mirror of
https://github.com/MobileGL-Dev/MobileGL
synced 2026-11-07 01:53:24 +09:00
[Docs] close Magma RD32 cache regression with host and device evidence
This commit is contained in:
@@ -36,6 +36,13 @@ their existing waits; a test that removes all waits would violate the protocol.
|
||||
is keyed to submit indices, not an assumed number of frames or client progress.
|
||||
Swapchain replacement/shutdown drains these after device idle, including abandoned
|
||||
recordings. Independent blits never overwrite a descriptor set still in flight.
|
||||
Wire sampled views and draw framebuffer objects may be reused within a frame slot:
|
||||
exact keys include native storage identity/generation and every view/window value,
|
||||
and the slot is cleared only after its actual fence completes. Overflow still uses
|
||||
submit-index retirement. A cache hit never skips the image dependencies or pass boundaries.
|
||||
Pipelines use renderer-lifetime, non-recycled identities interned from exact render-pass
|
||||
compatibility values. A creator's temporary native render-pass handle is not a wire
|
||||
pipeline cache key or an eviction proof for another compatible pass using that pipeline.
|
||||
5. GPU writes retain their reverse notifications; readback waits for actual GPU
|
||||
completion and returns owned bytes. Removing a per-draw queue idle does not
|
||||
remove explicit GL synchronization or image/buffer memory dependencies.
|
||||
|
||||
@@ -17,6 +17,10 @@
|
||||
|
||||
## 2. P5f 最终验收
|
||||
|
||||
**Magma RD32 性能补修已完成**（`70fb6689`）：inproc平均从86.85提升至116.73 FPS，
|
||||
同包monolith为115.01 FPS。重复native对象与pipeline缓存键已修正，保留同步/角色边界；
|
||||
7项新缓存场景、Vulkan同步验证和主机全门通过。见 [报告](notes/p5f/magma-rd32-performance-fix.md)。
|
||||
|
||||
**Magma run-ahead 已完成**（最终行为头 `194382c9`，2026-09-20）：异步 GPU 生命周期、
|
||||
真实 client lead、present credit 1/3、负控与 Vulkan 同步验证已通过；实机 VAO 哈希碰撞
|
||||
已修复。Redmi 同一 `.mgdebug.debug` APK 的 Magma RA=1 / RA=0、GLES RA=1 均完成
|
||||
|
||||
@@ -6,6 +6,10 @@
|
||||
> 阶段终审见 [close-review.md](close-review.md)。原 P5f 行为验收头 `cfca93c7`；
|
||||
> 后续 Magma run-ahead 的最终行为头为 `194382c9`，见下。
|
||||
>
|
||||
> **Magma RD32 性能补修（2026-09-20，`70fb6689`）**：修复重复view/framebuffer创建销毁和
|
||||
> 临时render-pass句柄造成的pipeline缓存退化；inproc **116.73 FPS**，同包monolith **115.01 FPS**。
|
||||
> 缓存回归、Vulkan同步验证及主机全门通过。见 [性能报告](magma-rd32-performance-fix.md)。
|
||||
>
|
||||
> **Magma run-ahead 已完成（2026-09-20）**：独立 readiness 已开启，GPU 提交/资源
|
||||
> 退休时序、真实排队/credit 1/3、负控、Vulkan 同步验证均通过。Redmi 同一
|
||||
> `.mgdebug.debug` APK 的 Magma RA=1 / RA=0 与 GLES RA=1 三臂均完成至少60秒
|
||||
|
||||
@@ -0,0 +1,163 @@
|
||||
{
|
||||
