A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
wip
+52
-13
+41
-10
libvektor/include/vektor/vektor.h
+41
-10
libvektor/include/vektor/vektor.h
···
19
19
Transfer
20
20
};
21
21
22
+
enum class Stage {
23
+
Transfer,
24
+
Compute,
25
+
RasterColorOut,
26
+
PixelShader,
27
+
VertexShader
28
+
};
29
+
enum class Signal {
30
+
AtomicSet,
31
+
AtomicMax,
32
+
AtomicOr,
33
+
};
34
+
enum class Op {
35
+
Never,
36
+
Less,
37
+
Equal,
38
+
};
39
+
40
+
enum HazardFlags {
41
+
// @todo
42
+
DrawArguments = 1 << 0,
43
+
Descriptors = 1 << 1
44
+
};
45
+
46
+
enum class Memory { Default, Gpu, Readback };
47
+
48
+
typedef uint64_t gpuptr_t;
22
49
typedef void *Device;
23
50
typedef void *Queue;
24
51
typedef void *CommandList;
25
-
26
-
typedef uint64_t gpuptr_t;
27
-
28
-
Version version();
29
-
30
-
Device create();
31
-
void destroy(Device);
32
-
33
-
enum class Memory { Default, Gpu, Readback };
52
+
typedef void *Semaphore;
34
53
35
54
struct Allocation {
36
55
void *cpu;
···
40
59
};
41
60
42
61
62
+
Version version();
63
+
64
+
Device create();
65
+
void destroy(Device);
66
+
43
67
Allocation malloc(Device, std::size_t size, Memory memory = Memory::Default);
44
68
Allocation malloc(Device, std::size_t size, std::size_t align, Memory memory = Memory::Default);
45
69
void free(Device, Allocation &);
46
70
71
+
Semaphore create_semaphore(uint64_t value);
72
+
47
73
Queue create_queue(Device, QueueType);
48
74
CommandList start_recording(Queue);
49
75
50
-
void submit(Queue, CommandList);
76
+
void submit(Queue, CommandList, Semaphore = nullptr, uint64_t value = 0);
51
77
52
78
void memset(CommandList, gpuptr_t addr, std::size_t size, uint32_t value);
79
+
80
+
void signal_after(CommandList, Stage before, gpuptr_t ptr, uint64_t value, Signal);
81
+
void wait_before(CommandList, Stage after, gpuptr_t ptr, uint64_t value, Op, HazardFlags hazard = 0, uint64 mask = ~0);
82
+
void wait_semaphore(Semaphore, uint64_t value);
83
+
void signal_semaphore(Semaphore, uint64_t value);
53
84
54
85
};
+3
-1
libvektor/src/amdgpu/cmdstream.h
+3
-1
libvektor/src/amdgpu/cmdstream.h
···
40
40
struct Submission {
41
41
uint32_t start_dw;
42
42
uint32_t end_dw;
43
-
amdgpu_cs_fence fence;
43
+
uint64_t point;
44
44
};
45
45
46
46
void wait_for_space(uint32_t target_dw_offset);
···
53
53
amdgpu_bo_handle m_bo_handle;
54
54
uint64_t m_gpu_va;
55
55
uint32_t* m_cpu_map;
56
+
57
+
uint64_t m_timeline_counter = 0;
56
58
57
59
uint32_t m_write_cursor_dw = 0;
58
60
std::deque<Submission> m_history;
+4
libvektor/src/vektor_impl.h
+4
libvektor/src/vektor_impl.h
+4
-2
libvektor/src/vektor_queue.cpp
+4
-2
libvektor/src/vektor_queue.cpp
···
48
48
return cl;
49
49
}
50
50
51
-
void submit(Queue pq, CommandList pcl) {
51
+
void submit(Queue pq, CommandList pcl, Semaphore sem, uint64_t value) {
52
52
auto *queue = (QueueImpl *)pq;
53
53
auto *cl = (CommandListImpl *)pcl;
54
54
assert(cl->queue == queue, "submit: commandlist from foreign queue");
55
55
56
-
queue->cmd_ring->submit(cl->cs);
56
+
auto *semaphore = (SemaphoreImpl *)sem;
57
+
58
+
queue->cmd_ring->submit(cl->cs, semaphore, value);
57
59
58
60
// @todo: to free commandlist, we want to be sure that it is no longer mapped and stuff.
59
61
// then, we can freely-free it. But i think this needs some deferred-cleanup, as