grumpycoders/pcsx-redux
1
0
Fork 0
mirror of https://github.com/grumpycoders/pcsx-redux.git synced 2026-10-12 02:15:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

spu tests: a negative per-voice-volume fixture for the I2S rig

Browse Source 
Every revstream fixture so far has set voice volume to a positive value, so
the fixed-mode sign bit has never been put in front of silicon. NEGVOICE
sets voice 1 to +1024 left and 0x7c00 (-1024) right: same magnitude,
opposite sign, well below saturation so no clamp can enter the answer.
Pair it with DRY, which removes the reverb tick-phase draw and with it the
3-run vote. Under the fixed decode Redux returns max|L+R| = 0 and
corr(L,R) = -1.000000; an unfixed build renders the right channel at +15360
instead of -1024.

Signed-off-by: Nicolas 'Pixel' Noble <[email protected]>
This commit is contained in:
Nicolas 'Pixel' Noble
2026-09-04 20:56:31 -07:00
parent 1591bc9ba3
commit 32f6e8192b
2 changed files with 25 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/mips/tests/spu-revstream/Makefile
+4
View File
@@ -41,6 +41,10 @@ ifeq ($(NEGOUT),true)
CPPFLAGS += -DREVSTREAM_NEGOUT
endif
ifeq ($(NEGVOICE),true)
CPPFLAGS += -DREVSTREAM_NEGVOICE
endif
ifeq ($(PITCH),true)
CPPFLAGS += -DREVSTREAM_PITCH
endif
src/mips/tests/spu-revstream/revstream.c
+21 -1
View File
@@ -445,7 +445,27 @@ int main(void) {
for (unsigned i = 0; i < 32; i++) regs[i] = SPU_REVERB_PRESET[i];
}
#ifdef REVSTREAM_PANNED
#if defined(REVSTREAM_NEGVOICE)
/* PER-VOICE VOLUME, NEGATIVE ON ONE SIDE. Every fixture in this arc has set
voice volume to a positive value, so the fixed-mode sign bit (bit 14, with
bit 15 clear) has never been exercised against silicon at all. NEGOUT
covers the reverb OUTPUT volume's sign; this covers the voice's.
0x0400 = +1024 left. 0x7c00 = bit15 clear so fixed mode, bit14 set so
negative: 0x7c00 - 0x8000 = -1024 right. Same magnitude, opposite sign, so
silicon's two channels should come back in antiphase at equal level and
the question is decided by one subtraction. Deliberately not full scale -
a small level keeps the mix far from the saturation point, so nothing in
the answer can be a clamp.
This is the fixture for the `vol &= 0x3fff` bug: that mask threw away the
sign the decode had just reconstructed, so an unfixed build renders the
right channel at +15360 rather than -1024, fifteen times the level and in
phase. Pair it with DRY=true - the reverb is not under test here, and a dry
capture has no 22.05kHz tick phase to draw, so it needs no 3-run vote. */
SPU_VOICES[1].volumeLeft = 0x0400;
SPU_VOICES[1].volumeRight = 0x7c00;
#elif defined(REVSTREAM_PANNED)
/* Hard-left voice, so Lin != Rin. EVERY capture in this arc until 2026-08-14
used a symmetric stimulus - dry L and dry R came back byte-identical - and
with L == R the different-side reflection produces the same thing as the