"production_head": "70fb6689",
|
||||
"source_artifact": {
|
||||
"source": "70fb6689",
|
||||
"apk": "C:\\Users\\geekerwan\\.codex\\tmp\\magma-perf-fix\\FCL-magma-perf-70fb6689.apk",
|
||||
"apk_sha256": "bfa352bcb3402d9a31252503bacf6fcdc21e28ffaaa929fd8c80e6f2e2781e4e",
|
||||
"library_sha256": "11f26ca1f1d6c9a3d89134c31b2348306e669b1d9ebcc2fc77c7f5def2b4f818"
|
||||
},
|
||||
"baseline_fps": {
|
||||
"Magma-inproc": 86.85293459363007,
|
||||
"Magma-monolith": 116.07722925432384
|
||||
},
|
||||
"clock_control": "stock-dvfs",
|
||||
"render_distance": 32,
|
||||
"simulation_distance": 12,
|
||||
"fan_target": 2,
|
||||
"after": [
|
||||
{
|
||||
"name": "Magma-inproc",
|
||||
"backend": "DirectVulkan",
|
||||
"transport": "inproc",
|
||||
"performance": {
|
||||
"fps": 116.73138991101129,
|
||||
"fps_median": 117.0,
|
||||
"fps_p10": 115.0,
|
||||
"fps_min": 106.0,
|
||||
"fps_max": 119.0,
|
||||
"fps_stdev": 1.8717476401816064,
|
||||
"seconds": 61.02900004386902,
|
||||
"intervals": 61,
|
||||
"frames": 7124,
|
||||
"method": "sum(reset-counter frames) / sum(actual adjacent-log seconds)"
|
||||
},
|
||||
"hardware": {
|
||||
"clock_control": "stock-dvfs",
|
||||
"samples": 11,
|
||||
"temperature_c_min": 57.0,
|
||||
"temperature_c_max": 58.1,
|
||||
"fan": {
|
||||
"fan_rpm": [
|
||||
"14487",
|
||||
"14591",
|
||||
"14676",
|
||||
"14697",
|
||||
"14804"
|
||||
],
|
||||
"fan_target": [
|
||||
"2"
|
||||
]
|
||||
},
|
||||
"frequencies": {
|
||||
"policy0_cur": {
|
||||
"min": 1555200,
|
||||
"max": 1555200
|
||||
},
|
||||
"policy6_cur": {
|
||||
"min": 1209600,
|
||||
"max": 1689600
|
||||
},
|
||||
"gpu": {
|
||||
"min": 832000000,
|
||||
"max": 900000000
|
||||
},
|
||||
"busy": {
|
||||
"min": 82,
|
||||
"max": 84
|
||||
}
|
||||
}
|
||||
},
|
||||
"boot_id": "e69d0329-2264-4a2d-9b00-e1a4ffb6154d",
|
||||
"actual_runtime": {
|
||||
"backend_names": [
|
||||
"Direct (Vulkan)"
|
||||
],
|
||||
"ipc_lines": [
|
||||
"Config: IPC ring=8MiB stage=256MiB spin=50us persistent-block=64KiB adopt-tier=2 verb-barrier=1 run-ahead=1 present-credit=1 strict=1 audit=0 role-split-state=1 affinity='auto'"
|
||||
],
|
||||
"run_ahead_armed": true,
|
||||
"caps_call_masks": [
|
||||
"0x0",
|
||||
"0x3fff00000400"
|
||||
],
|
||||
"run_ahead_cap_bit_seen": true,
|
||||
"stats_rows": 237,
|
||||
"residual_pulls_max": 0.0
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Magma-monolith",
|
||||
"backend": "DirectVulkan",
|
||||
"transport": "monolith",
|
||||
"performance": {
|
||||
"fps": 115.00581747506094,
|
||||
"fps_median": 115.0,
|
||||
"fps_p10": 113.0,
|
||||
"fps_min": 112.0,
|
||||
