Skip to content

Commit a15d1da

Browse files
devshgraphicsprogrammingdevshgraphicsprogramming
committed
CUDA Handler initialization and device finding
1 parent ae353c4 commit a15d1da

File tree

8 files changed

+826
-706
lines changed

8 files changed

+826
-706
lines changed

‎examples_tests/36.CUDAInterop/main.cpp

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -6,8 +6,6 @@
66
#include <nabla.h>
77

88

9-
#include "../src/nbl/video/CCUDAHandler.h"
10-
119
/**
1210
This example just shows a screen which clears to red,
1311
nothing fancy, just to show that Irrlicht links fine

‎include/nbl/video/CCUDADevice.h

Lines changed: 195 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,195 @@
1+
// Copyright (C) 2018-2020 - DevSH Graphics Programming Sp. z O.O.
2+
// This file is part of the "Nabla Engine".
3+
// For conditions of distribution and use, see copyright notice in nabla.h
4+
#ifndef _NBL_VIDEO_C_CUDA_DEVICE_H_
5+
#define _NBL_VIDEO_C_CUDA_DEVICE_H_
6+
7+
8+
#include "nbl/video/IPhysicalDevice.h"
9+
10+
11+
#ifdef _NBL_COMPILE_WITH_CUDA_
12+
13+
#include "cuda.h"
14+
#include "nvrtc.h"
15+
#if CUDA_VERSION < 9000
16+
#error "Need CUDA 9.0 SDK or higher."
17+
#endif
18+
19+
// useful includes in the future
20+
//#include "cudaEGL.h"
21+
//#include "cudaVDPAU.h"
22+
23+
namespace nbl::video
24+
{
25+
class CCUDAHandler;
26+
27+
class CCUDADevice : public core::IReferenceCounted
28+
{
29+
public:
30+
enum E_VIRTUAL_ARCHITECTURE
31+
{
32+
EVA_30,
33+
EVA_32,
34+
EVA_35,
35+
EVA_37,
36+
EVA_50,
37+
EVA_52,
38+
EVA_53,
39+
EVA_60,
40+
EVA_61,
41+
EVA_62,
42+
EVA_70,
43+
EVA_72,
44+
EVA_75,
45+
EVA_80,
46+
EVA_COUNT
47+
};
48+
static inline constexpr const char* virtualArchCompileOption[] = {
49+
"-arch=compute_30",
50+
"-arch=compute_32",
51+
"-arch=compute_35",
52+
"-arch=compute_37",
53+
"-arch=compute_50",
54+
"-arch=compute_52",
55+
"-arch=compute_53",
56+
"-arch=compute_60",
57+
"-arch=compute_61",
58+
"-arch=compute_62",
59+
"-arch=compute_70",
60+
"-arch=compute_72",
61+
"-arch=compute_75",
62+
"-arch=compute_80"
63+
};
64+
inline E_VIRTUAL_ARCHITECTURE getVirtualArchitecture() {return m_virtualArchitecture;}
65+
66+
inline const auto& geDefaultCompileOptions() const {return m_defaultCompileOptions;}
67+
68+
// TODO/REDO Vulkan: https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__EXTRES__INTEROP.html
69+
// https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#vulkan-interoperability
70+
// Watch out, use Driver API (`cu` functions) NOT the Runtime API (`cuda` functions)
71+
// Also maybe separate this out into its own `CCUDA` class instead of nesting it here?
72+
#if 0
73+
template<typename ObjType>
74+
struct GraphicsAPIObjLink
75+
{
76+
GraphicsAPIObjLink() : obj(nullptr), cudaHandle(nullptr), acquired(false)
77+
{
78+
asImage = {nullptr};
79+
}
80+
GraphicsAPIObjLink(core::smart_refctd_ptr<ObjType>&& _obj) : GraphicsAPIObjLink()
81+
{
82+
obj = std::move(_obj);
83+
}
84+
GraphicsAPIObjLink(GraphicsAPIObjLink&& other) : GraphicsAPIObjLink()
85+
{
86+
operator=(std::move(other));
87+
}
88+
89+
GraphicsAPIObjLink(const GraphicsAPIObjLink& other) = delete;
90+
GraphicsAPIObjLink& operator=(const GraphicsAPIObjLink& other) = delete;
91+
GraphicsAPIObjLink& operator=(GraphicsAPIObjLink&& other)
92+
{
93+
std::swap(obj,other.obj);
94+
std::swap(cudaHandle,other.cudaHandle);
95+
std::swap(acquired,other.acquired);
96+
std::swap(asImage,other.asImage);
97+
return *this;
98+
}
99+
100+
~GraphicsAPIObjLink()
101+
{
