diff options
Diffstat (limited to 'indra/newview/llviewertexture.cpp')
| -rw-r--r-- | indra/newview/llviewertexture.cpp | 195 |
1 files changed, 160 insertions, 35 deletions
diff --git a/indra/newview/llviewertexture.cpp b/indra/newview/llviewertexture.cpp index 727c0510f6..e55e4016ec 100644 --- a/indra/newview/llviewertexture.cpp +++ b/indra/newview/llviewertexture.cpp @@ -88,7 +88,8 @@ S32 LLViewerTexture::sAuxCount = 0; LLFrameTimer LLViewerTexture::sEvaluationTimer; F32 LLViewerTexture::sDesiredDiscardBias = 0.f; F32 LLViewerTexture::sBackgroundFactor = 0.f; -F32 LLViewerTexture::sMemoryPressureFactor = 0.f; +F32 LLViewerTexture::sMemoryPressureMultiplier = 1.f; +F32 LLViewerTexture::sLastDitchMinDiscard = 0.f; U32 LLViewerTexture::sBiasTexturesUpdated = 0; S32 LLViewerTexture::sMaxSculptRez = 128; //max sculpt image size @@ -532,26 +533,119 @@ void LLViewerTexture::updateClass() F32 over_pct = (used - target) / target; - // VRAM-pressure feedback loop with progressive backoff. Ramp starts at - // backoff_start x target, not at the budget cliff. Proportional in both - // directions: ramp = (over-1)(1-factor), decay = factor; converges at - // factor = 1 - 1/over_at_backoff_target. + // VRAM-pressure controller. Drives sMemoryPressureMultiplier from + // PREDICTED VRAM (used + in-flight refetch growth - in-flight downscale + // shrinkage) rather than instantaneous used. The feedback loop chasing + // instantaneous VRAM sawtooths because eviction is fast but refetch is + // slow; the prediction lets the controller see the equilibrium directly, + // so mult converges to a target value instead of cycling. { static LLCachedControl<F32> backoff_start(gSavedSettings, "TextureMemoryPressureBackoffStart", 0.85f); + static LLCachedControl<F32> max_mult(gSavedSettings, "TextureMemoryPressureMaxMultiplier", 64.f); + static LLCachedControl<F32> prediction_gain(gSavedSettings, "TextureMemoryPressurePredictionGain", 10.f); + static LLCachedControl<F32> smoothing_rate(gSavedSettings, "TextureMemoryPressureSmoothingRate", 4.f); + F32 backoff_target = target * llclamp((F32)backoff_start, 0.05f, 1.f); - F32 over = used / llmax(backoff_target, 1.f); - static LLCachedControl<F32> pressure_ramp_rate(gSavedSettings, "TextureMemoryPressureRampRate", 3.0f); - F32 dt = gFrameIntervalSeconds; - if (over > 1.f) + + // Walk the texture list once per frame, summing pending size deltas. + // Approximation: bytes(d) = (w>>d) * (h>>d) * 4 * 4/3. Units match + // `used` after the /1024/512 conversion that produced it. Coarse but + // proportionally correct - the gain knob tunes absolute magnitude. + S64 pending_bytes_increase = 0; + S64 pending_bytes_decrease = 0; { - sMemoryPressureFactor += - (over - 1.f) * (1.f - sMemoryPressureFactor) * (F32)pressure_ramp_rate * dt; + LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("vt - in-flight predict"); + for (auto& imagep : gTextureList) + { + if (imagep.isNull()) continue; + S32 fw = imagep->getFullWidth(); + S32 fh = imagep->getFullHeight(); + if (fw <= 0 || fh <= 0) continue; + S32 desired = imagep->getDesiredDiscardLevel(); + S32 current = imagep->getDiscardLevel(); + if (desired < 0 || current < 0 || desired == current) continue; + + S32 wd = llmax(1, fw >> desired); + S32 hd = llmax(1, fh >> desired); + S32 wc = llmax(1, fw >> current); + S32 hc = llmax(1, fh >> current); + // bpp=4, mip pyramid overhead 4/3 + S64 size_d = (S64)wd * hd * 4 * 4 / 3; + S64 size_c = (S64)wc * hc * 4 * 4 / 3; + + if (desired < current) + pending_bytes_increase += (size_d - size_c); + else + pending_bytes_decrease += (size_c - size_d); + } } - else + + // Match the unit conversion used to produce `used` (texture_bytes_alloc + // divided by 1024/512). 