Vulkan rendering backend. Early Work-In-Progress #8601

hrydgard · 2016-02-19T11:31:26Z

Please note - this is not for general use, it's only an early preview. It's very rough, only works in non-buffered (crashes in buffered) and is missing many features and optimizations, so it's not even fast yet.

EDIT: At this point, it already beats OpenGL in nearly everything I've tried, though is not feature complete of course.

But, it's a start, and this will hopefully let us take performance to the next level on Android, eventually - once drivers supporting Vulkan are released, and present on devices. Some devices will get Vulkan with a Marshmallow update (nVidia Shield/TV) while others will have to wait longer or won't get it at all.

Anuskuss · 2016-02-19T19:48:54Z

Awesome 👍

gamelaster · 2016-02-19T20:18:50Z

D3D11 backend WIP too?

hrydgard · 2016-02-19T21:23:30Z

@gamelaster D3D11? only vague plans, nothing concrete. Just made space in the enum, heh. Might remove it from the list.

VIRGINKLM · 2016-02-20T06:24:35Z

Just a small note. Samsung Galaxy A6 (2016) supports Vulkan straight out of the box and NVidia Shield devices began recieving Dev buils supporting Vulkan. Also I bet Galaxy S7 variants will have straight out of the box support for it.

RinMaru · 2016-02-20T09:05:25Z

So will this work on ogl backend in the future and can this hold potential for better performance on damanding games on hardware that couldnt run at full speed previously?

zminhquanz · 2016-02-20T13:50:54Z

https://en.wikipedia.org/wiki/Vulkan_(API)
Here is the list for the GPU from NVIDIA , AMD and PowerVR , Mali , Adreno , Intel to support Vulkan API
Device can be update to Android 6.0 to support it

unknownbrackets · 2016-02-21T05:18:56Z

Common/GraphicsContext.h

@@ -22,6 +22,9 @@ class GraphicsContext {

 	virtual void Resize() = 0;

+	// Needs casting to the appropriate type, unfortunately. Should find a better solution..
+	virtual void *GetAPIContext() { return nullptr; }


Seems like this is no longer used?

-[Unknown]

unknownbrackets · 2016-02-27T20:24:12Z

Where is the graphicsContext supposed to be populated? I had to add PSP_CoreParameter().graphicsContext = graphicsContext; to NativeRender() to fix some crashes.

Well, now I'm crashing in vkCmdClearAttachments, but it's progress.

-[Unknown]

hrydgard · 2016-02-27T23:12:27Z

@unknownbrackets Remember to turn off buffered rendering, it will crash :)

Don't recall having problems with graphicsContext... hm

hrydgard · 2016-02-27T23:36:33Z

@unknownbrackets turns out I broke depth buffer init, too, when messing with the shield

unknownbrackets · 2016-03-06T19:34:31Z

Windows/PPSSPP.sln

 		Debug|Win32 = Debug|Win32
 		Debug|x64 = Debug|x64
+		Release CRT DLL|Win32 = Release CRT DLL|Win32
+		Release CRT DLL|x64 = Release CRT DLL|x64


Do we want to add these, or were they just added by something else?

-[Unknown]

Hm, they were probably added when I tried to import the CMake-generated project for glslang, before I made my own instead.

unknownbrackets · 2016-03-06T22:55:46Z

unittest/UnitTests.vcxproj

@@ -123,7 +122,7 @@
    <Link>
      <SubSystem>Console</SubSystem>
      <GenerateDebugInformation>true</GenerateDebugInformation>
-      <AdditionalDependencies>Ws2_32.lib;winmm.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;opengl32.lib;dsound.lib;glu32.lib;avcodec.lib;avformat.lib;avutil.lib;swresample.lib;swscale.lib;comctl32.lib;d3d9.lib;dxguid.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>..\Windows\x64\Debug\glslang.lib;..\ext\vulkan\vulkan-1.lib;Ws2_32.lib;winmm.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;opengl32.lib;dsound.lib;glu32.lib;avcodec.lib;avformat.lib;avutil.lib;swresample.lib;swscale.lib;comctl32.lib;d3d9.lib;dxguid.lib;%(AdditionalDependencies)</AdditionalDependencies>


Any reason only this one gets the glslang lib?

-[Unknown]

It's not supposed to be there, it's linked through "References". this was an old thing that slipped through some cleanup.

unknownbrackets · 2016-03-12T17:53:53Z

@hrydgard

Refactor away nextTexture_. Not needed on Vulkan.

Do you mean because you plan to remove the hardware transform -> software fallback and move vertex decoding earlier so that it can be done before SetTexture()? Because that is the reason nextTexture_ exists. We don't really need it for D3D9 either, technically, I suppose.

