我正在将Vulkan与GLSL计算着色器一起使用,以对在OpenGL中渲染的纹理进行后期处理。它在大多数情况下都可以正常工作,但是当纹理的垂直分辨率在一定范围内时,我会看到非常奇怪的异常,其中gl_GlobalInvocationID似乎被弄乱了。
一个示例计算着色器,演示了如下情况
Sub test()
Dim user As String
Dim pwd As String
Dim path As String
user = ""
pwd = ""
path = ""
Debug.Print httpGET(path, user, pwd)
End Sub
Public Function httpGET(fn As String, _
Optional authUser As String = vbNullString, _
Optional authPass As String = vbNullString) As String
pHtml = fn
Dim oHttp As Object
Set oHttp = CreateObject("Microsoft.XMLHTTP")
Call oHttp.Open("GET", pHtml, False)
If (authUser <> vbNullString) Then
' need to do basic authentication
' acknowledgement to http://pastie.org/1192157
oHttp.SetRequestHeader "Content-Type", "application/json"
oHttp.SetRequestHeader "Accept", "application/json"
oHttp.SetRequestHeader "Authorization", "Basic " + _
EncodeBase64(authUser + ":" + authPass)
End If
Call oHttp.Send("")
httpGET = oHttp.ResponseText
Set oHttp = Nothing
End Function
Function EncodeBase64(text As String) As String
Dim arrData() As Byte
arrData = StrConv(text, vbFromUnicode)
Dim objXML As MSXML2.DOMDocument
Dim objNode As MSXML2.IXMLDOMElement
Set objXML = New MSXML2.DOMDocument
Set objNode = objXML.createElement("b64")
objNode.DataType = "bin.base64"
objNode.nodeTypedValue = arrData
EncodeBase64 = Application.Clean(objNode.text)
Set objNode = Nothing
Set objXML = Nothing
End Function
根据垂直分辨率,我得到奇怪的结果:482 +,241-256、121-128、55-64可以正常工作,但是257-481、129-240、65-121和<55不正确。见下文。
我通过以下方式创建纹理:
#version 450
#extension GL_ARB_separate_shader_objects : enable
#define WORKGROUP_SIZE 32
layout(local_size_x=WORKGROUP_SIZE, local_size_y=WORKGROUP_SIZE, local_size_z=1) in;
layout(binding=0, rgba8) uniform image2D input_tex;
layout(binding=1, rgba8) uniform image2D output_tex;
void main()
{
vec4 red = vec4(1.0, 0.0, 0.0, 1.0);
vec4 black = vec4(0.0, 0.0, 0.0, 1.0);
if(gl_GlobalInvocationID.x == 0 || gl_GlobalInvocationID.y == 0)
{
imageStore(output_tex, ivec2(gl_GlobalInvocationID.xy), red);
}
else
{
imageStore(output_tex, ivec2(gl_GlobalInvocationID.xy), black);
}
}
我在OpenGL中渲染为纹理,然后使用vk::ImageCreateInfo image_create_info(
vk::ImageCreateFlags(),
vk::ImageType::e2D,
vk::Format::eR8G8B8A8Unorm,
{width, height, 1},
1,
1,
vk::SampleCountFlagBits::e1,
vk::ImageTiling::eOptimal,
vk::ImageUsageFlagBits::eColorAttachment |
vk::ImageUsageFlagBits::eSampled |
vk::ImageUsageFlagBits::eTransferSrc |
vk::ImageUsageFlagBits::eStorage
);
image = device.createImage(image_create_info);
vk::MemoryRequirements memory_requirements = device.getImageMemoryRequirements(image);
memory_size = memory_requirements.size;
vk::ExportMemoryAllocateInfo export_allocate_info(
vk::ExternalMemoryHandleTypeFlagBits::eOpaqueFd);
vk::MemoryAllocateInfo memory_allocate_info(
memory_size,
memory_type // Standard way of finding this using eDeviceLocal
);
memory_allocate_info.pNext = &export_allocate_info;
device_memory = device.allocateMemory(memory_allocate_info);
device.bindImageMemory(image, device_memory, 0);
vk::MemoryGetFdInfoKHR memory_info(device_memory, vk::ExternalMemoryHandleTypeFlagBits::eOpaqueFd);
opaque_handle = device.getMemoryFdKHR(memory_info, dynamic_loader);
// Create OpenGL texture
glCreateTextures(GL_TEXTURE_2D, 1, &opengl_tex);
glCreateMemoryObjectsEXT(1, &opengl_memory);
glImportMemoryFdEXT(opengl_memory, memory_size, GL_HANDLE_TYPE_OPAQUE_FD_EXT, opaque_handle);
glTextureStorageMem2DEXT(opengl_tex, 1, GL_RGBA8, width, height, opengl_memory, 0);
从上方将渲染的图像复制到glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_COLOR_BUFFER_BIT, GL_NEAREST)
,然后执行包含上述着色器的命令缓冲区。之后,我将纹理回弹并在OpenGL中渲染到屏幕上。 (使用信号量进行同步)。
命令缓冲区的创建看起来像
opengl_tex
我检查了好几次,我并没有对某个地方的大小(除了工作组大小)进行硬编码。
我还制作了一个实用程序,可以读取png图像并执行相同的着色器并写出输出(绕过OpenGL互操作),并且无论尺寸大小都可以工作。
我需要能够处理480x480图像,因此我们将不胜感激。如果您需要其他任何信息,我可以提供更多代码。
(仅供参考,我在Linux和Vulkan 1.1.85.0上使用带有NVidia驱动程序410.78的GTX 1080)
编辑: 使用以下着色器代码段:
vk::CommandBufferBeginInfo begin_info(
vk::CommandBufferUsageFlagBits());
command_buffer.begin(begin_info);
command_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline);
command_buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute, pipeline_layout, 0, 1, &descriptor_set, 0, nullptr);
command_buffer.dispatch((uint32_t)ceil(width/float(WORKGROUP_SIZE), (uint32_t)ceil(height/float(WORKGROUP_SIZE)), 1);
command_buffer.end();
我得到结果
480x482图片工作:
480x480图像损坏: