starred/shadPS4

Fork 0

mirror of https://github.com/shadps4-emu/shadPS4.git synced 2026-04-26 00:05:58 +03:00

[PR #2335] [MERGED] Reduce redundant calls to sceKernelIsNeoMode, improve BB frame times #2658

New issue

Closed

opened 2026-02-27 22:00:45 +03:00 by kerem · 0 comments

kerem commented

2026-02-27 22:00:45 +03:00

Owner

📋 Pull Request Information

Original PR: https://github.com/shadps4-emu/shadPS4/pull/2335
Author: @hspir404
Created: 2/3/2025
Status: ✅ Merged
Merged: 2/3/2025
Merged by: @squidbus

Base: main ← Head: reduce-is-neo-mode-calls

📝 Commits (1)

0cb888d Remove log line that was consuming as much as 0.6ms frame time

📊 Changes

1 file changed (+0 additions, -1 deletions)

View changed files

📝 src/core/libraries/kernel/process.cpp (+0 -1)

📄 Description

I found this call to sceKernelIsNeoMode while poking around in the profiler. I saw that calls are very frequent (especially inside the conversion of texture descriptors to the Vulkan renderer). They are taking up some 750+ ms out of a 10 second sample. I sampled in several different places, and these calls were often only taking up as much as 400ms, depending on the scene being rendered, but still a not small chunk.

This change makes the code very slightly less appealing, but it's reducing 6040 calls to 65 (at least, those are the ones the sample-based profiler caught in a 10 second sample). Specifically, that sample was taken at the top of the DLC Nightmare Catherdral, notably the worst performing part of the game I've been able to find.

Before:

After:

I made an automation build here and tested both EXEs side by side in several parts of the game, and have screenshots from those tests. I loaded the emulator and loaded the save game, straight into these areas, and did not move the camera. I let the frame rate settle before taking screenshots, and attempted to capture the screenshot that got the highest frame rate I was seeing. If I missed the screenshot, I let it settle a bit again before taking another one.

Test Configuration -

Game Mods:

Vertex explosions fix (modifies base game files)
Skip intro
Disable motion blur
Disable chromatic aberration
Disable HTTP calls
60fps patch
vsync off
vblank divider 2

Box:

Windows 10
64GB RAM
nVidia RTX 3080 10GB VRAM
AMD 5950x (16-core processor)

In almost every scene there is a slight performance improvement.

Some other folks in chat also did some A/B testing, took some screenshots, and were getting a bit of a performance increase.

Before:

After:

Before:

After:

Before:

After:

Before:

After

Before:

After:

Before:

After:

Before:

After:

Before:

After:

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/shadps4-emu/shadPS4/pull/2335 **Author:** [@hspir404](https://github.com/hspir404) **Created:** 2/3/2025 **Status:** ✅ Merged **Merged:** 2/3/2025 **Merged by:** [@squidbus](https://github.com/squidbus) **Base:** `main` ← **Head:** `reduce-is-neo-mode-calls` --- ### 📝 Commits (1) - [`0cb888d`](https://github.com/shadps4-emu/shadPS4/commit/0cb888d01323c366d43abe25510934b78732704f) Remove log line that was consuming as much as 0.6ms frame time ### 📊 Changes **1 file changed** (+0 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `src/core/libraries/kernel/process.cpp` (+0 -1) </details> ### 📄 Description I found this call to `sceKernelIsNeoMode` while poking around in the profiler. I saw that calls are very frequent (especially inside the conversion of texture descriptors to the Vulkan renderer). They are taking up some 750+ ms out of a 10 second sample. I sampled in several different places, and these calls were often only taking up as much as 400ms, depending on the scene being rendered, but still a not small chunk. This change makes the code very slightly less appealing, but it's reducing 6040 calls to 65 (at least, those are the ones the sample-based profiler caught in a 10 second sample). Specifically, that sample was taken at the top of the DLC Nightmare Catherdral, notably the worst performing part of the game I've been able to find. Before: ![image](https://github.com/user-attachments/assets/ed42988a-8f70-40e9-992d-3180c6613fa0) After: ![image](https://github.com/user-attachments/assets/7e37a6dc-c037-4734-9cc9-177ac01e9aaa) I made an automation build here and tested both EXEs side by side in several parts of the game, and have screenshots from those tests. I loaded the emulator and loaded the save game, straight into these areas, and did not move the camera. I let the frame rate settle before taking screenshots, and attempted to capture the screenshot that got the highest frame rate I was seeing. If I missed the screenshot, I let it settle a bit again before taking another one. Test Configuration - Game Mods: - Vertex explosions fix (modifies base game files) - Skip intro - Disable motion blur - Disable chromatic aberration - Disable HTTP calls - 60fps patch - vsync off - vblank divider 2 Box: - Windows 10 - 64GB RAM - nVidia RTX 3080 10GB VRAM - AMD 5950x (16-core processor) In almost every scene there is a slight performance improvement. Some other folks in chat also did some A/B testing, took some screenshots, and were getting a bit of a performance increase. Before: ![1-before](https://github.com/user-attachments/assets/ec1b9a9f-1aec-420f-a4b9-b3875b61db5c) After: ![1-after](https://github.com/user-attachments/assets/e15773d7-08f2-4530-948d-f1ec7309c677) Before: ![2-before](https://github.com/user-attachments/assets/58c91f14-0b6b-4d49-ba6d-8d7bddc7049f) After: ![2-after](https://github.com/user-attachments/assets/64e7994f-7f70-450b-9c47-a9dbadcd2e09) Before: ![3-before](https://github.com/user-attachments/assets/c1950a3b-8b59-4640-94a2-501cbed93211) After: ![3-after](https://github.com/user-attachments/assets/934f9b68-2216-481a-9ecc-6e539f41ceb9) Before: ![4-before](https://github.com/user-attachments/assets/21bee82a-308c-4689-8589-28bc422b26af) After ![4-after](https://github.com/user-attachments/assets/e6749079-7527-4234-a648-786c8c25bab7) Before: ![5-before](https://github.com/user-attachments/assets/53683c79-0e46-41d6-9cb3-a2d30503c0aa) After: ![5-after](https://github.com/user-attachments/assets/894db2bd-0640-419e-b7c4-7259e08bbd12) Before: ![6-before](https://github.com/user-attachments/assets/27fa18cf-5bf6-4c36-8e79-0b5884a4761b) After: ![6-after](https://github.com/user-attachments/assets/9b9c8c33-2395-49dd-bff1-ac461793d7bc) Before: ![7-before](https://github.com/user-attachments/assets/c7a596f5-ffe1-4f3e-a412-49017dd92454) After: ![7-after](https://github.com/user-attachments/assets/d8553e98-8c83-4dd0-8a76-be5f6a9d7f08) Before: ![8-before](https://github.com/user-attachments/assets/7e47c58d-8c61-449a-bb2e-b9e8fd74555b) After: ![8-after](https://github.com/user-attachments/assets/335aa804-2e39-479e-b181-88b7b6f22751) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>