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: stop comparing the warmup region against a golden

Browse Source 
PcmTestHeader.warmup declares its prefix to be decode artifacts carried in
from whatever played before, so it is not reproducible across runs. The
comparison started at SPU_ONSET_SKIP and read almost all of it anyway, which
made a capture assert against the state its predecessor happened to leave
behind. That is why a golden could match silicon exactly on everything
repeatable and still fail: all four adpcm_edge predictor goldens are exact
rotations of a fresh hardware capture on their periodic tails and differ only
inside warmup.

SPU_ONSET_SKIP stays as a floor. Nothing starts passing - the failure count is
unchanged - but the residuals stop being masked: nine captures now read a
uniform -1, where before the first mismatch was warmup noise.

Signed-off-by: Nicolas 'Pixel' Noble <[email protected]>
This commit is contained in:
Nicolas 'Pixel' Noble
2026-08-03 09:11:53 -07:00
parent 841ad281df
commit 703ad59f6e
1 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/mips/tests/spu/spu.c
+11 -3
View File
@@ -303,6 +303,14 @@ static int spu_compare_golden(const char *name, const void *cap, const uint8_t *
const int warmupS = (int)(h->warmup / 2);
const int periodS = (int)(h->period / 2);
const int keptS = warmupS + periodS;
// The header declares the first warmupS samples to be decode artifacts carried
// in from whatever played before, so they are not reproducible across runs and
// never were: two captures of the same voice agree on the periodic tail and
// disagree here by construction. Comparing them made a capture assert against
// the state its predecessor happened to leave behind, which is why a golden
// could match silicon exactly on everything repeatable and still fail.
// SPU_ONSET_SKIP is a floor, not the start; the honest start is past the warmup.
const int startS = warmupS > SPU_ONSET_SKIP ? warmupS : SPU_ONSET_SKIP;
// A golden that carries no samples makes both loops below run zero iterations,
// so the test reports PASS having compared nothing - it cannot fail, which is
@@ -318,7 +326,7 @@ static int spu_compare_golden(const char *name, const void *cap, const uint8_t *
name, keptS, (int)h->warmup, (int)h->period);
return 1;
}
if (keptS <= SPU_ONSET_SKIP) {
if (keptS <= startS) {
// The golden is shorter than the number of leading samples the comparison
// throws away, so the loop below runs zero iterations. Note this is not
// "the golden is too short" on its own - it is an interaction between two
@@ -334,13 +342,13 @@ static int spu_compare_golden(const char *name, const void *cap, const uint8_t *
int bestS = 0, bestBad = 0x7fffffff;
for (int s = 0; s < 512; s++) {
int bad = 0;
for (int i = SPU_ONSET_SKIP; i < keptS; i++) {
for (int i = startS; i < keptS; i++) {
if (a[(s + i) & 511] != golden[i]) bad++;
}
if (bad < bestBad) { bestBad = bad; bestS = s; }
}
for (int i = SPU_ONSET_SKIP; i < keptS; i++) {
for (int i = startS; i < keptS; i++) {
int j = (bestS + i) & 511;
if (a[j] != golden[i]) {
ramsyscall_printf("%s mismatch at sample %d (ring offset %d): got 0x%04x, want 0x%04x\n",