1024*512 = 524288. + constexpr F32 BYTES_TO_USED_UNITS = 1.f / 524288.f; + F32 predicted_used = used + + (F32)pending_bytes_increase * BYTES_TO_USED_UNITS + - (F32)pending_bytes_decrease * BYTES_TO_USED_UNITS; + F32 predicted_over = predicted_used / llmax(backoff_target, 1.f); + + // Direct target: at over=1, mult=1; growth governed by gain power. + // Smooth toward target so per-frame prediction noise doesn't jolt + // the controller. Lerp rate set so a step change in target reaches + // ~63% in 1/smoothing_rate seconds (default 0.25s). + F32 cap = llmax((F32)max_mult, 1.0001f); + F32 target_mult = llclamp(powf(llmax(predicted_over, 1.f), llmax((F32)prediction_gain, 0.0001f)), 1.f, cap); + F32 dt = (F32)gFrameIntervalSeconds; + F32 alpha = 1.f - expf(-llmax(dt, 0.f) * llmax((F32)smoothing_rate, 0.f)); + sMemoryPressureMultiplier += (target_mult - sMemoryPressureMultiplier) * alpha; + sMemoryPressureMultiplier = llclamp(sMemoryPressureMultiplier, 1.f, cap); + + // Last-ditch global discard floor. Ramps up when mult is pegged near + // cap and we are still over budget; decays toward 0 when fitting. Once + // the normal compression knobs are exhausted, this creeps the discard + // floor up integer step by step, mirroring how sDesiredDiscardBias + // pushes background textures. + { + static LLCachedControl<F32> ld_engage(gSavedSettings, "TextureLastDitchEngageProgress", 0.95f); + static LLCachedControl<F32> ld_ramp(gSavedSettings, "TextureLastDitchRampRate", 0.5f); + static LLCachedControl<F32> ld_decay(gSavedSettings, "TextureLastDitchDecayRate", 0.5f); + static LLCachedControl<F32> ld_max(gSavedSettings, "TextureLastDitchMinDiscardMax", 13.f); + F32 progress = llclampf((sMemoryPressureMultiplier - 1.f) / (cap - 1.f)); + bool mult_saturated = progress >= llclampf((F32)ld_engage); + if (mult_saturated && predicted_over > 1.f) + { + sLastDitchMinDiscard += llmax((F32)ld_ramp, 0.f) * dt; + } + else if (predicted_over < 1.f) + { + sLastDitchMinDiscard -= llmax((F32)ld_decay, 0.f) * dt; + } + sLastDitchMinDiscard = llclamp(sLastDitchMinDiscard, 0.f, llmax((F32)ld_max, 0.f)); + } + + F32 over = used / llmax(backoff_target, 1.f); // legacy alias used below + + // Throttled bisection log. Once per second. + static LLFrameTimer s_pressure_log_timer; + if (s_pressure_log_timer.getElapsedTimeF32() > 1.f) { - sMemoryPressureFactor -= sMemoryPressureFactor * (F32)pressure_ramp_rate * dt; + s_pressure_log_timer.reset(); + F32 mult_progress = llclampf((sMemoryPressureMultiplier - 1.f) / (cap - 1.f)); + LL_INFOS("TextureStream") << "pressure" + << " mult=" << sMemoryPressureMultiplier + << " target_mult=" << target_mult + << " progress=" << mult_progress + << " used=" << used + << " predicted=" << predicted_used + << " target=" << target + << " over=" << over + << " pred_over=" << predicted_over + << " in+=" << (S32)(pending_bytes_increase / 1024 / 1024) + << "MB in-=" << (S32)(pending_bytes_decrease / 1024 / 1024) + << "MB bias=" << sDesiredDiscardBias + << " ldmin=" << sLastDitchMinDiscard + << " dsq=" << (S32)gTextureList.mDownScaleQueue.size() + << LL_ENDL; } - sMemoryPressureFactor = llclampf(sMemoryPressureFactor); } bool is_sys_low = isSystemMemoryLow(); @@ -2181,12 +2275,11 @@ bool LLViewerFetchedTexture::updateFetch() make_request = false; } else if (mDesiredDiscardLevel > (S32)mCodecMaxDiscardLevel && - current_discard >= 0 && - current_discard <= (S32)mCodecMaxDiscardLevel) + current_discard >= 0) { - // scaleDown can serve this from the GL pyramid. (If current is - // already past codec_max, fall through so a zoom-in can rebuild — - // scaleDown only goes deeper.) + // Desired is past codec_max. Only scaleDown can satisfy it. + // Applies even when current is also past codec_max (post-scaleDown); + // re-fetching at codec_max then scaleDown-ing again is pure thrash. LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("vftuf - desired > codec max"); make_request = false; } @@ -3177,6 +3270,16 @@ void LLViewerLODTexture::processTextureStats() F32 combined = distance_factor * size_factor; + // VRAM pressure: multiply the combined signal and clamp to 0..1. + // Compresses the effective draw range and picks up close-coverage + // textures (small combined) too. Applied before the channel + // exponent so subsequent transforms see a normalized 0..1 value. + // Avatar bakes exempt. + if (!isAgentAvatarBoost(mBoostLevel) && sMemoryPressureMultiplier > 1.f) + { + combined = llmin(combined * sMemoryPressureMultiplier, 1.f); + } + // Per-channel exponent. 