1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
|
# SPDX-License-Identifier: MIT
from ..utils import chexdump
from ..malloc import Heap
from construct.core import *
from ..fw.agx.channels import *
from ..fw.agx.cmdqueue import *
from ..fw.agx.microsequence import *
from ..hw.uat import MemoryAttr
from .object import *
import textwrap
class GPUContext:
def __init__(self, agx):
self.agx = agx
self.uat = self.agx.uat
self.u = self.agx.u
self.p = self.agx.p
self.verbose = False
#self.job_list = agx.kshared.new(JobList)
#self.job_list.first_job = 0
#self.job_list.last_head = self.job_list._addr # Empty list has self as last_head
#self.job_list.unkptr_10 = 0
#self.job_list.push()
self.gpu_context = agx.kobj.new(GPUContextData).push()
self.ttbr0_base = self.u.memalign(self.agx.PAGE_SIZE, self.agx.PAGE_SIZE)
self.p.memset32(self.ttbr0_base, 0, self.agx.PAGE_SIZE)
self.objects = {}
# 32K VA pages since buffer manager needs that
self.uobj = GPUAllocator(agx, "Userspace", 0x1600000000, 0x100000000, ctx=None,
guard_pages=16,
va_block=32768, nG=1, AP=0, PXN=1, UXN=1)
self.gobj = GPUAllocator(agx, "GEM", 0x1500000000, 0x100000000, ctx=None,
guard_pages=16, nG=1, AP=0, PXN=1, UXN=1)
self.pipeline_base = 0x1100000000
self.pipeline_size = 1 << 32
self.pobj = GPUAllocator(agx, "Pipelines", self.pipeline_base + 0x10000, self.pipeline_size,
ctx=None, guard_pages=1, nG=1, AP=0, PXN=1, UXN=1)
def bind(self, ctx_id):
self.ctx = ctx_id
self.uobj.ctx = ctx_id
self.gobj.ctx = ctx_id
self.pobj.ctx = ctx_id
self.uat.bind_context(ctx_id, self.ttbr0_base)
self.thing = self.buf_at(0x6fffff8000, 0, 0x4000, "thing")
def make_stream(self, base):
return self.uat.iostream(self.ctx, base, recurse=False)
def new_at(self, addr, objtype, name=None, track=True, **flags):
obj = GPUObject(self, objtype)
obj._stream = self.make_stream
if name is not None:
obj._name = name
size_align = align_up(obj._size, self.agx.PAGE_SIZE)
obj._addr = addr
obj._paddr = self.agx.u.memalign(self.agx.PAGE_SIZE, size_align)
#if isinstance(obj.val, ConstructClassBase):
#obj.val._addr = obj._addr
self.agx.log(f"[Context@{self.gpu_context._addr:#x}] Map {obj._name} size {obj._size:#x} @ {obj._addr:#x} ({obj._paddr:#x})")
flags2 = {"AttrIndex": MemoryAttr.Shared}
flags2.update(flags)
obj._map_flags = flags2
obj._size_align = size_align
self.agx.uat.iomap_at(self.ctx, obj._addr, obj._paddr, size_align, **flags2)
self.objects[obj._addr] = obj
self.agx.reg_object(obj, track=track)
return obj
def buf_at(self, addr, is_pipeline, size, name=None, track=True):
return self.new_at(addr, Bytes(size), name, track=track,
AttrIndex=MemoryAttr.Shared, PXN=1,
nG=1, AP=(1 if is_pipeline else 0))
def load_blob(self, addr, is_pipeline, filename, track=True):
data = open(filename, "rb").read()
obj = self.new_at(addr, Bytes(len(data)), filename, track=track,
AttrIndex=MemoryAttr.Shared, PXN=1,
nG=1, AP=(1 if is_pipeline else 0))
obj.val = data
obj.push()
return obj
def free(self, obj):
obj._dead = True
self.agx.uat.iomap_at(self.ctx, obj._addr, 0, obj._size_align, VALID=0)
del self.objects[obj._addr]
self.agx.unreg_object(obj)
def free_at(self, addr):
self.free(self.objects[obj._addr])
class GPUWorkQueue:
def __init__(self, agx, context, job_list):
self.agx = agx
self.u = agx.u
self.p = agx.p
self.context = context
self.info = agx.kobj.new(CommandQueueInfo)
self.pointers = agx.kshared.new(CommandQueuePointers).push()
self.pmap = CommandQueuePointerMap(self.u, self.pointers._paddr)
self.rb_size = self.pointers.rb_size
self.ring = agx.kobj.new_buf(8 * self.rb_size, "GPUWorkQueue.RB")
self.info.pointers = self.pointers
self.info.rb_addr = self.ring._addr
self.info.job_list = job_list
self.info.gpu_buf_addr = agx.kobj.buf(0x2c18, "GPUWorkQueue.gpu_buf")
self.info.gpu_context = context.gpu_context
self.info.push()
self.wptr = 0
self.first_time = True
self.agx.uat.flush_dirty()
def submit(self, work):
work.push()
self.p.write64(self.ring._paddr + 8 * self.wptr, work._addr)
self.wptr = (self.wptr + 1) % self.rb_size
self.agx.uat.flush_dirty()
self.pmap.CPU_WPTR.val = self.wptr
def wait_empty(self):
while self.wptr != self.pmap.GPU_DONEPTR.val:
self.agx.work()
class GPU3DWorkQueue(GPUWorkQueue):
TYPE = 1
class GPUTAWorkQueue(GPUWorkQueue):
TYPE = 0
class GPUMicroSequence:
def __init__(self, agx):
self.agx = agx
self.off = 0
self.ops = []
self.obj = None
def append(self, op):
off = self.off
self.ops.append(op)
self.off += op.sizeof()
return off
def finalize(self):
self.ops.append(EndCmd())
self.size = sum(i.sizeof() for i in self.ops)
self.obj = self.agx.kobj.new_buf(self.size, "GPUMicroSequence", track=False)
self.obj.val = b"".join(i.build() for i in self.ops)
self.obj.push()
return self.obj
def dump(self):
chexdump(self.agx.iface.readmem(self.obj._paddr, self.size))
print(MicroSequence.parse_stream(self.agx.uat.iostream(0, self.obj._addr)))
def __str__(self):
s = f"GPUMicroSequence: {len(self.ops)} ops\n"
for i, op in enumerate(self.ops):
op_s = textwrap.indent(str(op), ' ' * 4)
s += f"[{i:2}:{op.sizeof():#x}] = {op!s}\n"
return s
class GPUBufferManager:
def __init__(self, agx, context, blocks=8):
self.agx = agx
self.ctx = context
self.block_ctl_obj = agx.kshared.new(BufferManagerBlockControl)
self.block_ctl_obj.total = blocks
self.block_ctl_obj.wptr = 0
self.block_ctl_obj.unk = 0
self.block_ctl = self.block_ctl_obj.push().regmap()
self.counter_obj = agx.kshared.new(BufferManagerCounter)
self.counter_obj.count = 0
self.counter = self.counter_obj.push().regmap()
self.misc_obj = agx.kshared.new(BufferManagerMisc)
self.misc_obj.cpu_flag = 1
self.misc = self.misc_obj.push().regmap()
self.page_size = 0x8000
self.pages_per_block = 4
self.block_size = self.pages_per_block * self.page_size
self.page_list = context.uobj.new(Array(0x10000 // 4, Int32ul), "BM PageList", track=False)
self.block_list = context.uobj.new(Array(0x8000 // 4, Int32ul), "BM BlockList", track=False)
self.info = info = agx.kobj.new(BufferManagerInfo)
info.page_list_addr = self.page_list._addr
info.page_list_size = self.page_list._size
info.page_count = self.block_ctl_obj.total * 4
info.block_count = self.block_ctl_obj.total
info.block_list_addr = self.block_list._addr
info.block_ctl = self.block_ctl_obj
info.last_page = info.page_count - 1
info.block_size = self.block_size
info.counter = self.counter_obj
self.populate()
self.block_ctl_obj.pull()
self.block_list.push()
self.page_list.push()
info.push()
def increment(self):
self.counter_obj.count += 1
self.counter_obj.push()
def populate(self):
idx = self.block_ctl.wptr.val
total = self.block_ctl.total.val
while idx < total:
block = self.ctx.uobj.new_buf(self.block_size, "BM Block", track=False)
self.block_list[idx * 2] = block._addr // self.page_size
page_idx = idx * self.pages_per_block
for i in range(self.pages_per_block):
self.page_list[page_idx + i] = block._addr // self.page_size + i
idx += 1
self.block_ctl.wptr.val = idx
|