"fps_max": 118.0,
|
||||
"fps_stdev": 1.310503381393822,
|
||||
"seconds": 61.02300000190735,
|
||||
"intervals": 61,
|
||||
"frames": 7018,
|
||||
"method": "sum(reset-counter frames) / sum(actual adjacent-log seconds)"
|
||||
},
|
||||
"hardware": {
|
||||
"clock_control": "stock-dvfs",
|
||||
"samples": 11,
|
||||
"temperature_c_min": 55.4,
|
||||
"temperature_c_max": 58.9,
|
||||
"fan": {
|
||||
"fan_rpm": [
|
||||
"14282",
|
||||
"14363",
|
||||
"14384",
|
||||
"14487"
|
||||
],
|
||||
"fan_target": [
|
||||
"2"
|
||||
]
|
||||
},
|
||||
"frequencies": {
|
||||
"policy0_cur": {
|
||||
"min": 1363200,
|
||||
"max": 1555200
|
||||
},
|
||||
"policy6_cur": {
|
||||
"min": 1689600,
|
||||
"max": 1689600
|
||||
},
|
||||
"gpu": {
|
||||
"min": 734000000,
|
||||
"max": 900000000
|
||||
},
|
||||
"busy": {
|
||||
"min": 73,
|
||||
"max": 76
|
||||
}
|
||||
}
|
||||
},
|
||||
"boot_id": "b52edced-93b3-416e-890b-5d9a2ea4e8d2",
|
||||
"actual_runtime": {
|
||||
"backend_names": [
|
||||
"Direct (Vulkan)"
|
||||
],
|
||||
"ipc_lines": [],
|
||||
"run_ahead_armed": false,
|
||||
"caps_call_masks": [],
|
||||
"run_ahead_cap_bit_seen": false,
|
||||
"stats_rows": 250,
|
||||
"residual_pulls_max": 0.0
|
||||
}
|
||||
}
|
||||
],
|
||||
"restoration": {
|
||||
"original_world_files_unchanged": 61,
|
||||
"config": {
|
||||
"mg_env.txt": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
|
||||
"mg_transport.txt": "1d80cf637af2284a1763196d824b41f6f12bea849aa81f95ab8db0e5ce043c95",
|
||||
"options.txt": "4d1cd0cb595819d924f8d5823910b643c865f12b3f59b6900abd54b04e2ee646"
|
||||
},
|
||||
"fan_restored": true,
|
||||
"installed_library_sha256": "11f26ca1f1d6c9a3d89134c31b2348306e669b1d9ebcc2fc77c7f5def2b4f818"
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,105 @@
|
||||
# Magma RD32 inproc 性能修复
|
||||
|
||||
**已完成，2026-09-20**。生产行为头 `70fb6689`，之后仅追加 CI 脚本、契约与报告。
|
||||
原 [RD32 四组合记录](rd32-fourway.md) 的 Magma inproc 86.85 FPS / monolith 116.08 FPS
|
||||
是本次基线，保留原数据。修复后的同包对照已回到约115–117 FPS，同一场景的25%差距消失。
|
||||
|
||||
## 结果
|
||||
|
||||
| 模式 | 修复前平均 FPS | 修复后平均 FPS | 修复后区间中位数 | 修复后区间 P10 |
|
||||
|---|---:|---:|---:|---:|
|
||||
| Magma-monolith | 116.08 | 115.01 | 115.00 | 113.00 |
|
||||
| Magma-inproc | 86.85 | 116.73 | 117.00 | 115.00 |
|
||||
|
||||
inproc 相比旧实现平均提升 **34.4%**；同包 inproc 比 monolith 高约1.5%。只有单轮，
|
||||
不能据此认定稳定领先；本场景已不再呈现原先约25%的差距，两者均接近120 FPS量级。
|
||||
不宣称无上限吞吐或零架构开销。
|
||||
P10为约一秒计数区间的FPS分位，不是逐帧1% low。
|
||||
|
||||
## 原因与改动
|
||||
|
||||
实机30秒 `simpleperf cpu-clock` 采样中，服务端仍反复
|
||||
构造和销毁 native GPU 对象：`vkCreateFramebuffer` inclusive占7.91%，
|
||||