To clarify, ApplyTexture() MUST be called AFTER the vertices have been decoded. I'm not sure, but right now this isn't happening, and this could be contributing to rendering bugs. At the very least, it's ignoring the output of the vertex decode (texture bounds.)

-[Unknown]

hrydgard · 2016-03-12T19:54:03Z

Yeah, I intend to move things around a bit, I've just been swamped. And yes, UV scale/offset seems a bit broken, among other things, as expected...

unknownbrackets · 2016-03-13T17:01:31Z

Core/Config.h

@@ -50,10 +50,14 @@ enum BufferFilter {
 enum class GPUBackend {
 	OPENGL = 0,
 	DIRECT3D9 = 1,
+  DIRECT3D11 = 2,
+  VULKAN = 3,


Indent nitpick.

-[Unknown]

unknownbrackets · 2016-03-13T20:06:15Z

Personally, I would prefer this convention:

  VkFramebufferCreateInfo fb_info = { VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO };
  fb_info.renderPass = surface_render_pass_;

Instead of:

  VkFramebufferCreateInfo fb_info = {};
  fb_info.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO;
  fb_info.pNext = NULL;
  fb_info.renderPass = surface_render_pass_;

What do you think?

-[Unknown]

hrydgard · 2016-03-13T20:20:49Z

Hm, not bad. I guess the compiler will eliminate the extra stores (first zeroing the struct, then filling it out) anyway. Let's do that.

unknownbrackets · 2016-03-16T04:14:39Z

Do you think it would be a good idea to decode directly onto the push buffer (from the jit), and then if we're caching, issue a cmd to copy that to the cache buffer (which would then be used for subsequent draws?)

By the way, I'm seeing something funny with vertex coords in FF4, seems to be why only the first letter is drawing in some places. Still debugging, but this area has me thinking about vertex decode is all.

-[Unknown]

unknownbrackets · 2016-03-16T04:33:49Z

GPU/Vulkan/ShaderManagerVulkan.cpp

@@ -162,7 +162,7 @@ static void ConvertProjMatrixToVulkan(Matrix4x4 &in, bool invertedX, bool invert
 	if (invertedY)
 		yoff = -yoff;

-	const Vec3 trans(xoff, yoff, gstate_c.vpZOffset + 0.5f);
+	const Vec3 trans(xoff, yoff, gstate_c.vpZOffset * 0.5f + 0.5f);


Wait wait wait, does Vulkan use the half-pixel offset insanity that Direct3D9 does? It doesn't right?

-[Unknown]

… flexible.

Optimizing this number (by avoiding redundant UBO uploads etc) will probably help performance.

We have to clear both at the same time. I think it makes sense to consider this part of the "separate alpha" flag, since alpha has to match both color and depth.

It's definitely not a float, oops.

Otherwise we get wrong colors - also we are still converting colors.

Some places were not handling it as a power of two.

…g directly into the pushbuffers easier, plus other benefits later. (For example, we'll often be able to avoid rebinding the vertex and index buffers at new offsets by just keep counting upwards between draws, if the vertex format is the same but other state changed)

…ic growth

Turns out it's only OK to leave out if rendering only to color, otherwise behaviour is undefined - which means it may still be enabled.

…ing optimization.

… when possible.

…ot waste space (nVidia needs 256-byte alignment)

Vulkan rendering backend. Early Work-In-Progress

unknownbrackets reviewed Feb 21, 2016
View reviewed changes

hrydgard force-pushed the vulkan branch from 66c0a83 to 71efcf3 Compare February 21, 2016 18:03

hrydgard force-pushed the vulkan branch from 223d10c to 59a231a Compare February 28, 2016 12:29

unknownbrackets reviewed Mar 6, 2016
View reviewed changes

unknownbrackets mentioned this pull request Mar 6, 2016

Bring over non-Vulkan changes from Vulkan branch #8627

Merged

hrydgard force-pushed the vulkan branch from 59a231a to 6531ace Compare March 6, 2016 22:45

unknownbrackets reviewed Mar 6, 2016
View reviewed changes

unknownbrackets reviewed Mar 13, 2016
View reviewed changes

hrydgard force-pushed the vulkan branch from 04f8896 to 921ea64 Compare March 13, 2016 17:45

unknownbrackets reviewed Mar 16, 2016
View reviewed changes

hrydgard force-pushed the vulkan branch from 8f66a0b to cfa4df1 Compare March 17, 2016 22:37

hrydgard and others added 23 commits March 20, 2016 19:31

Upload PSP textures through a push buffer instead of image copy. More…

8a62724

… flexible.

Vulkan: Enable mipmapping

84e649f

Show how much pushbuffer space is used each frame.

e45c24b

Optimizing this number (by avoiding redundant UBO uploads etc) will probably help performance.

Get rid of strange offset in Vulkan matrix converter

a995dd2

Vulkan: Make sure depth==stencil when clearing.

f289e84

We have to clear both at the same time. I think it makes sense to consider this part of the "separate alpha" flag, since alpha has to match both color and depth.

