summaryrefslogtreecommitdiffstats
path: root/src/video_core/shader
diff options
context:
space:
mode:
authorYuri Kunde Schlesner <yuriks@yuriks.net>2017-02-12 22:29:14 +0100
committerYuri Kunde Schlesner <yuriks@yuriks.net>2017-02-12 22:29:14 +0100
commite10b11a5d06de6efd1ccb39a0ed6bb602761df6d (patch)
tree1fef798fc03ecfd0a789cc17d7f6ab57304e3f5c /src/video_core/shader
parentMerge pull request #2550 from yuriks/pica-refactor2 (diff)
downloadyuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar.gz
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar.bz2
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar.lz
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar.xz
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.tar.zst
yuzu-e10b11a5d06de6efd1ccb39a0ed6bb602761df6d.zip
Diffstat (limited to 'src/video_core/shader')
-rw-r--r--src/video_core/shader/shader_jit_x64_compiler.cpp8
1 files changed, 8 insertions, 0 deletions
diff --git a/src/video_core/shader/shader_jit_x64_compiler.cpp b/src/video_core/shader/shader_jit_x64_compiler.cpp
index 92b35dbc0..2dbc8b147 100644
--- a/src/video_core/shader/shader_jit_x64_compiler.cpp
+++ b/src/video_core/shader/shader_jit_x64_compiler.cpp
@@ -295,14 +295,22 @@ void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
}
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);
cmpordps(scratch, src2);
mulps(src1, src2);
+ // Set src2 to mask of (result == NaN)
movaps(src2, src1);
cmpunordps(src2, src2);
+ // Clear components where scratch != src2 (i.e. if result is NaN where neither source was NaN)
xorps(scratch, src2);
andps(src1, scratch);
}