`vkDestroyFramebuffer` 6.21%，`SetupWireDraw` 21.15%。这些是各自采样份额，存在调用层级，
|
||||
不能把父子项全部相加。客户端与服务端都存在工作/等待交替，不能把两个线程CPU时间相加当帧延迟。
|
||||
|
||||
本次修复：
|
||||
|
||||
1. **采样/storage image view复用**：UniformManager按帧槽缓存完整创建参数，包括root handle/gen、
|
||||
native image与allocation epoch、format/type、swizzle、aspect及mip/layer窗口。逐字段比较，
|
||||
不把相同hash当相同视图。句柄稳定后，原descriptor内容缓存可以复用。
|
||||
2. **draw framebuffer对象复用**：每个frame slot最多缓存32个inactive pass/view/framebuffer组合，
|
||||
精确匹配附件身份、窗口、格式、samples、draw-buffer映射和交换链信息。命中仍执行原有
|
||||
End/Begin、image transition及memory dependency；溢出走原submit-index退休。
|
||||
清理依赖真实slot fence或完整idle证明，发生在texture manager释放旧image之前。
|
||||
3. **pipeline兼容性身份修复**：原PipelineFactory把临时 `VkRenderPass` 句柄放进缓存key，
|
||||
跨帧重建容易重复创建pipeline，句柄回收也存在旧key别名风险。现在用精确formats/samples、
|
||||
有序color refs/holes与depth ref分配非零身份，renderer生命周期内不复用ID，
|
||||
与native句柄域分离。原句柄只供Vulkan创建调用；wire pipeline不再按creator pass的销毁
|
||||
事件回收，继续沿program/age/整体销毁生命周期，避免误删另一个兼容pass在途使用的pipeline。
|
||||
4. **其余热路径**：缓存物理设备不变的vertex-format支持结果；退休队列由逐项erase的二次搬移
|
||||
改为一次稳定压缩。没有移除GPU同步、reply等待、present credit或角色守卫。
|
||||
|
||||
修复后独立post-window profile中，framebuffer创建采样份额降到3.56%；旧pipeline创建/
|
||||
`inflate`热点不再出现在0.1%阈值以上的服务端采样项。仍有每帧必要对象工作，未声称零创建。
|
||||
服务端样本18,017→13,106，但两份profile处于动态频率、不同运行时刻，不能读成精确同频周期倍率。
|
||||
计时窗口与profile分开，profile开销没有混入上表的FPS。
|
||||
|
||||
## 验证
|
||||
|
||||
| 门 | PASS / skip / failed |
|
||||
|---|---:|
|
||||
| unit | 2304 / 10 / 0 |
|
||||
| GPU | 1119 / 296 / 0 |
|
||||
| strict split | 176 / 4 / 0 |
|
||||
| 双块 split + Magma | 246 / 6 / 0 |
|
||||
| cache专项（含Vulkan同步验证） | 7 / 0 / 0 |
|
||||
| run-ahead专项（含Vulkan同步验证） | 14 / 0 / 0 |
|
||||
| buffer专项 | 19 / 0 / 0 |
|
||||
|
||||
- 新7项公开GL场景覆盖sampler swizzle A→B→A、respec、mip/layer/view type、
|
||||
FBO A→B→A、重挂附件、sRGB切换及attachment respec。所有draw先排队，之后才统一读像素。
|
||||
- 新2项unit覆盖兼容性身份稳定、raw句柄换代与wire/native分域。临时禁止生产ComputeHash的
|
||||
wire key分支时，same-ID/different-native-RP断言实际rc1；逐字节恢复后完整构建和该case rc0。
|
||||
- cache7与run-ahead14均证明Khronos layer实际插入，零VUID/SYNC-HAZARD/validation error。
|
||||
既有caps枚举 `GL_STENCIL_INDEX8` 应用诊断仍分列，未掩盖Vulkan错误。
|
||||
- run-ahead=0负控实际等apply并失去queue lead后失败，恢复1转绿。strict两类marker棘轮为空。
|
||||
- GPU skip集合精确为旧282加新cache generic双后端14；专用cache7全部执行。无删名或额外skip。
|
||||
- pull/push构建与生成器、field/dirty self-test、4个include probes全过。
|
||||
G1 section-size/defined-symbol比较全0，`.text 10806051`、symbols增/删/变长/改名0/0/0/0；
|
||||
不是跨构建原始机器码逐字节承诺。G2 **3019==3019**；G14 **3803→3826，+23/-0**。
|
||||
- cache同步验证已加入CI，使用 `scripts/ci/magma_cache_checks.py`；入仓版本也独立跑过7/7。
|
||||
|
||||
## 实机条件与交付
|
||||
|
||||
Redmi `2f7cbe2e`，FCL `com.tungsten.fcl.mgdebug.debug`，Minecraft `26.3-rc-3`，
