GitLab

I have a trace of Age of Empires IV (https://cloud.igalia.com/s/o3ZyFEoY66FYDKj) which on frame 653 results in an unrecoverable hang (the board completely locks up and reboots after a few seconds).

I use this changes https://gitlab.freedesktop.org/Danil/mesa/-/commits/TEST/vkd3d to run it. To run the trace do:

gfxrecon-replay --remove-unsupported -m rebind --sfa age_of_empire_iv_full.gfxr --pause-frame 652

(The hang will happen on the 653 frame).

@flto found that the hang could be avoided if we don't use tiling with R8G8 images without UBWC and suggested that "probably R8G8 non-UBWC tiled is buggy, since media formats are always either linear or UBWC".

if (image->vk_format == VK_FORMAT_R8G8_UNORM && tile_mode == TILE6_3 && !ubwc_enabled)
   tile_mode = TILE6_LINEAR;

(This in tu_CreateImage would prevent the hang)

Digging deeper I found the command which hangs (I'm mostly sure of it):

t7			opcode: CP_BLIT (2c) (2 dwords)
				{ OP = BLIT_OP_SCALE }
			mode: RM6_BYPASS
			skip_ib2: g=0, l=0
			draw[2354] register values
!+	10f00f80			GRAS_2D_BLIT_CNTL: { ROTATE = ROTATE_0 | SOLID_COLOR | COLOR_FORMAT = FMT6_8_8_UNORM | MASK = 0xf | IFMT = R2D_UNORM8 | RASTER_MODE = TYPE_TILED }
 +	00000000			GRAS_2D_DST_TL: { X = 0 | Y = 0 }
!+	04ff04ff			GRAS_2D_DST_BR: { X = 1279 | Y = 1279 }
!+	10f00f80			RB_2D_BLIT_CNTL: { ROTATE = ROTATE_0 | SOLID_COLOR | COLOR_FORMAT = FMT6_8_8_UNORM | MASK = 0xf | IFMT = R2D_UNORM8 | RASTER_MODE = TYPE_TILED }
 +	00000000			RB_2D_UNKNOWN_8C01: 0
!+	0000030f			RB_2D_DST_INFO: { COLOR_FORMAT = FMT6_8_8_UNORM | TILE_MODE = TILE6_3 | COLOR_SWAP = WZYX | SAMPLES = MSAA_ONE }
!+	2cb30280			RB_2D_DST: 0x2cb30280
 +	00000000			RB_2D_DST_HI: 0
!+	00000028			RB_2D_DST_PITCH: 2560
!+	2cb30280			RB_2D_DST_FLAGS: 0x2cb30280
 +	00000000			RB_2D_DST_FLAGS_HI: 0
 +	00000000			RB_2D_DST_FLAGS_PITCH: 0
 +	00000000			RB_2D_SRC_SOLID_C0: 0
 +	00000000			RB_2D_SRC_SOLID_C1: 0
 +	00000000			RB_2D_SRC_SOLID_C2: 0
 +	00000000			RB_2D_SRC_SOLID_C3: 0
!+	0000f078			SP_2D_DST_FORMAT: { COLOR_FORMAT = FMT6_8_8_UNORM | MASK = 0xf }

Command stream dump: age_of_empire_iv_full_near_hang_a660.tar.gz (commands were instrumented to find the exact place of the hang). The hang happens between CP_WAIT_REG_MEM REF = with values of 0x4d245 and 0x4d246.

The blit seem to be caused by the clear of a render target by an empty render pass with VK_ATTACHMENT_LOAD_OP_CLEAR.

The image is:

"[43888]vkCreateImage": {
    "return": "VK_SUCCESS",
    "device": "0x0x170",
    "pCreateInfo": {
        "sType": "VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO",
        "pNext": {
            "sType": "VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO",
            "pNext": null,
            "viewFormatCount": 4,
            "pViewFormats": [
                "VK_FORMAT_R8G8_UNORM",
                "VK_FORMAT_R8G8_UINT",
                "VK_FORMAT_R8G8_SINT",
                "VK_FORMAT_R8G8_SNORM"
            ]
        },
        "flags": 8,
        "imageType": "VK_IMAGE_TYPE_2D",
        "format": "VK_FORMAT_R8G8_UNORM",
        "extent": {
            "width": 1280,
            "height": 1280,
            "depth": 1
        },
        "mipLevels": 1,
        "arrayLayers": 1,
        "samples": "VK_SAMPLE_COUNT_1_BIT",
        "tiling": "VK_IMAGE_TILING_OPTIMAL",
        "usage": 23,
        "sharingMode": "VK_SHARING_MODE_EXCLUSIVE",
        "queueFamilyIndexCount": 0,
        "pQueueFamilyIndices": [],
        "initialLayout": "VK_IMAGE_LAYOUT_UNDEFINED"
    },
    "pAllocator": null,
    "[out]pImage": "0x0xd93"
},

Anyway, I tried to reproduce the issue by re-creating renderpass I found by dumping trace's commands, and wasn't able to reproduce anything...

We could just disable the tiling for non-ubwc R8G8 (or all two channel formats?) checking what blob does beforehand and call it a day. But I guess we want to know the exact extents of the issue.