void check_special_values() {
V4SF vx;
vx.f[0] = -1000;
vx.f[1] = -100;
vx.f[2] = 100;
vx.f[3] = 1000;
printf("exp("); print4(vx.v); printf(") = "); print4(exp_ps(vx.v)); printf("\n");
vx.f[0] = QNAN.f;
vx.f[1] = PINF.f;
vx.f[2] = MINF.f;
vx.f[3] = QNAN2.f;
printf("exp("); print4(vx.v); printf(") = "); print4(exp_ps(vx.v)); printf("\n");
vx.f[0] = 0;
vx.f[1] = -10;
vx.f[2] = 1e30f;
vx.f[3] = 1e-42f;
printf("log("); print4(vx.v); printf(") = "); print4(log_ps(vx.v)); printf("\n");
vx.f[0] = QNAN.f;
vx.f[1] = PINF.f;
vx.f[2] = MINF.f;
vx.f[3] = QNAN2.f;
printf("log("); print4(vx.v); printf(") = "); print4(log_ps(vx.v)); printf("\n");
printf("sin("); print4(vx.v); printf(") = "); print4(sin_ps(vx.v)); printf("\n");
printf("cos("); print4(vx.v); printf(") = "); print4(cos_ps(vx.v)); printf("\n");
vx.f[0] = -1e30;
vx.f[1] = -100000;
vx.f[2] = 1e30;
vx.f[3] = 100000;
printf("sin("); print4(vx.v); printf(") = "); print4(sin_ps(vx.v)); printf("\n");
printf("cos("); print4(vx.v); printf(") = "); print4(cos_ps(vx.v)); printf("\n");
}
// SIMD cos
__SIMD _SIMD_cos_ps(__SIMD a)
{
#ifdef USE_SSE
return cos_ps(a);
#elif defined USE_AVX
return cos256_ps(a);
#endif
}
float Cos(float angleRadians)
{
#ifdef MATH_USE_SINCOS_LOOKUPTABLE
return cos_lookuptable(angleRadians);
#elif defined(MATH_SSE2)
// Do range reduction by 2pi before calling cos - this enchances precision of cos_ps a lot
return s4f_x(cos_ps(modf_ps(setx_ps(angleRadians), pi2)));
#else
return cosf(angleRadians);
#endif
}
//-----------------------------------------------------------------------------------
ArrayReal MathlibNEON::Cos4( ArrayReal x )
{
// Map arbitrary angle x to the range [-pi; +pi] without using division.
// Code taken from MSDN's HLSL trick. Architectures with fused mad (i.e. NEON)
// can replace the add, the sub, & the two muls for two mad
ArrayReal integralPart;
x = vaddq_f32( vmulq_f32( x, ONE_DIV_2PI ), HALF );
x = Modf4( x, integralPart );
x = vsubq_f32( vmulq_f32( x, TWO_PI ), PI );
return cos_ps( x );
}
ArrayFloat MathlibSSE2::Cos4( ArrayFloat x )
{
// Map arbitrary angle x to the range [-pi; +pi] without using division.
// Code taken from MSDN's HLSL trick. Architectures with fused mad (i.e. NEON)
// can replace the add, the sub, & the two muls for two mad
ArrayFloat integralPart;
x = _mm_add_ps( _mm_mul_ps( x, ONE_DIV_2PI ), HALF );
x = Modf4( x, integralPart );
x = _mm_sub_ps( _mm_mul_ps( x, TWO_PI ), PI );
return cos_ps( x );
}
int check_sincos_precision(float xmin, float xmax) {
unsigned nb_trials = 100000;
printf("checking sines on [%g*Pi, %g*Pi]\n", xmin, xmax);
float max_err_sin_ref = 0, max_err_sin_cep = 0, max_err_sin_x = 0;
float max_err_cos_ref = 0, max_err_cos_cep = 0, max_err_cos_x = 0;
float max_err_sum_sqr_test = 0;
float max_err_sum_sqr_ref = 0;
xmin *= M_PI; xmax *= M_PI;
unsigned i;
for (i=0; i < nb_trials; ++i) {
V4SF vx, sin4, cos4, sin4_2, cos4_2;
vx.f[0] = i*(xmax-xmin)/(nb_trials-1) + xmin;
vx.f[1] = (i+.5)*(xmax-xmin)/(nb_trials-1) + xmin;
vx.f[2] = frand()*(xmax-xmin);
vx.f[3] = (i / 32)*M_PI/((i%32)+1);
if (vx.f[3] < xmin || vx.f[3] > xmax) vx.f[3] = frand()*(xmax-xmin);
/*
vx.f[0] = M_PI/2;
vx.f[1] = M_PI;
vx.f[2] = M_PI/3;
vx.f[3] = M_PI/4;
*/
sin4.v = sin_ps(vx.v);
cos4.v = cos_ps(vx.v);
sincos_ps(vx.v, &sin4_2.v, &cos4_2.v);
unsigned j;
for (j=0; j < 4; ++j) {
float x = vx.f[j];
float sin_test = sin4.f[j];
float cos_test = cos4.f[j];
if (sin_test != sin4_2.f[j]) {
printf("sin / sincos mismatch at x=%g\n", x);
exit(1); return 1;
}
if (cos_test != cos4_2.f[j]) {
printf("cos / sincos mismatch at x=%g\n", x);
return 1;
}
float sin_ref = sinf(x);
float sin_cep = cephes_sinf(x);
float err_sin_ref = fabs(sin_ref - sin_test);
float err_sin_cep = fabs(sin_cep - sin_test);
if (err_sin_ref > max_err_sin_ref) {
max_err_sin_ref = err_sin_ref;
max_err_sin_x = x;
}
max_err_sin_cep = MAX(max_err_sin_cep, err_sin_cep);
float cos_ref = cosf(x);
float cos_cep = cephes_cosf(x);
float err_cos_ref = fabs(cos_ref - cos_test);
float err_cos_cep = fabs(cos_cep - cos_test);
if (err_cos_ref > max_err_cos_ref) {
max_err_cos_ref = err_cos_ref;
max_err_cos_x = x;
}
max_err_cos_cep = MAX(max_err_cos_cep, err_cos_cep);
float err_sum_sqr_test = fabs(1 - cos_test*cos_test - sin_test*sin_test);
float err_sum_sqr_ref = fabs(1 - cos_ref*cos_ref - sin_ref*sin_ref);
max_err_sum_sqr_ref = MAX(max_err_sum_sqr_ref, err_sum_sqr_ref);
max_err_sum_sqr_test = MAX(max_err_sum_sqr_test, err_sum_sqr_test);
//printf("sin(%g) = %g %g err=%g\n", x, sin_ref, sin_test, err_sin_ref);
}
}
printf("max deviation from sinf(x): %g at %14.12g*Pi, max deviation from cephes_sin(x): %g\n",
max_err_sin_ref, max_err_sin_x/M_PI, max_err_sin_cep);
printf("max deviation from cosf(x): %g at %14.12g*Pi, max deviation from cephes_cos(x): %g\n",
max_err_cos_ref, max_err_cos_x/M_PI, max_err_cos_cep);
printf("deviation of sin(x)^2+cos(x)^2-1: %g (ref deviation is %g)\n",
max_err_sum_sqr_test, max_err_sum_sqr_ref);
if (max_err_sum_sqr_ref < 2e-7 && max_err_sin_ref < 2e-7 && max_err_cos_ref < 2e-7) {
printf(" ->> precision OK for the sin_ps / cos_ps / sincos_ps <<-\n\n");
return 0;
} else {
printf("\n WRONG PRECISION !! there is a problem\n\n");
return 1;
}
}