|
||||
同一冻结初始世界副本、固定相机、RD32/simulation12、2620×1280、VSync off/maxFps260。
|
||||
每臂独立重启、入场CPU温度低于40°C、进入世界后预热180秒，再采至少60秒。
|
||||
最终所选两臂采样窗全部风扇观测均档位2、RPM>0；原始FCL计数按实际相邻日志间隔归一化。
|
||||
|
||||
| 模式 | CPU温度 °C | 大核观测 GHz | GPU观测 MHz |
|
||||
|---|---:|---:|---:|
|
||||
| Magma-monolith | 55.4–58.9 | 1.690–1.690 | 734–900 |
|
||||
| Magma-inproc | 57.0–58.1 | 1.210–1.690 | 832–900 |
|
||||
|
||||
使用默认DVFS，没有关闭温控。初次候选inproc计时为116.77 FPS，但风扇被系统关闭，
|
||||
该次计时不纳入最终配对，只保留其post-window profile。最终另跑一次fan2 inproc，
|
||||
没有把两次结果平均或挑选最好值。新增fan变化检查会重新预热，采样窗若漂移则拒绝该次结果。
|
||||
|
||||
两组actual backend/transport/capability、全部stats `rsp=0`、PID与前后画面均核查；
|
||||
inproc实际ARMED、strict/role-split=1；monolith无IPC配置、vbs=0。
|
||||
|
||||
- APK：`C:/Users/geekerwan/.codex/tmp/magma-perf-fix/FCL-magma-perf-70fb6689.apk`
|
||||
- APK SHA256：`bfa352bcb3402d9a31252503bacf6fcdc21e28ffaaa929fd8c80e6f2e2781e4e`
|
||||
- 已安装库SHA256：`11f26ca1f1d6c9a3d89134c31b2348306e669b1d9ebcc2fc77c7f5def2b4f818`
|
||||
- 最终配对证据：`C:/Users/geekerwan/.codex/tmp/magma-perf-fix/final-evidence/`，
|
||||
结论文件 `magma-pair.json`；包含source selection、原始计数/日志、4张验图、真实恢复核验。
|
||||
通用四臂分析器的 `performance.json` 标2/4是因为本次只测Magma两臂，不能冒称又重测了Espryt。
|
||||
- profile：`../magma-perf-baseline-profile/` 与 `../magma-perf-candidate70fb-bench/`，
|
||||
含原始perf.data、匹配Build ID的符号缓存及报告。测试计时不使用profile期间的FPS。
|
||||
- 主机证据：`/home/swung/w7/magma-perf-logs/`，前缀 `final-gates-70fb6689`、
|
||||
`syncval-caches-70fb6689`、`syncval-runahead-70fb6689`、`negative-70fb6689`、
|
||||
`pipeline-key-red-once-70fb`、`ci-cache7-ac1f`。
|
||||
|
||||
原世界61个文件已逐项SHA256核对不变，options/mg_env/mg_transport逐字节恢复、风扇恢复0，
|
||||
只保留新APK安装。已有Gradle修改和父FCL冲突索引未动。没有将缓存优化写成P7全量完成。
|
||||
@@ -2,6 +2,9 @@
|
||||
|
||||
**已完成，2026-09-20**。最终生产与实机行为头 `194382c96a8a5f51ef23f412fdccfbd40a82bf78`。
|
||||
|
||||
后续 `70fb6689` 已修复 RD32 下的 Magma inproc 性能差距：**86.85→116.73 FPS**，
|
||||
同包 monolith **115.01 FPS**；根因、缓存生命周期与复测见 [性能修复报告](magma-rd32-performance-fix.md)。
|
||||
|
||||
后续同包的 **render distance 32 四组合性能测量**见 [rd32-fourway.md](rd32-fourway.md)。
|
||||
该轮默认动态频率下，Espryt monolith/inproc 为62.18/64.02 FPS，Magma 为116.08/86.85 FPS；
|
||||
不要将下面的功能验证窗口替代这份高负载性能记录。
|
||||
|
||||
@@ -1,5 +1,8 @@
|
||||
# Redmi RD32 四组合性能记录
|
||||
|
||||
> 本文保留 `194382c9` 的基线。后续 `70fb6689` 已修复 Magma inproc 的缓存退化，
|
||||
> 同RD32场景实测 inproc **116.73 FPS**、monolith **115.01 FPS**；见 [性能修复报告](magma-rd32-performance-fix.md)。
|
||||
|
||||
2026-09-20。用户要求把 Espryt / Magma × monolith / inproc 的 render distance 都设为32，实测一轮。
|
||||
最终每组合选入一次完整窗口；原始数据、排除项与复算方法均保留。
|
||||
|
||||
|
||||
Reference in New Issue
Block a user