mirror of
https://github.com/par274/sharpemu.git
synced 2026-07-22 19:06:15 +08:00
[shader_recompiler] Fix guest image byte count calculation for Vulkan video presenter (#395)
This commit is contained in:
Berk
@@ -1667,6 +1667,32 @@ internal static unsafe class VulkanVideoPresenter
|
||||
}
|
||||
}
|
||||
|
||||
internal static ulong GetGuestImageByteCount(uint format, uint width, uint height)
|
||||
{
|
||||
var blockBytes = format switch
|
||||
{
|
||||
169 or 170 or 175 or 176 => 8UL,
|
||||
171 or 172 or 173 or 174 or
|
||||
177 or 178 or 179 or 180 or 181 or 182 => 16UL,
|
||||
_ => 0UL,
|
||||
};
|
||||
if (blockBytes != 0)
|
||||
{
|
||||
return checked(((ulong)width + 3) / 4 * (((ulong)height + 3) / 4) * blockBytes);
|
||||
}
|
||||
|
||||
var bytesPerPixel = format switch
|
||||
{
|
||||
1 => 1UL,
|
||||
2 or 3 or 16 or 17 or 19 => 2UL,
|
||||
11 or 12 => 8UL,
|
||||
13 => 12UL,
|
||||
14 => 16UL,
|
||||
_ => 4UL,
|
||||
};
|
||||
return checked((ulong)width * height * bytesPerPixel);
|
||||
}
|
||||
|
||||
private static byte[]? TakeGuestImageInitialData(ulong address)
|
||||
{
|
||||
lock (_gate)
|
||||
@@ -2811,6 +2837,8 @@ internal static unsafe class VulkanVideoPresenter
|
||||
private readonly HashSet<(ulong Address, uint Width, uint Height, Format Format)> _tracedTextureUploads = new();
|
||||
private readonly HashSet<(ulong Address, uint Width, uint Height, uint Format)> _dumpedTextures = new();
|
||||
private readonly HashSet<(ulong Address, uint Width, uint Height, uint Format)> _tracedTextureUploadContents = new();
|
||||
private readonly HashSet<(ulong Address, int ActualSize, ulong ExpectedSize, Format Format)>
|
||||
_rejectedGuestImageUploads = new();
|
||||
private readonly HashSet<(ulong Address, int Size)> _tracedGlobalBuffers = new();
|
||||
private readonly HashSet<(ulong Address, ulong Size)> _tracedGlobalWritebacks = new();
|
||||
private readonly HashSet<(ulong Shader, uint X, uint Y, uint Z, string Reason)>
|
||||
@@ -5583,7 +5611,20 @@ internal static unsafe class VulkanVideoPresenter
|
||||
DependencyCount = 1,
|
||||
PDependencies = &dependency,
|
||||
};
|
||||
Check(_vk.CreateRenderPass(_device, &renderPassInfo, null, out _renderPass), "vkCreateRenderPass");
|
||||
Check(
|
||||
_vk.CreateRenderPass(
|
||||
_device,
|
||||
&renderPassInfo,
|
||||
null,
|
||||
out var swapchainRenderPass),
|
||||
"vkCreateRenderPass");
|
||||
if (swapchainRenderPass.Handle == 0)
|
||||
{
|
||||
throw new InvalidOperationException(
|
||||
"vkCreateRenderPass returned a null swapchain render pass");
|
||||
}
|
||||
|
||||
_renderPass = swapchainRenderPass;
|
||||
|
||||
_swapchainImageViews = new ImageView[_swapchainImages.Length];
|
||||
_framebuffers = new Framebuffer[_swapchainImages.Length];
|
||||
@@ -5610,7 +5651,7 @@ internal static unsafe class VulkanVideoPresenter
|
||||
var framebufferInfo = new FramebufferCreateInfo
|
||||
{
|
||||
SType = StructureType.FramebufferCreateInfo,
|
||||
RenderPass = _renderPass,
|
||||
RenderPass = swapchainRenderPass,
|
||||
AttachmentCount = 1,
|
||||
PAttachments = &imageView,
|
||||
Width = _extent.Width,
|
||||
@@ -8221,22 +8262,25 @@ internal static unsafe class VulkanVideoPresenter
|
||||
var shadow = allocation.Shadow.AsSpan(checked((int)guestOffset), guestBuffer.Length);
|
||||
if (!source.SequenceEqual(shadow))
|
||||
{
|
||||
var sharedReadOnly = string.Equals(
|
||||
allocation.QueueName,
|
||||
SharedReadOnlyGuestBufferQueue,
|
||||
StringComparison.Ordinal);
|
||||
if (sharedReadOnly &&
|
||||
(allocation.LastUseTimeline > _completedTimeline ||
|
||||
IsGuestBufferAllocationReferencedByOpenBatch(allocation)))
|
||||
{
|
||||
return CreateVersionedReadOnlyGlobalBufferResource(
|
||||
guestBuffer,
|
||||
expectedBias,
|
||||
size);
|
||||
}
|
||||
|
||||
// HOST_COHERENT does not permit racing a mapped CPU write with
|
||||
// an in-flight shader access. Retire prior users, publish their
|
||||
// dirty ranges to guest memory, then upload the current guest
|
||||
// bytes (which may be newer than the parser's captured array).
