#lcd grid v2 | Explore Tumblr Posts and Blogs

pokefan531 · 4 years

Text

Legacy AMD APU Llano laptop for Emulation - Part 4

Radeon HD 6520g GPU Emulation tests

Since I covered the CPU for emulation, let's test out the GPU. The Radeon HD 6520g. It is based on Terascale 2 on lower HD 6000s. It supports DX11, and OpenGL 4.5. Like many Terascale 2 cards, it doesn't natively support FP64, so it's emulated. That means advanced tessellation performance is not good, and Mesa Drivers doesn't have the emulation yet, so it's stuck at OpenGL 3.3 for now. You can force OpenGL versions on games and apps that doesn't need FP64. It's best to try OpenGL 4.3 or 4.4 from commands.

We are gonna test out emulators on graphic scaling, shaders, and effects. To not have CPU bottleneck, there are few methods. One is to run a light game that only uses software rendering on the emulator, like SNES or Genesis, and use less accurate emulator or setting. Two is to find something that has a lot of things on screen that doesn't affect the speed of the emulator from CPU bottleneck. Third is to always use overclocking to reduce or remove any CPU bottleneck. For putting effects like SSAO, Per-Pixel, or traditional anti-aliasing, it may be rarely use since it's from the first gen APU. I couldn't guess the performance for A4s since I don't have any lower APUs than A6-3420m. For A8s or new gen APUs, I can guess they are faster than the results below, in most situations. My laptop comes with single memory slot, with DDR3, and unlikely be a bottleneck.

Shaders: I am testing out shaders and loading Sonic 3 and Knuckles. We're using 320x224 game to have enough pixels for low res games. We are upscaling the screen at 3x since I'm using a 768p monitor with Integer scaling, which many shaders looks better with. Some shaders will have a + on the fps for having a dynamic that goes faster on more simple details and slower on many details on screen. I let the shaders play out, and find the lowest fps. Tested on Retroarcch without vsync, audio sync, and put on Threaded Video.

720p or 768p: None: 400.0+ Bicubic: 330.0+ xBR-lv2: 174.0 xBR-lv3: 160.0 xBR-lv3-multipass: 268.0 xBR-mlv4: 158.0 Super-xbr-2P: 158.0 Super-xbr-3P: 126.0 xbrz-freescale: 75.0+ xbrx-freescale-multipass: 170.0 4x-xbrz: 50.0+ fxaa: 10.0 (set at 1x: 47.0+) smaa: 170.0+ ddt: 290.0+ crt-geom: 167.0 crt-royale: 64.0 mame-ntsc: 54.0 gtu: 100.0 lcd-grid-v2: 200.0 pixellate: 290.0+ motionblur: 174.0 nedi: 81.0 nnedi3-2x-luma: 70.0 nnedi3-2x-rgb: 35.0 omniscale-legacy: 24.0 scalefx: 193.0 scalefx-hybrid: 157.0 scalefx-rAA: 63.0 4xScaleHQ: 155.0 hq4x: 234.0 adaptive-sharpen: 11.0 (set at 1x: 85.0)

At HD Ready resolutions, a lot of the shaders play well from 320x224 at 3x. Scaling shaders like Bicubic runs well. xBR series runs pretty good, and multipass is preferred for faster results. Fixed xBRz is not at fullspeed. FXAA should be set to 1x since it's only postprocessing an image and not scaling higher. Somehow, it's still slower than SMAA. Could be how the shader is ported or something? Many CRT shaders works pretty great. Even crt-royale standard works over 60fps, but some crt shaders may not look fine at 3x or 720p resolutions. Hyllian, geom, or cgwg looks fine. Mame-crt one is the most demanding. LCD shaders plays well, and same with sharp upscales at non-integer like Pixellate runs fast. NNEDI3 upscale works fullspeed with only luma, but don't use anything above like RGB, higher neurons, or 4x. Omniscale are pretty slow, and legacy was less than half. ScaleFX worls pretty well. The hybrid and rAA looks similar, but hybrid is preferred for prerendered game for performance. Adaptive sharpen should be set at 1x. Not sure why few shader presets above doesn't use 1x for pure postprocess shaders that has nothing for upscaling. Overall, many shaders runs pretty good for 720p or 768p. It really looks good.

1080p: Bicubic: 179.0 xBR-lv2: 85.0 xBR-lv3: 78.0 xBR-lv3-multipass: 147.0 xBR-mlv4: 95.0 super-xbr-3p: 79.0 super-xbr-6p: 36.0 xbrz-freescale: 41.0+ xbrz-freescale-multipass: 92.0 ddt: 221.0+ crt-royale: 37.0 crt-royale-fake-bloom-intel: 48.0 gtu: 54.0 lcd-grid-v2: 100.0 pixellate: 199.0 nedi: 57.0 nnedi3-2x-luma: 52.0 scalefx: 137.0 scalefx-hybrid: 115.0 4xscaleHQ: 115.0 hq4x: 162.0 jinc2-sharper: 125.0

At 1080p, many of the shaders holds up well from 720p or 3x to 1080p or 4x. xBR family plays at fullspeed, especially the multipass presets. Super-Xbr ones works great at 2x with 3 passes. Trying to do 4x with six pass preset is almost down to half fullspeed. crt-royale is the chore for great crt shader. I tried the one with fake bloom and for intel windows driver, but it's still below fullspeed, and royale is mostly made for 4k. Gtu is also below, but the rest of the crt shaders should perform fine, and some that don't look as good at 720p should look better at 1080p. Nedi and NNEDI3 also loses their speed for reaching 60fps. Scalefx performs pretty good. Default scalefx scales at 3x. Generally, many shaders above still performs at fullspeed. For fast forwarding, I would choose faster alternative, like the multipass presets of a targeted shader for more speed. It's amazing how the 6520g performs good with shaders. It was a bit lower when I used to use glsl before Retroarch has slang and DX11 and glcore video driver. It's a big improvement from gl driver with glsl shaders and cg shaders, to glcore (linux) and DX11 (Windows) with slang shaders.

Since I want to compare my Tegra K1 tablet, I'll force 5x on 1080p display, and with 5x, it will crop parts of the top and bottom, almost to how most CRTs crops the image.

5x at 1080P (overscan): None: 237.0 Bicubic: 151.0 xBR-lv2: 62.0 xBR-lv3: 57.0 xBR-lv3-multipass: 119.0 xBR-mlv4: 76.0 super-xbr-3p: 65.0 xbrz-freescale-multipass: 69.0 ddt: 210.0+ crt-geom: 59.0 lcd-grid-v2: 71.0 pixellate: 173.0 scalefx: 137.0 scalefx-hybrid: 115.0 4xscaleHQ: 114.0 hq4x: 162.0 jinc2-sharper: 96.0

First of all, take a look at few framemeister videos about 5x and scanlines and integers like Phonedork. He explains about 5x integer for 1080p. My Tablet is 1200p, which shows full 5x. Anyway, Bicubic and Jinc runs pretty good. xBR multipass and xbrz has the sharpest antialiasing on the filter, and multipass versions are faster. I do recommend using lv3-multipass. Super-xbr-3p is above 60fps, but other filters are preferred. DDT performs very fast and can look a bit better than bilinear for upscaling. Crt-geom is barely, so I do suggest cgwg or hyllian for CRT shaders. LCD from 240p to 5x still performs good. Presets from interpolation folder like pixellate is still very efficient. Although AANN is very heavy, but that's for non-integer nearest scaling to see pixels. scalefx and hq4x didn't change from 4x, but 9x scalefx is too heavy. I'm surprise a lot of those shaders performs fine at 5x upscale. Although, 6520g has no support for 4k, and 1440p may be a tight minimum for 512MB of VRAM.

