int kernel(lwp_functions* pf,
void* params,
void* inout,
unsigned int iter,
unsigned int n)
{
Ternary_params* p = (Ternary_params*)params;
switch (p->cmd)
{
case AM:
{
int length = p->length / 4;
vector float *a = (vector float *)inout;
vector float *b = a + length;
vector float *c = a + 2 * length;
unsigned int i;
for (i = 0; i != length; ++i, ++a, ++b, ++c)
*a = spu_mul(spu_add(*a, *b), *c);
return 0;
}
case MA:
{
int length = p->length / 4;
vector float *a = (vector float *)inout;
vector float *b = a + length;
vector float *c = a + 2 * length;
unsigned int i;
for (i = 0; i != length; ++i, ++a, ++b, ++c)
*a = spu_madd(*a, *b, *c);
return 0;
}
case CAM:
{
static vector unsigned char lo =
(vector unsigned char) { 0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23};
static vector unsigned char hi =
(vector unsigned char) { 8, 9, 10, 11, 24, 25, 26, 27,
12, 13, 14, 15, 28, 29, 30, 31};
int length = p->length / 4;
float *a = (float *)inout;
float *b = a + 8 * length;
float *c = a + 16 * length;
unsigned int i;
// (a + b) * c:
// r.r = (a.r+b.r)*c.r - (a.i+b.i)*c.i
// r.i = (a.r+b.r)*c.i + (a.i+b.i)*c.r
for (i = 0; i != length; ++i, a+=8, b+=8, c+=8)
{
vector float av = {*a, *(a+2), *(a+4), *(a+6)}; // a.r
vector float bv = {*b, *(b+2), *(b+4), *(b+6)}; // b.r
vector float cv = {*c, *(c+2), *(c+4), *(c+6)}; // c.r
vector float dv = {*(a+1), *(a+3), *(a+5), *(a+7)}; // a.i
vector float ev = {*(b+1), *(b+3), *(b+5), *(b+7)}; // b.i
vector float fv = {*(c+1), *(c+3), *(c+5), *(c+7)}; // c.i
vector float trv = spu_add(av, bv); // a.r+b.r
vector float tiv = spu_add(dv, ev); // a.i+b.i
vector float sv = spu_mul(trv, cv); // (a.r+b.r)*c.r
vector float tv = spu_mul(trv, fv); // (a.r+b.r)*c.i
vector float real = spu_nmsub(tiv, fv, sv); // r.r
vector float imag = spu_madd(tiv, cv, tv); // r.i
// interleave result
*(vector float *)a = spu_shuffle(real, imag, lo);
*(vector float *)(a+4) = spu_shuffle(real, imag, hi);
}
return 0;
}
case CMA:
{
static vector unsigned char lo =
(vector unsigned char) { 0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23};
static vector unsigned char hi =
(vector unsigned char) { 8, 9, 10, 11, 24, 25, 26, 27,
12, 13, 14, 15, 28, 29, 30, 31};
int length = p->length / 4;
float *a = (float *)inout;
float *b = a + 8 * length;
float *c = a + 16 * length;
unsigned int i;
// a * b + c:
// r.r = a.r*b.r + c.r - a.i*b.i
// r.i = a.r*b.i + c.i + a.i*b.r
for (i = 0; i != length; ++i, a+=8, b+=8, c+=8)
{
vector float av = {*a, *(a+2), *(a+4), *(a+6)}; // a.r
vector float bv = {*b, *(b+2), *(b+4), *(b+6)}; // b.r
vector float cv = {*c, *(c+2), *(c+4), *(c+6)}; // c.r
vector float dv = {*(a+1), *(a+3), *(a+5), *(a+7)}; // a.i
vector float ev = {*(b+1), *(b+3), *(b+5), *(b+7)}; // b.i
vector float fv = {*(c+1), *(c+3), *(c+5), *(c+7)}; // c.i
vector float real = spu_nmsub(dv, ev, spu_madd(av, bv, cv)); // r.r
vector float imag = spu_madd(dv, bv, spu_madd(av, ev, fv)); // r.i
// interleave result
*(vector float *)a = spu_shuffle(real, imag, lo);
*(vector float *)(a+4) = spu_shuffle(real, imag, hi);
}
return 0;
}
case ZAM:
{
int length = p->length / 4;
float *a_re = (float *)inout;
float *a_im = a_re + 4 * length;
float *b_re = a_re + 8 * length;
float *b_im = a_re + 12 * length;
float *c_re = a_re + 16 * length;
float *c_im = a_re + 20 * length;
unsigned int i;
// (a + b) * c:
// r.r = (a.r+b.r)*c.r - (a.i+b.i)*c.i