|
||||
if (string.Equals(
|
||||
allocation.QueueName,
|
||||
SharedReadOnlyGuestBufferQueue,
|
||||
StringComparison.Ordinal))
|
||||
{
|
||||
// Shared read-only storage can still be referenced by any
|
||||
// logical guest queue, so replacing its mapped contents
|
||||
// requires retiring every prior reader.
|
||||
WaitForAllGuestSubmissionsForCpuVisibility();
|
||||
WriteBackAllDirtyGuestBuffers();
|
||||
}
|
||||
else
|
||||
if (!sharedReadOnly)
|
||||
{
|
||||
// Writable aliases are private to one logical guest queue.
|
||||
// Retiring unrelated queues here recreates the global FIFO
|
||||
@@ -8287,6 +8331,72 @@ internal static unsafe class VulkanVideoPresenter
|
||||
};
|
||||
}
|
||||
|
||||
private bool IsGuestBufferAllocationReferencedByOpenBatch(
|
||||
GuestBufferAllocation allocation)
|
||||
{
|
||||
if (!_batchOpen)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
foreach (var resources in _batchResources)
|
||||
{
|
||||
foreach (var globalBuffer in resources.GlobalMemoryBuffers)
|
||||
{
|
||||
if (ReferenceEquals(globalBuffer?.Allocation, allocation))
|
||||
{
|
||||
return true;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
private GlobalBufferResource CreateVersionedReadOnlyGlobalBufferResource(
|
||||
GuestMemoryBuffer guestBuffer,
|
||||
ulong byteBias,
|
||||
ulong guestSize)
|
||||
{
|
||||
var descriptorSize = checked((guestSize + byteBias + 3) & ~3UL);
|
||||
var descriptorLength = checked((int)descriptorSize);
|
||||
var snapshot = GuestDataPool.Rent(descriptorLength);
|
||||
try
|
||||
{
|
||||
var snapshotData = snapshot.AsSpan(0, descriptorLength);
|
||||
snapshotData.Clear();
|
||||
guestBuffer.Data.AsSpan(0, guestBuffer.Length).CopyTo(
|
||||
snapshotData[checked((int)byteBias)..]);
|
||||
|
||||
var buffer = CreateHostBuffer(
|
||||
snapshotData,
|
||||
BufferUsageFlags.StorageBufferBit,
|
||||
out var memory,
|
||||
out var mapped);
|
||||
return new GlobalBufferResource
|
||||
{
|
||||
BaseAddress = guestBuffer.BaseAddress,
|
||||
Writable = false,
|
||||
WriteBackToGuest = false,
|
||||
Buffer = buffer,
|
||||
Memory = memory,
|
||||
Mapped = mapped + checked((nint)byteBias),
|
||||
Offset = 0,
|
||||
Size = descriptorSize,
|
||||
GuestOffset = byteBias,
|
||||
GuestSize = guestSize,
|
||||
};
|
||||
}
|
||||
finally
|
||||
{
|
||||
GuestDataPool.Return(snapshot);
|
||||
if (guestBuffer.Pooled)
|
||||
{
|
||||
GuestDataPool.Return(guestBuffer.Data);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private GlobalBufferResource CreateTransientGlobalBufferResource(
|
||||
GuestMemoryBuffer guestBuffer)
|
||||
{
|
||||
@@ -9029,23 +9139,59 @@ internal static unsafe class VulkanVideoPresenter
|
||||
12 => 8UL,
|
||||
13 => 12UL,
|
||||
14 => 16UL,
|
||||
16 => 2UL,
|
||||
17 => 2UL,
|
||||
19 => 2UL,
|
||||
_ => 4UL,
|
||||
};
|
||||
|
||||
private static ulong GetTextureByteCount(uint format, uint width, uint height)
|
||||
=> GetGuestImageByteCount(format, width, height);
|
||||
|
||||
private static ulong GetVulkanImageByteCount(Format format, uint width, uint height)
|
||||
{
|
||||
var blockBytes = format switch
|
||||
{
|
||||
// BC1 (169/170) and BC4 (175/176) are 8 bytes per 4x4 block;
|
||||
// BC2/BC3/BC5/BC6H/BC7 are 16 bytes per block.
