lvella/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

video_core/shader: Document sanitized MUL operation

Browse source
This commit is contained in:
Yuri Kunde Schlesner 2017-02-12 13:29:14 -08:00
parent 443bb3d522
commit e10b11a5d0
1 changed files with 8 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
src/video_core/shader/shader_jit_x64_compiler.cpp
View file

@ -295,14 +295,22 @@ void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
} }
void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) { void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
// 0 * inf and inf * 0 in the PICA should return 0 instead of NaN. This can be implemented by
// checking for NaNs before and after the multiplication. If the multiplication result is NaN
// where neither source was, this NaN was generated by a 0 * inf multiplication, and so the
// result should be transformed to 0 to match PICA fp rules.
// Set scratch to mask of (src1 != NaN and src2 != NaN)
movaps(scratch, src1); movaps(scratch, src1);
cmpordps(scratch, src2); cmpordps(scratch, src2);
mulps(src1, src2); mulps(src1, src2);
// Set src2 to mask of (result == NaN)
movaps(src2, src1); movaps(src2, src1);
cmpunordps(src2, src2); cmpunordps(src2, src2);
// Clear components where scratch != src2 (i.e. if result is NaN where neither source was NaN)
xorps(scratch, src2); xorps(scratch, src2);
andps(src1, scratch); andps(src1, scratch);
} }