2, 3, 2)));
}
}
}
/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos_ps(v4sfu *xptr, v4sfu *sptr, v4sfu *cptr) {
__m128 x=*((__m128 *)xptr), *s=(__m128 *)sptr, *c=(__m128 *)cptr, xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
#ifdef USE_SSE2
__m128i emm0, emm2, emm4;
#else
__m64 mm0, mm1, mm2, mm3, mm4, mm5;
#endif
sign_bit_sin = x;
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm_and_ps(sign_bit_sin, *(__m128*)_ps_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in emm2 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm4 = emm2;
/* get the swap sign flag for the sine */
emm0 = _mm_and_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
__m128 swap_sign_bit_sin = _mm_castsi128_ps(emm0);
/* get the polynom selection mask for the sine*/
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm2:mm3 */
xmm3 = _mm_movehl_ps(xmm3, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm3);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm4 = mm2;
mm5 = mm3;
/* get the swap sign flag for the sine */
mm0 = _mm_and_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_and_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
__m128 swap_sign_bit_sin;
COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit_sin);
/* get the polynom selection mask for the sine */
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 poly_mask;
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
#ifdef USE_SSE2
emm4 = _mm_sub_epi32(emm4, *(__m128i*)_pi32_2);
emm4 = _mm_andnot_si128(emm4, *(__m128i*)_pi32_4);
emm4 = _mm_slli_epi32(emm4, 29);
__m128 sign_bit_cos = _mm_castsi128_ps(emm4);
#else
/* get the sign flag for the cosine */
mm4 = _mm_sub_pi32(mm4, *(__m64*)_pi32_2);
mm5 = _mm_sub_pi32(mm5, *(__m64*)_pi32_2);
mm4 = _mm_andnot_si64(mm4, *(__m64*)_pi32_4);
mm5 = _mm_andnot_si64(mm5, *(__m64*)_pi32_4);
mm4 = _mm_slli_pi32(mm4, 29);
mm5 = _mm_slli_pi32(mm5, 29);
__m128 sign_bit_cos;
COPY_MM_TO_XMM(mm4, mm5, sign_bit_cos);
_mm_empty(); /* good-bye mmx */
#endif
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
__m128 z = _mm_mul_ps(x,x);
y = *(__m128*)_ps_coscof_p0;
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
__m128 ysin2 = _mm_and_ps(xmm3, y2);
__m128 ysin1 = _mm_andnot_ps(xmm3, y);
y2 = _mm_sub_ps(y2,ysin2);
y = _mm_sub_ps(y, ysin1);
xmm1 = _mm_add_ps(ysin1,ysin2);
xmm2 = _mm_add_ps(y,y2);
/* update the sign */
*s = _mm_xor_ps(xmm1, sign_bit_sin);
*c = _mm_xor_ps(xmm2, sign_bit_cos);
}
void SWModelRenderer::RenderInner(spades::draw::SWModel *model,
const client::ModelRenderParam ¶m) {
auto& mat = param.matrix;
auto origin = mat.GetOrigin();
auto axis1 = mat.GetAxis(0);
auto axis2 = mat.GetAxis(1);
auto axis3 = mat.GetAxis(2);
auto *rawModel = model->GetRawModel();
auto rawModelOrigin = rawModel->GetOrigin();
rawModelOrigin += 0.1f;
origin += axis1 * rawModelOrigin.x;
origin += axis2 * rawModelOrigin.y;
origin += axis3 * rawModelOrigin.z;
int w = rawModel->GetWidth();
int h = rawModel->GetHeight();
//int d = rawModel->GetDepth();
// evaluate brightness for each normals