1.0 = baseline; <1.0 pushes combined // toward 1 (max attenuation) faster. Edges are preserved: // pow(0, p) = 0, pow(1, p) = 1. @@ -3225,16 +3328,6 @@ void LLViewerLODTexture::processTextureStats() combined = llmax(combined, bg); } - // VRAM pressure: pow(combined, 1 - factor) bends the curve - // without flattening it - pow(0, p) = 0 so close textures - // (combined ~ 0) stay near 0 while mid/far push toward 1. - // Avatar bakes exempt. - if (!isAgentAvatarBoost(mBoostLevel) && sMemoryPressureFactor > 0.f) - { - F32 pressure_exp = llmax(1.f - sMemoryPressureFactor, 0.0001f); - combined = powf(combined, pressure_exp); - } - discard_level = combined * dim_max_for_image; } @@ -3252,21 +3345,36 @@ void LLViewerLODTexture::processTextureStats() mDesiredDiscardLevel = llmin(effective_cap, (S32)discard_level); - // Apply the setMinDiscardLevel cap, relaxed proportionally under - // VRAM pressure - at factor=1 the cap reaches dim_max so capped - // textures (terrain, etc.) participate fully in eviction. Caps of - // 0 (thumbnails) and avatar bakes are preserved. + // Apply the setMinDiscardLevel cap, relaxed under VRAM pressure + // (cap_relax = 1 - 1/mult: 0 at mult=1, ~0.5 at mult=2, ~0.9 at + // mult=10). Caps of 0 (thumbnails) and avatar bakes are preserved. S32 effective_min_cap = mMinDesiredDiscardLevel; - if (sMemoryPressureFactor > 0.f && + if (sMemoryPressureMultiplier > 1.f && mMinDesiredDiscardLevel > 0 && mMinDesiredDiscardLevel < S8_MAX && !isAgentAvatarBoost(mBoostLevel)) { + F32 cap_relax = 1.f - 1.f / sMemoryPressureMultiplier; F32 room = (F32)dim_max_for_image_i - (F32)mMinDesiredDiscardLevel; - effective_min_cap += (S32)(sMemoryPressureFactor * room); + effective_min_cap += (S32)(cap_relax * room); effective_min_cap = llmin(effective_min_cap, dim_max_for_image_i); } mDesiredDiscardLevel = llmin((S8)effective_min_cap, mDesiredDiscardLevel); + // Last-ditch global discard floor, driven by sLastDitchMinDiscard. + // That state creeps up step-by-step while mult is pegged and we are + // still over budget (see updateClass), and decays when we fit. Force + // every non-avatar-bake texture to at least floor(sLastDitchMinDiscard) + // discard, capped at the per-texture max. + if (!isAgentAvatarBoost(mBoostLevel)) + { + S32 forced = (S32)floorf(sLastDitchMinDiscard); + forced = llclamp(forced, 0, dim_max_for_image_i); + if (forced > mDesiredDiscardLevel) + { + mDesiredDiscardLevel = (S8)forced; + } + } + // // At this point we've calculated the quality level that we want, @@ -3310,6 +3418,23 @@ bool LLViewerLODTexture::scaleDown() return false; } + // Hard structural blocks only. Per-texture policy (icons pinned to full + // res, etc.) lives in processTextureStats; if that policy is later + // relaxed (e.g. honor mKnownDrawWidth for icons rendered at 8x8 in a + // friend list) the scaleDown path stays open. + if (!mUseMipMaps || mDontDiscard || mBoostLevel >= LLGLTexture::BOOST_HIGH) + { + // No mip pyramid to drop into, texture is explicitly pinned full res, + // or BOOST_HIGH+ emergency-out (currently only the GLTF "force full + // res" hack hits this). + return false; + } + // Avatar bakes are exempt from mid-bake eviction (cloud avatar risk). + if (isAgentAvatarBoost(mBoostLevel)) + { + return false; + } + if (!mDownScalePending) { mDownScalePending = true; |