Let's test 480p image for measurements. Likely in Retroarch, you will play 240p games, but certain PS1, Saturn, N64, and many Dreamcast games have 640x480. A generic 480p image is tested for GPU measurements.

480P to 1080P: None: 189.0 Super-xbr-2p: 47.0 (without jinc: 62.0) Super-xbr-3p: 37.0 (without jinc: 45.0) ddt: 180.0+ nedi: 22.5 (without jinc: 25.0) nnedi3-2x-luma: 19.3 (without jinc: 21.0) reverse-aa: 126.0 xbr-lv3-multipass: 90.5 jinc2-sharper: 94.0

I test less shaders since it's more demanding for 480p image to 1080p. I only tested a simple image for shaders. Super-xbr shaders seems to perform all right. By default, it will perform lower than 60fps from 480p. I removed jinc from the last pass and let bilinear scale from 960p to 1080p, and I do get a barebones on 2 pass one. The three pass still is under 60fps. This results can vary on Super-xbr on media players, such as MadVR or MPV. MadVR uses 2P, and Super-Xbr hasn't been looked for MPV, but outperformed by Ravu. I would love to test Ravu on Retroarch, but it doesn't have it. Super-xbr performs more like default 2p one on media players, even without jinc. Back to shaders, DDT performs pretty high, and it does make some edge lines more smooth. NEDI and NNEDI3 are below 30FPS, and they aren't useful for media players based on my tests. Reverse-AA performs pretty good, and it can make digitalized 480p games look good. However, pixelated areas will stay pixellated, so you will get mixed look on reverse-aa, but can look decent. xBR with multipass will help out 480p images that has little to no antialiasing that are prerendered, and it can look pretty great on early 480p games. Letting xBR upscale from 480p to 1080p is not integer scaling, so use 2x on xBR and let bilinear scale it. jinc2 shaders performs fast enough that it delivers high quality scaling than bilinear, bicubic, and lanczos. Overall, the shaders above can make 480p games look great, depending on the content. Likely, you would use super-xbr, nnedi3, and jinc to upscale 480p content, but as for the 6520g, go with jinc2 or reverse-aa.

Since we covered the shaders for Retroarch, let's go to internal resolution on emulators.

PSX: PCSXR-PGXP will perform really good on 1080p screen, especially using PGXP will make the polygons look even better. Shaders on Pete's OpenGL2 2.9 and Tweak works very well. Although when installing tweak version, make sure you find the ini file on ini folder of PCSXR directory, and change scaling from 8x to 4x. 8x heavily tanks the performance due to 512MB VRAM limit, but 4x is very fine on any situation.

N64: Using GlideN64 4.0 (Linux) at 1080p with default setting, except using 3-point filtering, depth buffer to 0 for CPU performance, no MSAA or FXAA, and testing the very far and wide view of icy mountains of hailfire peaks from Banjo Tooie. By using the settings, it goes to 36fps (22) on the screen, and I know it should look less demanding, but with framebuffer effects being more accurate, it involves GPU limits. Turning on Legacy Blending and set frame buffer mode to 1 (VI origin) goes up to 47fps (28), and it's still not fullspeed. 3-point filtering has no effect. However, lowering the resolution to 3x native resolution to 720p, or playing it on my internal 768p monitor, it gets me around 54fps, even with the said setting changes on the scene. For 480p or native resolution that you find in Angrylion, it goes to 60fps, as long as you at least have FBE to VI origin. Don't use MSAA, because it is much more demanding than just going few times the native resolution. It should play fine overall. Many games should run fine at 1080p, but demanding parts on late big games can get tricky. Go down to 3x or 2x if you have to. Using Project64 with Jabo's D3D8 will not have any GPU bottleneck due to its age. Glide64 is the same thing, if you can ignore the CPU bottleneck caused by Glide wrappers for Glide64.