uint8_t brights[3*3*3];
{
auto lightVec = MakeVector3(0.f, -0.707f, -0.707f);
float dot1 = Vector3::Dot(axis1, lightVec) * fastRSqrt(axis1.GetPoweredLength());
float dot2 = Vector3::Dot(axis2, lightVec) * fastRSqrt(axis2.GetPoweredLength());
float dot3 = Vector3::Dot(axis3, lightVec) * fastRSqrt(axis3.GetPoweredLength());
for(int x = 0; x < 3; x++){
float d;
int cnt;
switch(x){
case 0: d = -dot1; cnt = 1; break;
case 1: d = 0.f; cnt = 0; break;
case 2: d = dot1; cnt = 1; break;
}
for(int y = 0; y < 3; y++){
auto d2 = d;
auto cnt2 = cnt;
switch(y){
case 0: d2 -= dot2; cnt2++; break;
case 1: break;
case 2: d2 += dot2; cnt2++; break;
}
for(int z = 0; z < 3; z++) {
auto d3 = d;
auto cnt3 = cnt2;
switch(y){
case 0: d3 -= dot3; cnt3++; break;
case 1: break;
case 2: d3 += dot3; cnt3++; break;
}
switch(cnt3){
case 2:
d3 *= 0.707f;
break;
case 3:
d3 *= 0.57735f;
break;
}
d3 = 192.f + d3 * 62.f;
brights[x + y * 3 + z * 9]
= static_cast<uint8_t>(d3);
}
}
}
}
// compute center coord. for culling
{
auto center = origin;
auto localCenter = model->GetCenter();
center += axis1 * localCenter.x;
center += axis2 * localCenter.y;
center += axis3 * localCenter.z;
float largestAxis = axis1.GetPoweredLength();
largestAxis = std::max(largestAxis, axis2.GetPoweredLength());
largestAxis = std::max(largestAxis, axis3.GetPoweredLength());
if(!r->SphereFrustrumCull(center, model->GetRadius() * sqrtf(largestAxis)))
return;
}
Bitmap *fbmp = r->fb;
auto *fb = fbmp->GetPixels();
int fw = fbmp->GetWidth();
int fh = fbmp->GetHeight();
auto *db = r->depthBuffer.data();
Matrix4 viewproj = r->GetProjectionViewMatrix();
Vector4 ndc2scrscale = {fw * 0.5f, -fh * 0.5f, 1.f, 1.f};
//Vector4 ndc2scroff = {fw * 0.5f, fh * 0.5f, 0.f, 0.f};
int ndc2scroffX = fw >> 1;
int ndc2scroffY = fh >> 1;
// render each points
auto tOrigin = viewproj * MakeVector4(origin.x, origin.y, origin.z, 1.f);
auto tAxis1 = viewproj * MakeVector4(axis1.x, axis1.y, axis1.z, 0.f);
auto tAxis2 = viewproj * MakeVector4(axis2.x, axis2.y, axis2.z, 0.f);
auto tAxis3 = viewproj * MakeVector4(axis3.x, axis3.y, axis3.z, 0.f);
tOrigin *= ndc2scrscale;
tAxis1 *= ndc2scrscale;
tAxis2 *= ndc2scrscale;
tAxis3 *= ndc2scrscale;
float pointDiameter;// = largestAxis * 0.55f * fh * 0.5f;
{
float largestAxis = tAxis1.GetPoweredLength();
largestAxis = std::max(largestAxis, tAxis2.GetPoweredLength());
largestAxis = std::max(largestAxis, tAxis3.GetPoweredLength());
pointDiameter = sqrtf(largestAxis);
}
uint32_t customColor;
customColor =
ToFixed8(param.customColor.z) |
(ToFixed8(param.customColor.y) << 8) |
(ToFixed8(param.customColor.x) << 16);
auto v1 = tOrigin;
float zNear = r->sceneDef.zNear;
for(int x = 0; x < w; x++) {
auto v2 = v1;
for(int y = 0; y < h; y++) {
auto *mp = &model->renderData
[model->renderDataAddr[x + y * w]];
while(*mp != -1) {
uint32_t data = *(mp++);
uint32_t normal = *(mp++);
int z = static_cast<int>(data >> 24);
//SPAssert(z < d);
SPAssert(z >= 0);
auto vv = v2 + tAxis3 * zvals[z];
if(vv.z < zNear) continue;
// save Z value (don't divide this by W!)