102+
assert(!acquired); // you've fucked up, there's no way for us to fix it, you need to release the objects on a proper stream
103+
if (obj)
104+
CCUDAHandler::cuda.pcuGraphicsUnregisterResource(cudaHandle);
105+
}
106+
107+
//
108+
auto* getObject() const {return obj.get();}
109+
110+
private:
111+
core::smart_refctd_ptr<ObjType> obj;
112+
CUgraphicsResource cudaHandle;
113+
bool acquired;
114+
115+
friend class CCUDAHandler;
116+
public:
117+
union
118+
{
119+
struct
120+
{
121+
CUdeviceptr pointer;
122+
} asBuffer;
123+
struct
124+
{
125+
CUmipmappedArray mipmappedArray;
126+
CUarray array;
127+
} asImage;
128+
};
129+
};
130+
131+
//
132+
static CUresult registerBuffer(GraphicsAPIObjLink<video::IGPUBuffer>* link, uint32_t flags = CU_GRAPHICS_REGISTER_FLAGS_NONE);
133+
static CUresult registerImage(GraphicsAPIObjLink<video::IGPUImage>* link, uint32_t flags = CU_GRAPHICS_REGISTER_FLAGS_NONE);
134+
135+
136+
template<typename ObjType>
137+
static CUresult acquireResourcesFromGraphics(void* tmpStorage, GraphicsAPIObjLink<ObjType>* linksBegin, GraphicsAPIObjLink<ObjType>* linksEnd, CUstream stream)
138+
{
139+
auto count = std::distance(linksBegin,linksEnd);
140+
141+
auto resources = reinterpret_cast<CUgraphicsResource*>(tmpStorage);
142+
auto rit = resources;
143+
for (auto iit=linksBegin; iit!=linksEnd; iit++,rit++)
144+
{
145+
if (iit->acquired)
146+
return CUDA_ERROR_UNKNOWN;
147+
*rit = iit->cudaHandle;
148+
}
149+
150+
auto retval = cuda.pcuGraphicsMapResources(count,resources,stream);
151+
for (auto iit=linksBegin; iit!=linksEnd; iit++)
152+
iit->acquired = true;
153+
return retval;
154+
}
155+
template<typename ObjType>
156+
static CUresult releaseResourcesToGraphics(void* tmpStorage, GraphicsAPIObjLink<ObjType>* linksBegin, GraphicsAPIObjLink<ObjType>* linksEnd, CUstream stream)
157+
{
158+
auto count = std::distance(linksBegin,linksEnd);
159+
160+
auto resources = reinterpret_cast<CUgraphicsResource*>(tmpStorage);
161+
auto rit = resources;
162+
for (auto iit=linksBegin; iit!=linksEnd; iit++,rit++)
163+
{
164+
if (!iit->acquired)
165+
return CUDA_ERROR_UNKNOWN;
166+
*rit = iit->cudaHandle;
167+
}
168+
169+
auto retval = cuda.pcuGraphicsUnmapResources(count,resources,stream);
170+
for (auto iit=linksBegin; iit!=linksEnd; iit++)
171+
iit->acquired = false;
172+
return retval;
173+
}
174+
175+
static CUresult acquireAndGetPointers(GraphicsAPIObjLink<video::IGPUBuffer>* linksBegin, GraphicsAPIObjLink<video::IGPUBuffer>* linksEnd, CUstream stream, size_t* outbufferSizes = nullptr);
176+
static CUresult acquireAndGetMipmappedArray(GraphicsAPIObjLink<video::IGPUImage>* linksBegin, GraphicsAPIObjLink<video::IGPUImage>* linksEnd, CUstream stream);
177+
static CUresult acquireAndGetArray(GraphicsAPIObjLink<video::IGPUImage>* linksBegin, GraphicsAPIObjLink<video::IGPUImage>* linksEnd, uint32_t* arrayIndices, uint32_t* mipLevels, CUstream stream);
178+
#endif
179+
180+
protected:
181+
friend class CCUDAHandler;
182+
CCUDADevice(core::smart_refctd_ptr<CVulkanConnection>&& _vulkanConnection, IPhysicalDevice* const _vulkanDevice, const E_VIRTUAL_ARCHITECTURE _virtualArchitecture);
183+
~CCUDADevice() = default;
184+
185+
std::vector<const char*> m_defaultCompileOptions;
186+
core::smart_refctd_ptr<CVulkanConnection> m_vulkanConnection;
187+
IPhysicalDevice* const m_vulkanDevice;
188+
E_VIRTUAL_ARCHITECTURE m_virtualArchitecture;
189+
};
190+
191+
}
192+
193+
#endif // _NBL_COMPILE_WITH_CUDA_
194+
195+
#endif

0 commit comments

Comments
 (0)