// r.i = (a.r+b.r)*c.i + (a.i+b.i)*c.r
for (i = 0; i != length;
++i, a_re+=4, b_re+=4, c_re+=4, a_im+=4, b_im+=4, c_im+=4)
{
vector float *av = (vector float *)a_re;
vector float *bv = (vector float *)b_re;
vector float *cv = (vector float *)c_re;
vector float *dv = (vector float *)a_im;
vector float *ev = (vector float *)b_im;
vector float *fv = (vector float *)c_im;
vector float trv = spu_add(*av, *bv); // a.r+b.r
vector float tiv = spu_add(*dv, *ev); // a.i+b.i
vector float sv = spu_mul(trv, *cv); // (a.r+b.r)*c.r
vector float tv = spu_mul(trv, *fv); // (a.r+b.r)*c.i
*av = spu_nmsub(tiv, *fv, sv); // r.r
*dv = spu_madd(tiv, *cv, tv); // r.i
}
return 0;
}
case ZMA:
{
int length = p->length / 4;
float *a_re = (float *)inout;
float *a_im = a_re + 4 * length;
float *b_re = a_re + 8 * length;
float *b_im = a_re + 12 * length;
float *c_re = a_re + 16 * length;
float *c_im = a_re + 20 * length;
unsigned int i;
// a * b + c:
// r.r = a.r*b.r + c.r - a.i*b.i
// r.i = a.r*b.i + c.i + a.i*b.r
for (i = 0; i != length;
++i, a_re+=4, b_re+=4, c_re+=4, a_im+=4, b_im+=4, c_im+=4)
{
vector float *av = (vector float *)a_re;
vector float *bv = (vector float *)b_re;
vector float *cv = (vector float *)c_re;
vector float *dv = (vector float *)a_im;
vector float *ev = (vector float *)b_im;
vector float *fv = (vector float *)c_im;
vector float tmp = spu_nmsub(*dv, *ev, spu_madd(*av, *bv, *cv));
*dv = spu_madd(*dv, *bv, spu_madd(*av, *ev, *fv));
*av = tmp;
}
return 0;
}
}
return 1;
}
vector double
__divv2df3 (vector double a_in, vector double b_in)
{
/* Variables */
vec_int4 exp, exp_bias;
vec_uint4 no_underflow, overflow;
vec_float4 mant_bf, inv_bf;
vec_ullong2 exp_a, exp_b;
vec_ullong2 a_nan, a_zero, a_inf, a_denorm, a_denorm0;
vec_ullong2 b_nan, b_zero, b_inf, b_denorm, b_denorm0;
vec_ullong2 nan;
vec_uint4 a_exp, b_exp;
vec_ullong2 a_mant_0, b_mant_0;
vec_ullong2 a_exp_1s, b_exp_1s;
vec_ullong2 sign_exp_mask;
vec_double2 a, b;
vec_double2 mant_a, mant_b, inv_b, q0, q1, q2, mult;
/* Constants */
vec_uint4 exp_mask_u32 = spu_splats((unsigned int)0x7FF00000);
vec_uchar16 splat_hi = (vec_uchar16) {
0,1,2,3, 0,1,2,3, 8, 9,10,11, 8,9,10,11
};
vec_uchar16 swap_32 = (vec_uchar16) {
4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11
};
vec_ullong2 exp_mask = spu_splats(0x7FF0000000000000ULL);
vec_ullong2 sign_mask = spu_splats(0x8000000000000000ULL);
vec_float4 onef = spu_splats(1.0f);
vec_double2 one = spu_splats(1.0);
vec_double2 exp_53 = (vec_double2)spu_splats(0x0350000000000000ULL);
sign_exp_mask = spu_or(sign_mask, exp_mask);
/* Extract the floating point components from each of the operands including
* exponent and mantissa.
*/
a_exp = (vec_uint4)spu_and((vec_uint4)a_in, exp_mask_u32);
a_exp = spu_shuffle(a_exp, a_exp, splat_hi);
b_exp = (vec_uint4)spu_and((vec_uint4)b_in, exp_mask_u32);
b_exp = spu_shuffle(b_exp, b_exp, splat_hi);
a_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)a_in, sign_exp_mask), 0);
a_mant_0 = spu_and(a_mant_0, spu_shuffle(a_mant_0, a_mant_0, swap_32));
b_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)b_in, sign_exp_mask), 0);
b_mant_0 = spu_and(b_mant_0, spu_shuffle(b_mant_0, b_mant_0, swap_32));
a_exp_1s = (vec_ullong2)spu_cmpeq(a_exp, exp_mask_u32);
b_exp_1s = (vec_ullong2)spu_cmpeq(b_exp, exp_mask_u32);
/* Identify all possible special values that must be accommodated including:
* +-denorm, +-0, +-infinity, and NaNs.