float zval = vv.z;
// use vv.z for point radius to be divided by W
vv.z = pointDiameter;
// perspective division
float scl = fastRcp(vv.w);
vv *= scl;
int ix = static_cast<int>(vv.x) + ndc2scroffX;
int iy = static_cast<int>(vv.y) + ndc2scroffY;
int idm = static_cast<int>(vv.z + .99f);
idm = std::max(1, idm);
int minX = ix - (idm >> 1);
int minY = iy - (idm >> 1);
if(minX >= fw || minY >= fh) continue;
int maxX = ix + idm;
int maxY = iy + idm;
if(maxX <= 0 || maxY <= 0) continue;
minX = std::max(minX, 0);
minY = std::max(minY, 0);
maxX = std::min(maxX, fw);
maxY = std::min(maxY, fh);
auto *fb2 = fb + (minX + minY * fw);
auto *db2 = db + (minX + minY * fw);
int w = maxX - minX;
uint32_t color = data & 0xffffff;
if(color == 0)
color = customColor;
SPAssert(normal < 27);
int bright = brights[normal];
#if ENABLE_SSE2
if(lvl == SWFeatureLevel::SSE2) {
auto m = _mm_setr_epi32(color, 0, 0, 0);
auto f = _mm_set1_epi16(bright << 8);
m = _mm_unpacklo_epi8(m, _mm_setzero_si128());
m = _mm_mulhi_epu16(m, f);
m = _mm_packus_epi16(m, m);
_mm_store_ss(reinterpret_cast<float*>(&color),
_mm_castsi128_ps(m));
}else
#endif
{
uint32_t c1 = color & 0xff00;
uint32_t c2 = color & 0xff00ff;
c1 *= bright;
c2 *= bright;
color = ((c1&0xff0000) | (c2&0xff00ff00)) >> 8;
}
for(int yy = minY; yy < maxY; yy++){
auto *fb3 = fb2;
auto *db3 = db2;
for(int xx = w; xx > 0; xx--) {
if(zval < *db3) {
*db3 = zval;
*fb3 = color;
}
fb3++; db3++;
}
fb2 += fw;
db2 += fw;
}
}
v2 += tAxis2;
}
v1 += tAxis1;
}
}
void spu_interpreter::FRSQEST(SPUThread& CPU, spu_opcode_t op)
{
const auto mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff));
CPU.GPR[op.rt].vf = _mm_rsqrt_ps(_mm_and_ps(CPU.GPR[op.ra].vf, mask));
}
static __m128 mm_pow_ps(__m128 a, __m128 b) {
// a^b = exp2(b * log2(a))
// exp2(x) and log2(x) are calculated using polynomial approximations.
__m128 log2_a, b_log2_a, a_exp_b;
// Calculate log2(x), x = a.
{
// To calculate log2(x), we decompose x like this:
// x = y * 2^n
// n is an integer
// y is in the [1.0, 2.0) range
//
// log2(x) = log2(y) + n
// n can be evaluated by playing with float representation.
// log2(y) in a small range can be approximated, this code uses an order
// five polynomial approximation. The coefficients have been
// estimated with the Remez algorithm and the resulting
// polynomial has a maximum relative error of 0.00086%.
// Compute n.
// This is done by masking the exponent, shifting it into the top bit of
// the mantissa, putting eight into the biased exponent (to shift/
// compensate the fact that the exponent has been shifted in the top/
// fractional part and finally getting rid of the implicit leading one
// from the mantissa by substracting it out.
static const ALIGN16_BEG int float_exponent_mask[4] ALIGN16_END = {
0x7F800000, 0x7F800000, 0x7F800000, 0x7F800000};
static const ALIGN16_BEG int eight_biased_exponent[4] ALIGN16_END = {
0x43800000, 0x43800000, 0x43800000, 0x43800000};
static const ALIGN16_BEG int implicit_leading_one[4] ALIGN16_END = {
0x43BF8000, 0x43BF8000, 0x43BF8000, 0x43BF8000};
static const int shift_exponent_into_top_mantissa = 8;
const __m128 two_n = _mm_and_ps(a, *((__m128*)float_exponent_mask));
const __m128 n_1 = _mm_castsi128_ps(_mm_srli_epi32(
_mm_castps_si128(two_n), shift_exponent_into_top_mantissa));
const __m128 n_0 = _mm_or_ps(n_1, *((__m128*)eight_biased_exponent));
const __m128 n = _mm_sub_ps(n_0, *((__m128*)implicit_leading_one));
// Compute y.
static const ALIGN16_BEG int mantissa_mask[4] ALIGN16_END = {
0x007FFFFF, 0x007FFFFF, 0x007FFFFF, 0x007FFFFF};
static const ALIGN16_BEG int zero_biased_exponent_is_one[4] ALIGN16_END = {
0x3F800000, 0x3F800000, 0x3F800000, 0x3F800000};
const __m128 mantissa = _mm_and_ps(a, *((__m128*)mantissa_mask));
const __m128 y =
_mm_or_ps(mantissa, *((__m128*)zero_biased_exponent_is_one));
// Approximate log2(y) ~= (y - 1) * pol5(y).
// pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
static const ALIGN16_BEG float ALIGN16_END C5[4] = {
-3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f};
static const ALIGN16_BEG float ALIGN16_END
C4[4] = {3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f};
static const ALIGN16_BEG float ALIGN16_END
C3[4] = {-1.2315303f, -1.2315303f, -1.2315303f, -1.2315303f};
static const ALIGN16_BEG float ALIGN16_END
C2[4] = {2.5988452f, 2.5988452f, 2.5988452f, 2.5988452f};
static const ALIGN16_BEG float ALIGN16_END
C1[4] = {-3.3241990f, -3.3241990f, -3.3241990f, -3.3241990f};
static const ALIGN16_BEG float ALIGN16_END
C0[4] = {3.1157899f, 3.1157899f, 3.1157899f, 3.1157899f};
const __m128 pol5_y_0 = _mm_mul_ps(y, *((__m128*)C5));
const __m128 pol5_y_1 = _mm_add_ps(pol5_y_0, *((__m128*)C4));
const __m128 pol5_y_2 = _mm_mul_ps(pol5_y_1, y);
const __m128 pol5_y_3 = _mm_add_ps(pol5_y_2, *((__m128*)C3));
const __m128 pol5_y_4 = _mm_mul_ps(pol5_y_3, y);
const __m128 pol5_y_5 = _mm_add_ps(pol5_y_4, *((__m128*)C2));
const __m128 pol5_y_6 = _mm_mul_ps(pol5_y_5, y);
const __m128 pol5_y_7 = _mm_add_ps(pol5_y_6, *((__m128*)C1));
const __m128 pol5_y_8 = _mm_mul_ps(pol5_y_7, y);
const __m128 pol5_y = _mm_add_ps(pol5_y_8, *((__m128*)C0));
const __m128 y_minus_one =
_mm_sub_ps(y, *((__m128*)zero_biased_exponent_is_one));
const __m128 log2_y = _mm_mul_ps(y_minus_one, pol5_y);
// Combine parts.
log2_a = _mm_add_ps(n, log2_y);
}
// b * log2(a)
b_log2_a = _mm_mul_ps(b, log2_a);
// Calculate exp2(x), x = b * log2(a).
{
// To calculate 2^x, we decompose x like this:
// x = n + y
// n is an integer, the value of x - 0.5 rounded down, therefore
// y is in the [0.5, 1.5) range
//
// 2^x = 2^n * 2^y
// 2^n can be evaluated by playing with float representation.
// 2^y in a small range can be approximated, this code uses an order two
// polynomial approximation. The coefficients have been estimated
// with the Remez algorithm and the resulting polynomial has a
// maximum relative error of 0.17%.
// To avoid over/underflow, we reduce the range of input to ]-127, 129].
static const ALIGN16_BEG float max_input[4] ALIGN16_END = {129.f, 129.f,
129.f, 129.f};
static const ALIGN16_BEG float min_input[4] ALIGN16_END = {
-126.99999f, -126.99999f, -126.99999f, -126.99999f};
const __m128 x_min = _mm_min_ps(b_log2_a, *((__m128*)max_input));
const __m128 x_max = _mm_max_ps(x_min, *((__m128*)min_input));
// Compute n.
static const ALIGN16_BEG float half[4] ALIGN16_END = {0.5f, 0.5f,
0.5f, 0.5f};
const __m128 x_minus_half = _mm_sub_ps(x_max, *((__m128*)half));
const __m128i x_minus_half_floor = _mm_cvtps_epi32(x_minus_half);
// Compute 2^n.
static const ALIGN16_BEG int float_exponent_bias[4] ALIGN16_END = {
127, 127, 127, 127};
static const int float_exponent_shift = 23;
const __m128i two_n_exponent =
_mm_add_epi32(x_minus_half_floor, *((__m128i*)float_exponent_bias));
const __m128 two_n =
_mm_castsi128_ps(_mm_slli_epi32(two_n_exponent, float_exponent_shift));
// Compute y.
const __m128 y = _mm_sub_ps(x_max, _mm_cvtepi32_ps(x_minus_half_floor));
// Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
static const ALIGN16_BEG float C2[4] ALIGN16_END = {
3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f};
static const ALIGN16_BEG float C1[4] ALIGN16_END = {
6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f};
static const ALIGN16_BEG float C0[4] ALIGN16_END = {1.0017247f, 1.0017247f,
1.0017247f, 1.0017247f};
const __m128 exp2_y_0 = _mm_mul_ps(y, *((__m128*)C2));
const __m128 exp2_y_1 = _mm_add_ps(exp2_y_0, *((__m128*)C1));
const __m128 exp2_y_2 = _mm_mul_ps(exp2_y_1, y);
const __m128 exp2_y = _mm_add_ps(exp2_y_2, *((__m128*)C0));
// Combine parts.