|
||||
169 or 170 or 175 or 176 => 8UL,
|
||||
171 or 172 or 173 or 174 or
|
||||
177 or 178 or 179 or 180 or 181 or 182 => 16UL,
|
||||
Format.BC1RgbUnormBlock or
|
||||
Format.BC1RgbSrgbBlock or
|
||||
Format.BC1RgbaUnormBlock or
|
||||
Format.BC1RgbaSrgbBlock or
|
||||
Format.BC4UnormBlock or
|
||||
Format.BC4SNormBlock => 8UL,
|
||||
Format.BC2UnormBlock or
|
||||
Format.BC2SrgbBlock or
|
||||
Format.BC3UnormBlock or
|
||||
Format.BC3SrgbBlock or
|
||||
Format.BC5UnormBlock or
|
||||
Format.BC5SNormBlock or
|
||||
Format.BC6HUfloatBlock or
|
||||
Format.BC6HSfloatBlock or
|
||||
Format.BC7UnormBlock or
|
||||
Format.BC7SrgbBlock => 16UL,
|
||||
_ => 0UL,
|
||||
};
|
||||
return blockBytes == 0
|
||||
? checked((ulong)width * height * GetTextureBytesPerPixel(format))
|
||||
: checked(((ulong)width + 3) / 4 * (((ulong)height + 3) / 4) * blockBytes);
|
||||
if (blockBytes != 0)
|
||||
{
|
||||
return checked(((ulong)width + 3) / 4 * (((ulong)height + 3) / 4) * blockBytes);
|
||||
}
|
||||
|
||||
var bitsPerTexel = GetFormatCompatibilityClass(format);
|
||||
if (bitsPerTexel == 0)
|
||||
{
|
||||
bitsPerTexel = format switch
|
||||
{
|
||||
Format.B5G6R5UnormPack16 or
|
||||
Format.R5G5B5A1UnormPack16 or
|
||||
Format.R4G4B4A4UnormPack16 => 16,
|
||||
Format.B8G8R8A8Unorm or
|
||||
Format.B8G8R8A8Srgb => 32,
|
||||
_ => 0,
|
||||
};
|
||||
}
|
||||
|
||||
return bitsPerTexel == 0
|
||||
? 0
|
||||
: checked((ulong)width * height * bitsPerTexel / 8);
|
||||
}
|
||||
|
||||
private bool SupportsColorAttachment(Format format)
|
||||
@@ -10184,7 +10330,10 @@ internal static unsafe class VulkanVideoPresenter
|
||||
TakeGuestImageInitialData(work.Targets[index].Address) is { } initialData &&
|
||||
!targets[index].Initialized &&
|
||||
(ulong)initialData.Length ==
|
||||
(ulong)targets[index].Width * targets[index].Height * 4)
|
||||
GetTextureByteCount(
|
||||
targetDescriptor.Format,
|
||||
targets[index].Width,
|
||||
targets[index].Height))
|
||||
{
|
||||
UploadGuestImageInitialData(targets[index], initialData);
|
||||
}
|
||||
@@ -10733,7 +10882,31 @@ internal static unsafe class VulkanVideoPresenter
|
||||
|
||||
private void UploadGuestImageInitialData(GuestImageResource target, byte[] pixels)
|
||||
{
|
||||
var byteCount = (ulong)pixels.Length;
|
||||
var guestDataFormat = (target.GuestFormat & 0x8000_0000u) != 0
|
||||
? (target.GuestFormat >> 8) & 0x1FFu
|
||||
: 0;
|
||||
var uploadPixels = guestDataFormat == 13
|
||||
? ExpandRgb32Pixels(pixels)
|
||||
: pixels;
|
||||
var expectedByteCount = GetVulkanImageByteCount(
|
||||
target.Format,
|
||||
target.Width,
|
||||
target.Height);
|
||||
if (expectedByteCount == 0 || (ulong)uploadPixels.Length != expectedByteCount)
|
||||
{
|
||||
if (_rejectedGuestImageUploads.Add(
|
||||
(target.Address, uploadPixels.Length, expectedByteCount, target.Format)))
|
||||
{
|
||||
Console.Error.WriteLine(
|
||||
$"[LOADER][WARN] Vulkan rejected incompatible guest image upload " +
|
||||
$"addr=0x{target.Address:X16} size={target.Width}x{target.Height} " +
|
||||
$"format={target.Format} bytes={uploadPixels.Length} expected={expectedByteCount}");
|
||||
}
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
var byteCount = (ulong)uploadPixels.Length;
|
||||
var staging = CreateBuffer(
|
||||
byteCount,
|
||||
BufferUsageFlags.TransferSrcBit,
|
||||
@@ -10745,9 +10918,13 @@ internal static unsafe class VulkanVideoPresenter
|
||||
Check(
|
||||
_vk.MapMemory(_device, stagingMemory, 0, byteCount, 0, &mapped),
|
||||
"vkMapMemory(guest image init)");
|
||||
fixed (byte* source = pixels)
|
||||
fixed (byte* source = uploadPixels)
|
||||
{
|
||||
System.Buffer.MemoryCopy(source, mapped, pixels.Length, pixels.Length);
|
||||
System.Buffer.MemoryCopy(
|
||||
source,
|
||||
mapped,
|
||||
uploadPixels.Length,
|
||||
uploadPixels.Length);
|
||||
}
|
||||
|
||||
_vk.UnmapMemory(_device, stagingMemory);
|
||||
|
||||
Reference in New Issue
Block a user