使用金属间接命令缓冲区时出错："Fragment shader cannot be used with indirect command buffers"

Question

我正在开发一个基于 MetalMTKView的应用程序，该应用程序利用 A11 TBDR 架构在单个渲染通道中进行延迟着色。我使用了Apple的延迟照明示例代码作为参考，效果很好。

我想尝试使用 A11 硬件上 Metal 2 的间接命令缓冲区功能将几何缓冲区传递更改为 GPU 驱动。

我一直在 GPU 示例代码上使用 Apple 的编码间接命令缓冲区作为我的主要参考点。我可以在我的iPhone XR上运行此示例(尽管可能偏离主题，滚动不流畅，它会抖动)。

但是，当我尝试将几何缓冲区传递移动到间接命令缓冲区时，我自己的代码遇到了困难。当我在几何缓冲区管道的MTLRenderPipelineDescriptor上将supportIndirectCommandBuffers设置为true时，device.makeRenderPipelineState失败并显示错误

AGXMetalA12 Code=3 "片段着色器不能与间接命令缓冲区一起使用">

我无法在文档中找到有关此错误的任何信息。我想知道，间接管道中是否不允许某些类型的片段操作，或者我忽略了对 GPU 驱动绘图的某种限制(也许是颜色附件的数量)？

SharedTypes.h

Header 由 Metal 和 Swift 共享

#ifndef SharedTypes_h
#define SharedTypes_h
#ifdef __METAL_VERSION__
#define NS_CLOSED_ENUM(_type, _name) enum _name : _type _name; enum _name : _type
#define NSInteger metal::int32_t
#else
#import <Foundation/Foundation.h>
#endif
#include <simd/simd.h>
typedef struct {
uint32_t meshId;
matrix_float3x3 normalViewMatrix;
matrix_float4x4 modelMatrix;
matrix_float4x4 shadowMVPTransformMatrix;
} InstanceData;
typedef struct {
vector_float3 cameraPosition;
float voxelScale;
float blockScale;
vector_float3 lightDirection;
matrix_float4x4 viewMatrix;
matrix_float4x4 projectionMatrix;
matrix_float4x4 projectionMatrixInverse;
matrix_float4x4 shadowViewProjectionMatrix;
} VoxelUniforms;
typedef NS_CLOSED_ENUM(NSInteger, BufferIndex)
{
BufferIndexInstances  = 0,
BufferIndexVertices = 1,
BufferIndexIndices = 2,
BufferIndexVoxelUniforms = 3,
};
typedef NS_CLOSED_ENUM(NSInteger, RenderTarget)
{
RenderTargetLighting = 0,
RenderTargetNormal_shadow = 1,
RenderTargetVoxelIndex = 2,
RenderTargetDepth = 3,
};
#endif /* SharedTypes_h */

GBuffer 着色器

#include <metal_stdlib>
using namespace metal;
#include "../SharedTypes.h"
struct VertexIn {
packed_half3 position;
packed_half3 texCoord3D;
half ambientOcclusion;
uchar normalIndex;
};
struct VertexInOut {
float4 position [[ position ]];
half3 worldPos;
half3 eyeNormal;
half3 localPosition;
half3 localNormal;
float eyeDepth;
float3 shadowCoord;
half3 texCoord3D;
};
vertex VertexInOut gBufferVertex(device InstanceData* instances [[ buffer( BufferIndexInstances ) ]],
device VertexIn* vertices [[ buffer( BufferIndexVertices ) ]],
constant VoxelUniforms &uniforms [[ buffer( BufferIndexVoxelUniforms ) ]],
uint vid [[ vertex_id ]],
ushort iid [[ instance_id ]])
{
InstanceData instance = instances[iid];
VertexIn vert = vertices[vid];
VertexInOut out;
float4 position = float4(float3(vert.position), 1);
float4 worldPos = instance.modelMatrix * position;
float4 eyePosition = uniforms.viewMatrix * worldPos;
out.position = uniforms.projectionMatrix * eyePosition;
out.worldPos = half3(worldPos.xyz);
out.eyeDepth = eyePosition.z;
half3 normal = normals[vert.normalIndex];
out.eyeNormal = half3(instance.normalViewMatrix * float3(normal));
out.shadowCoord = (instance.shadowMVPTransformMatrix * position).xyz;
out.localPosition = half3(vert.position);
out.localNormal = normal;
out.texCoord3D = half3(vert.texCoord3D);
return out;
}
fragment GBufferData gBufferFragment(VertexInOut in [[ stage_in ]],
constant VoxelUniforms &uniforms [[ buffer( BufferIndexVoxelUniforms ) ]],
texture3d<ushort, access::sample> voxelMap [[ texture(0) ]],
depth2d<float> shadowMap [[ texture(1) ]],
texture3d<half, access::sample> fogOfWarMap [[ texture(2) ]]
) {
// voxel index
half3 center = round(in.texCoord3D);
uchar voxIndex = voxelMap.read(ushort3(center)).r - 1;
// ambient occlusion
half3 neighborPos = center + in.localNormal;
half3 absNormal = abs(in.localNormal);
half2 texCoord2D = tc2d(in.localPosition / uniforms.voxelScale, absNormal);
half ao = getAO(voxelMap, neighborPos, absNormal.yzx, absNormal.zxy, texCoord2D);
// shadow
constexpr sampler shadowSampler(coord::normalized,
filter::linear,
mip_filter::none,
address::clamp_to_edge,
compare_func::less);
float shadow_sample = ambientLightingLevel;
for (short i = 0; i < shadowSampleCount; i++){
shadow_sample += shadowMap.sample_compare(shadowSampler, in.shadowCoord.xy + poissonDisk[i] * 0.002, in.shadowCoord.z - 0.0018) * shadowContributionPerSample;
}
shadow_sample = min(1.0, shadow_sample);
//fog-of-war
half fogOfWarSample = fogOfWarMap.sample(fogOfWarSampler, (float3(in.worldPos) / uniforms.blockScale) + float3(0.5, 0.4, 0.5)).r;
half notVisible = max(fogOfWarSample, 0.5h);
// output
GBufferData out;
out.normal_shadow = half4(in.eyeNormal, ao * half(shadow_sample) * notVisible);
out.voxelIndex = voxIndex;
out.depth = in.eyeDepth;
return out;
};

管道设置

extension RenderTarget {
var pixelFormat: MTLPixelFormat {
switch self {
case .lighting: return .bgra8Unorm
case .normal_shadow: return .rgba8Snorm
case .voxelIndex: return .r8Uint
case .depth: return .r32Float
}
}
static var allCases: [RenderTarget] = [.lighting, .normal_shadow, .voxelIndex, .depth]
}
public final class GBufferRenderer {
private let renderPipelineState: MTLRenderPipelineState
weak var shadowMap: MTLTexture?
public init(depthPixelFormat: MTLPixelFormat, colorPixelFormat: MTLPixelFormat, sampleCount: Int = 1) throws {
let library = try LibraryMonad.getLibrary()
let device = library.device
let descriptor = MTLRenderPipelineDescriptor()
descriptor.vertexFunction = library.makeFunction(name: "gBufferVertex")!
descriptor.fragmentFunction = library.makeFunction(name: "gBufferFragment")!
descriptor.depthAttachmentPixelFormat = depthPixelFormat
descriptor.stencilAttachmentPixelFormat = depthPixelFormat
descriptor.sampleCount = sampleCount
for target in RenderTarget.allCases {
descriptor.colorAttachments[target.rawValue].pixelFormat = target.pixelFormat
}
// uncomment below to trigger throw
// descriptor.supportIndirectCommandBuffers = true
renderPipelineState = try device.makeRenderPipelineState(descriptor: descriptor) // throws "Fragment shader cannot be used with indirect command buffers"
}
public convenience init(mtkView: MTKView) throws {
try self.init(depthPixelFormat: mtkView.depthStencilPixelFormat, colorPixelFormat: mtkView.colorPixelFormat, sampleCount: mtkView.sampleCount)
}
}

当以通常的方式从 CPU 触发绘制时，上述方法效果很好，但是当设置supportIndirectCommandBuffers以准备 GPU 绘制时，它会抛出错误。

我尝试剥离片段着色器以仅返回 GBuffer 的常量值，然后makeRenderPipelineState成功，但是当我重新添加纹理采样时，它又开始抱怨。我似乎无法确定它到底不喜欢碎片着色器的地方。

Answer 1

浏览代码以及 Metal 文档和 Metal 着色语言规范，我想我知道你为什么会收到这个错误。

如果您浏览 Metal 中metal_command_buffer标头中存在的render_command接口，您会发现要将参数传递给间接渲染命令，您只有以下功能：set_vertex_buffer和set_fragment_buffer，没有像MTLRenderCommandEncoder中那样set_vertex_texture或set_vertex_sampler。

但是，由于管道使用着色器，而着色器又使用纹理作为参数，并且你通过使用supportIndirectCommandBuffers指示要在间接命令中使用此管道，因此 Metal 别无选择，只能使管道创建失败。

相反，如果要将纹理或采样器传递给间接渲染命令，则应使用参数缓冲区，该缓冲区将传递给发出间接渲染命令的着色器，而着色器又将使用每个render_command的set_vertex_buffer和set_fragment_buffer绑定它们。

规范：金属着色语言规范(第 5.16 节)