a_exp_b = _mm_mul_ps(exp2_y, two_n);
}
return a_exp_b;
}
inline GPR_t si_andbi( GPR_t RA, uint8_t I10 )
{
return _mm_castsi128_ps( _mm_and_si128( _mm_castps_si128( RA ), _mm_set1_epi8( I10 ) ) );
}
GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x00000f0f, 0x00000000), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x000f0f0f, 0x00000000), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x00000000), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x0000000f), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x00000f0f), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x000f0f0f), GSVector4i(0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f), }; const GSVector4 GSVector4::m_ps0123(0.0f, 1.0f, 2.0f, 3.0f); const GSVector4 GSVector4::m_ps4567(4.0f, 5.0f, 6.0f, 7.0f); const GSVector4 GSVector4::m_half(0.5f); const GSVector4 GSVector4::m_one(1.0f); const GSVector4 GSVector4::m_two(2.0f); const GSVector4 GSVector4::m_four(4.0f); const GSVector4 GSVector4::m_x4b000000(_mm_castsi128_ps(_mm_set1_epi32(0x4b000000))); const GSVector4 GSVector4::m_x4f800000(_mm_castsi128_ps(_mm_set1_epi32(0x4f800000))); #if _M_SSE >= 0x500 const GSVector8 GSVector8::m_half(0.5f); const GSVector8 GSVector8::m_one(1.0f); const GSVector8 GSVector8::m_x7fffffff(_mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff))); const GSVector8 GSVector8::m_x80000000(_mm256_castsi256_ps(_mm256_set1_epi32(0x80000000))); const GSVector8 GSVector8::m_x4b000000(_mm256_castsi256_ps(_mm256_set1_epi32(0x4b000000))); const GSVector8 GSVector8::m_x4f800000(_mm256_castsi256_ps(_mm256_set1_epi32(0x4f800000))); #endif #if _M_SSE >= 0x501
inline GPR_t si_andi( GPR_t RA, int16_t I10 )
{
return _mm_castsi128_ps( _mm_and_si128( _mm_castps_si128( RA ), _mm_set1_epi32( static_cast<int32_t>(SignExtend( I10, 10 ) ) ) ) );
}
inline GPR_t si_ori( GPR_t RA, int64_t IMM )
{
return _mm_or_ps( RA, _mm_castsi128_ps( _mm_set1_epi32( (int32_t)IMM ) ) );
}
inline GPR_t si_nand( GPR_t RA, GPR_t RB )
{
return _mm_andnot_ps( _mm_and_ps( RA, RB ), _mm_castsi128_ps( _mm_set1_epi32( 0xffffffff ) ) );
}
__m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
int vdwioffset0;
__m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
__m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
__m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
int nvdwtype;
__m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
int *vdwtype;
real *vdwparam;
__m128 one_sixth = _mm_set1_ps(1.0/6.0);
__m128 one_twelfth = _mm_set1_ps(1.0/12.0);
__m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
real rswitch_scalar,d_scalar;
__m128 dummy_mask,cutoff_mask;
__m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
__m128 one = _mm_set1_ps(1.0);
__m128 two = _mm_set1_ps(2.0);
x = xx[0];
f = ff[0];
nri = nlist->nri;
iinr = nlist->iinr;
jindex = nlist->jindex;
jjnr = nlist->jjnr;
shiftidx = nlist->shift;
gid = nlist->gid;
shiftvec = fr->shift_vec[0];
fshift = fr->fshift[0];
nvdwtype = fr->ntype;
vdwparam = fr->nbfp;
//----------------------------------------------------------------
// Transforms the AABB vertices to screen space once every frame
// Also performs a coarse depth pre-test
//----------------------------------------------------------------
PreTestResult TransformedAABBoxAVX::TransformAndPreTestAABBox(__m128 xformedPos[], const __m128 cumulativeMatrix[4], const float *pDepthSummary)
{
// w ends up being garbage, but it doesn't matter - we ignore it anyway.
__m128 vCenter = _mm_loadu_ps(&mBBCenter.x);
__m128 vHalf = _mm_loadu_ps(&mBBHalf.x);
__m128 vMin = _mm_sub_ps(vCenter, vHalf);
__m128 vMax = _mm_add_ps(vCenter, vHalf);
// transforms
__m128 xRow[2], yRow[2], zRow[2];
xRow[0] = _mm_shuffle_ps(vMin, vMin, 0x00) * cumulativeMatrix[0];
xRow[1] = _mm_shuffle_ps(vMax, vMax, 0x00) * cumulativeMatrix[0];
yRow[0] = _mm_shuffle_ps(vMin, vMin, 0x55) * cumulativeMatrix[1];
yRow[1] = _mm_shuffle_ps(vMax, vMax, 0x55) * cumulativeMatrix[1];
zRow[0] = _mm_shuffle_ps(vMin, vMin, 0xaa) * cumulativeMatrix[2];
zRow[1] = _mm_shuffle_ps(vMax, vMax, 0xaa) * cumulativeMatrix[2];
__m128 zAllIn = _mm_castsi128_ps(_mm_set1_epi32(~0));
__m128 screenMin = _mm_set1_ps(FLT_MAX);
__m128 screenMax = _mm_set1_ps(-FLT_MAX);
for(UINT i = 0; i < AABB_VERTICES; i++)
{
// Transform the vertex
__m128 vert = cumulativeMatrix[3];
vert += xRow[sBBxInd[i]];
vert += yRow[sBByInd[i]];
vert += zRow[sBBzInd[i]];
// We have inverted z; z is in front of near plane iff z <= w.
__m128 vertZ = _mm_shuffle_ps(vert, vert, 0xaa); // vert.zzzz
__m128 vertW = _mm_shuffle_ps(vert, vert, 0xff); // vert.wwww
__m128 zIn = _mm_cmple_ps(vertZ, vertW);
zAllIn = _mm_and_ps(zAllIn, zIn);
// project
xformedPos[i] = _mm_div_ps(vert, vertW);
// update bounds
screenMin = _mm_min_ps(screenMin, xformedPos[i]);
screenMax = _mm_max_ps(screenMax, xformedPos[i]);
}
// if any of the verts are z-clipped, we (conservatively) say the box is in
if(_mm_movemask_ps(zAllIn) != 0xf)
return ePT_VISIBLE;
// Clip against screen bounds
screenMin = _mm_max_ps(screenMin, _mm_setr_ps(0.0f, 0.0f, 0.0f, -FLT_MAX));
screenMax = _mm_min_ps(screenMax, _mm_setr_ps((float) (SCREENW - 1), (float) (SCREENH - 1), 1.0f, FLT_MAX));
// Quick rejection test
if(_mm_movemask_ps(_mm_cmplt_ps(screenMax, screenMin)))
return ePT_INVISIBLE;
// Prepare integer bounds
__m128 minMaxXY = _mm_shuffle_ps(screenMin, screenMax, 0x44); // minX,minY,maxX,maxY
__m128i minMaxXYi = _mm_cvtps_epi32(minMaxXY);
__m128i minMaxXYis = _mm_srai_epi32(minMaxXYi, 3);
__m128 maxZ = _mm_shuffle_ps(screenMax, screenMax, 0xaa);
// Traverse all 8x8 blocks covered by 2d screen-space BBox;
// if we know for sure that this box is behind the geometry we know is there,
// we can stop.
int rX0 = minMaxXYis.m128i_i32[0];
int rY0 = minMaxXYis.m128i_i32[1];
int rX1 = minMaxXYis.m128i_i32[2];
int rY1 = minMaxXYis.m128i_i32[3];
__m128 anyCloser = _mm_setzero_ps();
for(int by = rY0; by <= rY1; by++)
{
const float *srcRow = pDepthSummary + by * (SCREENW/BLOCK_SIZE);
// If for any 8x8 block, maxZ is not less than (=behind) summarized
// min Z, box might be visible.
for(int bx = rX0; bx <= rX1; bx++)
{
anyCloser = _mm_or_ps(anyCloser, _mm_cmpnlt_ss(maxZ, _mm_load_ss(&srcRow[bx])));
}
if(_mm_movemask_ps(anyCloser))
{
return ePT_UNSURE; // okay, box might be in
}
}
// If we get here, we know for sure that the box is fully behind the stuff in the
// depth buffer.
return ePT_INVISIBLE;
}