*/
a_denorm0= (vec_ullong2)spu_cmpeq(a_exp, 0);
a_nan = spu_andc(a_exp_1s, a_mant_0);
a_zero = spu_and (a_denorm0, a_mant_0);
a_inf = spu_and (a_exp_1s, a_mant_0);
a_denorm = spu_andc(a_denorm0, a_zero);
b_denorm0= (vec_ullong2)spu_cmpeq(b_exp, 0);
b_nan = spu_andc(b_exp_1s, b_mant_0);
b_zero = spu_and (b_denorm0, b_mant_0);
b_inf = spu_and (b_exp_1s, b_mant_0);
b_denorm = spu_andc(b_denorm0, b_zero);
/* Scale denorm inputs to into normalized numbers by conditionally scaling the
* input parameters.
*/
a = spu_sub(spu_or(a_in, exp_53), spu_sel(exp_53, a_in, sign_mask));
a = spu_sel(a_in, a, a_denorm);
b = spu_sub(spu_or(b_in, exp_53), spu_sel(exp_53, b_in, sign_mask));
b = spu_sel(b_in, b, b_denorm);
/* Extract the divisor and dividend exponent and force parameters into the signed
* range [1.0,2.0) or [-1.0,2.0).
*/
exp_a = spu_and((vec_ullong2)a, exp_mask);
exp_b = spu_and((vec_ullong2)b, exp_mask);
mant_a = spu_sel(a, one, (vec_ullong2)exp_mask);
mant_b = spu_sel(b, one, (vec_ullong2)exp_mask);
/* Approximate the single reciprocal of b by using
* the single precision reciprocal estimate followed by one
* single precision iteration of Newton-Raphson.
*/
mant_bf = spu_roundtf(mant_b);
inv_bf = spu_re(mant_bf);
inv_bf = spu_madd(spu_nmsub(mant_bf, inv_bf, onef), inv_bf, inv_bf);
/* Perform 2 more Newton-Raphson iterations in double precision. The
* result (q1) is in the range (0.5, 2.0).
*/
inv_b = spu_extend(inv_bf);
inv_b = spu_madd(spu_nmsub(mant_b, inv_b, one), inv_b, inv_b);
q0 = spu_mul(mant_a, inv_b);
q1 = spu_madd(spu_nmsub(mant_b, q0, mant_a), inv_b, q0);
/* Determine the exponent correction factor that must be applied
* to q1 by taking into account the exponent of the normalized inputs
* and the scale factors that were applied to normalize them.
*/
exp = spu_rlmaska(spu_sub((vec_int4)exp_a, (vec_int4)exp_b), -20);
exp = spu_add(exp, (vec_int4)spu_add(spu_and((vec_int4)a_denorm, -0x34), spu_and((vec_int4)b_denorm, 0x34)));
/* Bias the quotient exponent depending on the sign of the exponent correction
* factor so that a single multiplier will ensure the entire double precision
* domain (including denorms) can be achieved.
*
* exp bias q1 adjust exp
* ===== ======== ==========
* positive 2^+65 -65
* negative 2^-64 +64
*/
exp_bias = spu_xor(spu_rlmaska(exp, -31), 64);
exp = spu_sub(exp, exp_bias);
q1 = spu_sel(q1, (vec_double2)spu_add((vec_int4)q1, spu_sl(exp_bias, 20)), exp_mask);
/* Compute a multiplier (mult) to applied to the quotient (q1) to produce the
* expected result. On overflow, clamp the multiplier to the maximum non-infinite
* number in case the rounding mode is not round-to-nearest.
*/
exp = spu_add(exp, 0x3FF);
no_underflow = spu_cmpgt(exp, 0);
overflow = spu_cmpgt(exp, 0x7FE);
exp = spu_and(spu_sl(exp, 20), (vec_int4)no_underflow);
exp = spu_and(exp, (vec_int4)exp_mask);
mult = spu_sel((vec_double2)exp, (vec_double2)(spu_add((vec_uint4)exp_mask, -1)), (vec_ullong2)overflow);
/* Handle special value conditions. These include:
*
* 1) IF either operand is a NaN OR both operands are 0 or INFINITY THEN a NaN
* results.
* 2) ELSE IF the dividend is an INFINITY OR the divisor is 0 THEN a INFINITY results.
* 3) ELSE IF the dividend is 0 OR the divisor is INFINITY THEN a 0 results.
*/
mult = spu_andc(mult, (vec_double2)spu_or(a_zero, b_inf));
mult = spu_sel(mult, (vec_double2)exp_mask, spu_or(a_inf, b_zero));
nan = spu_or(a_nan, b_nan);
nan = spu_or(nan, spu_and(a_zero, b_zero));
nan = spu_or(nan, spu_and(a_inf, b_inf));
mult = spu_or(mult, (vec_double2)nan);
/* Scale the final quotient */
q2 = spu_mul(q1, mult);
return (q2);
}
