SPU LLVM: Improve FNMS
Should be more accurate with postive/negative zero inputs according to docs while being more optimized. TODO: Check SPU precise interptreter.
This commit is contained in:
@@ -7465,9 +7465,9 @@ public:
|
|||||||
{
|
{
|
||||||
// See FMA.
|
// See FMA.
|
||||||
if (g_cfg.core.spu_accurate_xfloat)
|
if (g_cfg.core.spu_accurate_xfloat)
|
||||||
set_vr(op.rt4, -fmuladd(get_vr<f64[4]>(op.ra), get_vr<f64[4]>(op.rb), eval(-get_vr<f64[4]>(op.rc))));
|
set_vr(op.rt4, fmuladd(eval(-get_vr<f64[4]>(op.ra)), get_vr<f64[4]>(op.rb), get_vr<f64[4]>(op.rc)));
|
||||||
else if (g_cfg.core.spu_approx_xfloat)
|
else if (g_cfg.core.spu_approx_xfloat)
|
||||||
set_vr(op.rt4, -fma32x4(get_vr<f32[4]>(op.ra), get_vr<f32[4]>(op.rb), eval(-get_vr<f32[4]>(op.rc))));
|
set_vr(op.rt4, fma32x4(eval(-get_vr<f32[4]>(op.ra)), get_vr<f32[4]>(op.rb), get_vr<f32[4]>(op.rc)));
|
||||||
else
|
else
|
||||||
set_vr(op.rt4, get_vr<f32[4]>(op.rc) - get_vr<f32[4]>(op.ra) * get_vr<f32[4]>(op.rb));
|
set_vr(op.rt4, get_vr<f32[4]>(op.rc) - get_vr<f32[4]>(op.ra) * get_vr<f32[4]>(op.rb));
|
||||||
}
|
}
|
||||||
|
|||||||
Reference in New Issue
Block a user