Bug 106922 - radv: corruption issues with Tangrams demo on Polaris
Summary: radv: corruption issues with Tangrams demo on Polaris
Status: RESOLVED NOTOURBUG
Alias: None
Product: Mesa
Classification: Unclassified
Component: Drivers/Vulkan/radeon (show other bugs)
Version: git
Hardware: Other All
: medium normal
Assignee: mesa-dev
QA Contact: mesa-dev
URL:
Whiteboard:
Keywords:
Depends on:
Blocks:
 
Reported: 2018-06-14 21:05 UTC by Christoph Haag
Modified: 2018-09-21 07:08 UTC (History)
0 users

See Also:
i915 platform:
i915 features:


Attachments
corruption at small resolutions (208.63 KB, image/png)
2018-09-15 01:17 UTC, Christoph Haag
Details

Description Christoph Haag 2018-06-14 21:05:00 UTC
RX 480
mesa radv 41dabdc4753
llvm 7svn 165520

The demo is for windows only and I think closed source but it runs well with the most recent amdvlk. On radv it crashes.
You can get it here: http://www.pouet.net/prod.php?which=76464

Make a 64 bit wine prefix, run wine64 Tangrams.exe, click demo and it immediately crashes with this message:

LLVM ERROR: Cannot select: 0x7e8d8750: i16 = bitcast 0x7e8d8af8
  0x7e8d8af8: f32,ch = BUFFER_LOAD<(dereferenceable load 4 from custom TargetCustom7, align 1, addrspace 4)> 0x7d433f78, 0x7e8d7430, Constant:i32<0>, 0x7e8d7638, Constant:i1<0>, Constant:i1<0>
    0x7e8d7430: v4i32,ch = load<(dereferenceable invariant load 16 from %ir.56, addrspace 6)> 0x7d433f78, 0x7e8d72f8, undef:i32
      0x7e8d72f8: i32 = add 0x7e8ce2c8, Constant:i32<48>
        0x7e8ce2c8: i32,ch = CopyFromReg 0x7d433f78, Register:i32 %100
          0x7e8ce260: i32 = Register %100
        0x7e8d3e40: i32 = Constant<48>
      0x7e8cebb8: i32 = undef
    0x7e8ceb50: i32 = Constant<0>
    0x7e8d7638: i32 = add 0x7e8d73c8, Constant:i32<2>
      0x7e8d73c8: i32 = mul 0x7e96cb20, Constant:i32<6>
        0x7e96cb20: i32 = add 0x7e96c918, 0x7e96c9e8
          0x7e96c918: i32 = mulhs 0x7e8ce810, Constant:i32<1431655766>
            0x7e8ce810: i32 = add 0x7e8ce538, 0x7e8ce398
              0x7e8ce538: i32,ch = CopyFromReg 0x7d433f78, Register:i32 %103
                0x7e8ce4d0: i32 = Register %103
              0x7e8ce398: i32,ch = CopyFromReg 0x7d433f78, Register:i32 %101
                0x7e8ce330: i32 = Register %101
            0x7e8d7290: i32 = Constant<1431655766>
          0x7e96c9e8: i32 = srl 0x7e96c918, Constant:i32<31>
            0x7e96c918: i32 = mulhs 0x7e8ce810, Constant:i32<1431655766>
              0x7e8ce810: i32 = add 0x7e8ce538, 0x7e8ce398
                0x7e8ce538: i32,ch = CopyFromReg 0x7d433f78, Register:i32 %103
                  0x7e8ce4d0: i32 = Register %103
                0x7e8ce398: i32,ch = CopyFromReg 0x7d433f78, Register:i32 %101
                  0x7e8ce330: i32 = Register %101
              0x7e8d7290: i32 = Constant<1431655766>
            0x7e96c980: i32 = Constant<31>
        0x7e8d7360: i32 = Constant<6>
      0x7e8d75d0: i32 = Constant<2>
    0x7e8d3b68: i1 = Constant<0>
    0x7e8d3b68: i1 = Constant<0>
In function: main
Comment 1 Bas Nieuwenhuizen 2018-06-14 21:25:05 UTC
Can reproduce a LLVM error here.
Comment 2 Bas Nieuwenhuizen 2018-06-14 21:34:38 UTC
If this is any indication it may just be not checking exts:

SPIR-V WARNING:
    In file ../mesa/src/compiler/spirv/spirv_to_nir.c:3312
    Unsupported SPIR-V capability: SpvCapabilityInt16
    28 bytes into the SPIR-V binary
SPIR-V WARNING:
    In file ../mesa/src/compiler/spirv/spirv_to_nir.c:3394
    Unsupported SPIR-V capability: SpvCapabilityStorageBuffer16BitAccess
    36 bytes into the SPIR-V binary
SPIR-V WARNING:
    In file ../mesa/src/compiler/spirv/spirv_to_nir.c:3394
    Unsupported SPIR-V capability: SpvCapabilityUniformAndStorageBuffer16BitAccess
    44 bytes into the SPIR-V binary

Since we don't support these.

(or the undef is messing things up, let me find out where it gets introduced)
Comment 3 Bas Nieuwenhuizen 2018-06-14 21:46:28 UTC
So top level is a f32 -> i16 bitcast:

Cannot select: 0x7e8d8750: i16 = bitcast 0x7e8d8af8

which is not allowed. Is in LLVM source:

%76 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %75, i32 0, i32 %74, i1 false, i1 false) #3
  %77 = bitcast float %76 to i16

Happens in the nir:

vec1 16 ssa_115 = intrinsic load_ssbo (ssa_113, ssa_114) () ()

I'm really inclined to say it is our lacking 16-bit support.
Comment 4 Christoph Haag 2018-07-24 11:48:03 UTC
After seeing
https://www.phoronix.com/scan.php?page=news_item&px=RADV-VK_KHR_16bit_storage
I thought I'd try again.

I now get a much clearer error:

SPIR-V WARNING:
    In file ../src/compiler/spirv/spirv_to_nir.c:3436
    Unsupported SPIR-V capability: SpvCapabilityInt16
    28 bytes into the SPIR-V binary

The application then (unsurprisingly) crashes.
Comment 5 Samuel Pitoiset 2018-07-25 11:45:20 UTC
Missing 16-bit integer support.
Comment 6 Samuel Pitoiset 2018-09-14 12:57:47 UTC
The demo should now work if you apply:

https://patchwork.freedesktop.org/series/49701/
https://patchwork.freedesktop.org/series/49710/
Comment 7 Christoph Haag 2018-09-15 01:17:53 UTC
Created attachment 141569 [details]
corruption at small resolutions

Nice, with the patches it runs without crashing.

At 1920x1080 it looks good, but when choosing smaller resolutions, corruption appears. In the default 720x400 it's pretty bad. I have no idea if the corruption is in any way related to the 16 bit support though.

RX 480, llvm 8.0.0svn_r169421, latest radv git + patches
Comment 8 Samuel Pitoiset 2018-09-17 08:22:23 UTC
Thanks for checking. Apparently the corruption is Polaris specific. On vega it works nice, if you have the hardware can you confirm?
Comment 9 Samuel Pitoiset 2018-09-17 13:20:00 UTC
I pushed both series but still need to figure out the problem on Polaris.
Comment 10 Samuel Pitoiset 2018-09-21 07:08:07 UTC
So, this demo has several problems:

- Not enough memory allocated in a pool which ends up by crashing with RADV. It does work with AMDVLK because they always allocate more space than needed. I pushed a change for that this week, but it was actually wrong and has been reverted. So, expect a crash anyway.
- Use of 16-bit integers without checking if shaderInt16 is supported.
- Possibly a missing barrier which might explain the rendering glitches on GFX8 (syncing GFX shaders and invalidating L2 seems to fix the problem).

Tangrams is buggy in many ways and I don't think it's worth trying to investigate more.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.