Saturn: I have test a little bit of Yaba Sanshiro, but the APU's tesellation performance is not really great, so I'd use perspetive correction. At 1x and 2x, they perform identical. At 3x (720p), I do see it perform a little bit slower, so it will be more slower at 4x or higher. RBG is set to default for performance reasons. GPU Tesellation usually performs the fastest, but when you have a GPU that doesn't have native fp64 or uses two fp32 to emulate fp64, it would either perform slower than cpu tesellation or won't show any polygons, or quads on screen.

Dreamcast: Old deprecated NullDC can run at higher resolutions just fine. Reicast's fork, Flycast, can run at higher resolutions fine, but don't use Per-Pixel since it's not good for 6520g. If the game you're playing runs at fullspeed on the APU, you can play at higher resolution on the standalone Flycast. Redream only allows higher resolution on paid versions, but it can run fine. I haven't encounter any GPU bottleneck for Dreamcast yet.

PS2: PCSX2 can have GPU go high, even at 1x. Often, the CPU is the main bottleneck that even native res or play with hack presets couldn't show the GPU bottleneck as often. However, whenever the CPU isn't demanding for a certain game that is pretty light, the GPU becomes the bottleneck, and even cases on some 2D elements. If the game is pure 2D, use software rendering if it helps more. I tried lowering the settings of GS2X settings, and no improvements.

Gamecube/Wii: Haven't tested much on the laptop, but I'd advise to use 1x resolution, and don't use ubershaders. It's not enough for 6520g, so use skip drawings to avoid shader compilation stutter. Although, for light games such as Animal Crossing and Sonic Gem Collections, you can use higher resolution.

PSP: PPSSPP allows internal resolution increase. On God of War, not it's only CPU demanding, but it's also GPU demanding too on my laptop. GOW's titlescreen is slow when you are on higher resolution, even at 2x. Frameskipping may sound ideal, but it can produce glitches and input lag. You can't use frameskip for general use. GOW runs fine at 1x, but we're left with CPU, with overclocking, and default settings on graphics, to encounter small slowdown with fighting several enemies onscreen at the beginning of the game. For the rest of the PSP games, they can run fine at 2x. 2x runs fullspeed in most games. 3x can run fine on some games, but other games can go below fullspeed. 4x is mostly slower and shouldn't be used. 2x is the safe resolution for many PSP games to play. As for Shaders, the last three on the list are the only demanding ones I found for 6520g. 5xBR-lv2 plays around 41fps, Video Smoothing is around 21fps, and Supersampling AA one is really low to around 11fps. The rest plays fine at 1080p. It's only tested for native resolution, so higher resolution can changed a bit for other shaders, specifically scaling ones like 5xBR and 4xHQ.

NDS: As mentioned on CPU testing on A6-3420m, Software is faster than OpenGL, even using native resolution, and using software graphics to high settings. On any Desmume builds, using at least 2x will become slower. It's not worth testing Desmume for GPU performance since it all relates to CPU bottleneck. On MelonDS, it does a good job increasing performance when using OpenGL renderer, but as for increasing resolution, I haven't gone through with it yet. The emulator is still bottlenecked by CPU performance. However, you almost get same performance when you increase internal resolution when playing 3D games, and not become slower that Desmume has as a problem.

Citra: Since it is CPU demanding on the laptop, it won't reach GPU bottleneck until you choose 2x or 3x, but mostly 2x you find the bottleneck, even using hardware shaders and two accurate options turned off.

Overall, 6520g is a decent iGPU for retrogaming, and very good for retroarch shaders. The benchmark is made for test purposes and to show how the 1st gen APU perform in graphics. I didn't test old windows games since I don't have that much of them installed on the laptop.

PCSXR-PGXP: 4x IR + PGXP Jabo's D3D8 (Windows): Any Resolution Glide64 (Deprecated): Any Resolution GlideN64 (Linux): 3x IR + 3-point Bilinear + Legacy Blending Yaba Sanshiro: 2x Dreamcast: 1080p (No Per-Pixel) Dolphin: 1x (Except very light games), async skip PPSSPP: 1x (GoW), 2x

As a bonus, I'll give you the average of few windows games. Shown on both native Windows and Wine with Gallium Nine DX9. Crysis GPU Demo: 38-40 Average (Low 800x600) Fortnite: 20-30 Average (Very Low 360p)*

*Although partially playable, 3.5GB of RAM is not enough for Fortnite to get rid of constant lag spikes.