Vulkan: Properly set the alpha test ref.

dad64b3

It's definitely not a float, oops.

Vulkan: Simplify color testing.

baa7132

Correct the 5551 format for now.

3744008

Otherwise we get wrong colors - also we are still converting colors.

Split out VulkanTexture from VulkanContext.cpp/h into VulkanImage.cpp/h

da50370

Vulkan: Initial support for aniso filtering.

82f3df1

Cleanup differences in aniso handling.

add506a

Some places were not handling it as a power of two.

Vulkan: Fix our use of dual source blending, re-enable if available.

966cb89

Move VulkanPushBuffer into VulkanMemory.h

5d56537

Change the VulkanPushBuffer API to allow for adding support for dynam…

2c61b1e

…ic growth

Always specify depthstencil state in pipelines.

8e9acff

Turns out it's only OK to leave out if rendering only to color, otherwise behaviour is undefined - which means it may still be enabled.

Fix a resource leak. Request COHERENT memory for pushbuffers.

5d45413

Minor optimizations

18d78f9

No need to align vertex/index data. Will only be harmful for an upcom…

d0659f0

…ing optimization.

Vulkan: Remove support for software skinning to reduce complexity

d7ffc39

Decode vertex data directly into the vertex pushbuffer, saving a memcpy.

5d19f3d

Be more economical with UBO pushbuffer space by reusing the last data…

0b1cfaf

… when possible.

SoftGPU with Vulkan runs but displays black

29bc07e

hrydgard force-pushed the vulkan branch from ff3d23a to 29bc07e Compare March 20, 2016 19:05

hrydgard added 3 commits March 20, 2016 20:26

Vulkan: Add ugly temporary warning about buffered rendering not working

c33c3cf

Combined two uniforms to get the base UBO down to 512b, in order to n…

827481d

…ot waste space (nVidia needs 256-byte alignment)

Buildfix to last commit, add a comment

0e19927

hrydgard added a commit that referenced this pull request Mar 20, 2016

Merge pull request #8601 from hrydgard/vulkan

2b46ae5

Vulkan rendering backend. Early Work-In-Progress

hrydgard merged commit 2b46ae5 into master Mar 20, 2016

hrydgard deleted the vulkan branch March 20, 2016 20:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Vulkan rendering backend. Early Work-In-Progress #8601

Vulkan rendering backend. Early Work-In-Progress #8601

hrydgard commented Feb 19, 2016

Anuskuss commented Feb 19, 2016

gamelaster commented Feb 19, 2016

hrydgard commented Feb 19, 2016

VIRGINKLM commented Feb 20, 2016

RinMaru commented Feb 20, 2016

zminhquanz commented Feb 20, 2016

unknownbrackets Feb 21, 2016

unknownbrackets commented Feb 27, 2016

hrydgard commented Feb 27, 2016

hrydgard commented Feb 27, 2016

unknownbrackets Mar 6, 2016

hrydgard Mar 6, 2016

unknownbrackets Mar 6, 2016

hrydgard Mar 6, 2016

unknownbrackets commented Mar 12, 2016

hrydgard commented Mar 12, 2016

unknownbrackets Mar 13, 2016

unknownbrackets commented Mar 13, 2016

hrydgard commented Mar 13, 2016

unknownbrackets commented Mar 16, 2016

unknownbrackets Mar 16, 2016

Vulkan rendering backend. Early Work-In-Progress #8601

Vulkan rendering backend. Early Work-In-Progress #8601

Conversation

hrydgard commented Feb 19, 2016

Anuskuss commented Feb 19, 2016

gamelaster commented Feb 19, 2016

hrydgard commented Feb 19, 2016

VIRGINKLM commented Feb 20, 2016

RinMaru commented Feb 20, 2016

zminhquanz commented Feb 20, 2016

unknownbrackets Feb 21, 2016

Choose a reason for hiding this comment

unknownbrackets commented Feb 27, 2016

hrydgard commented Feb 27, 2016

hrydgard commented Feb 27, 2016

unknownbrackets Mar 6, 2016

Choose a reason for hiding this comment

hrydgard Mar 6, 2016

Choose a reason for hiding this comment

unknownbrackets Mar 6, 2016

Choose a reason for hiding this comment

hrydgard Mar 6, 2016

Choose a reason for hiding this comment

unknownbrackets commented Mar 12, 2016

hrydgard commented Mar 12, 2016

unknownbrackets Mar 13, 2016

Choose a reason for hiding this comment

unknownbrackets commented Mar 13, 2016

hrydgard commented Mar 13, 2016

unknownbrackets commented Mar 16, 2016

unknownbrackets Mar 16, 2016

Choose a reason for hiding this comment