Another bonus is for players like MPC-HC+MadVR or MPV can use shaders to make video image look better. I'll be testing MPV since I have Linux on the laptop as of this writing. I don't know when I'll be testing Windows to use MadVR. I feel like getting it to work on wine is challenging.

MPV Upscale from 640x360p to 1080p: Bilinear: 96.3 Lanczos: 71.8 Spline32: 71.8 Spline64: 63.9 Ewa_Robidoux: 14.3 Ewa_Lanczos: 8.0 Ravu-lite-r2: 52.0 Ravu-lite-r3: 52.2 Ravu-lite-r4: 50.0 Ravu-r2: 45.9 Ravu-r3: 45.0 Ravu-r4: 41.2 nnedi3-nns16-win8x4: 36.0 FSRCNN_x2_r2-8-0-2: 26.5 FSRCNNX_x2-8-0-4-1: 28.5

Tested common scalers and image double shaders. All tests are using gather .hook files and use bilinear scaling on luma and chroma be default. Gather shaders performs better than compute on the laptop. (Note: you must force OpenGL and GLSL version to 4.3 at least for gather to work.) No temporal scaling (motion blend) is used. Used FPS check from Gallium HUD, since MPV bar length or display stats can lower fps by few frames lower, however it doesn't happen on D3D11. Surprised that bilinear is over 90fps. Bilinear, Lanczos, and Spline performs pretty good. EWA based scalers performs the worst, and gets very low fps. On Ravu, it performs below 60fps on all of them. Ravu-lite r2 and r3 performs identical. Between r2 and r4, performance is not as different. Standard Ravu performs lower, but Ravu for 360p is good enough to use r4 on videos that are at most 30fps. NNEDI3-16, surprisingly performs a bit better than what I expected. It is below Ravu-r4. FSRCNN shaders are up. They're not useful for this GPU since they're below 30fps and can take a hit on videos that are higher resolution. However, FSRCNNX-8 performs slightly better and has more quality than FSRCNN-8.

MPV Upscale from 720x480p DVD to 1080p: Bilinear: 86.3 Lanczos: 59.9 Spline32: 59.9 Spline64: 54.0 Ravu-lite-r2: 40.7 Ravu-lite-r3: 40.3 Ravu-lite-r4: 39.3 Ravu-r2: 35.5 Ravu-r3: 34.4 Ravu-r4: 31.4 nnedi3-nns16-win8x4: 26.3 FSRCNN_x2_r2-8-0-2: 19.1 FSRCNNX_x2-8-0-4-1: 20.4

For 480p to 1080p, we see more results. Lanczos and Spline36 are identical and is borderline with 60fps. Although, bilinear should be stable and not have one framedrop per few seconds. Ravu shaders are appropriate for videos that are at most 30fps. Ravu-r4 is nearly down to 30fps, but r3 seems to be safe. r4 is better suited for 360p instead. NNEDI3 performs lower, and neither can any FSRCNN shaders give enough performance for general playback. FSRCNNX performs slightly better than standard FSRCNN again by little.

720p videos won't play well with those image doublers, so you're left with bilinear for 60fps playback and lanczos or spline shaders for 30fps or lower.

Overall, the entire laptop plays really fine with up to 5th gen home consoles and up to PSP portable consoles. It’s a really decent APU to use for emulation and it can have more tweaks with Linux. As a HTPC, it plays well on things up to 1080p. Playback is fine, but GPU performance for scaling can be basic or limiting for upscaling quality. As for games, it should play DirectX 9 era games really fine and early DirectX 11 ones too.

Previous Page on CPU Emulation tests.

#emulation #laptop #apu

0 notes