filmov
tv
George Hotz | Reverse engineering | switching from AQL to PM4 to avoid bugs on the AMD 7900XTX Part2
Показать описание
Date of the stream 6 Apr 2024.
Live-stream chat added as Subtitles/CC - English (Twitch Chat) - at the bottom - Show Transcript
Source:
Follow for notifications:
Support George:
Pre-order tinybox:
Chapters:
00:00:00 intro
00:01:50 meditation
00:04:10 questions, AQL packets, radeon southern islands
00:06:00 two code paths
00:10:10 amd vulkan, dispatch_indirect
00:18:30 packet queue = AQL packets, amdgpu eop
00:22:10 AQL_CONTROL
00:30:55 blog part 3
00:37:00 resnet training job to crash
00:40:00 DEBUG=2
00:43:00 getting rid of AQL
00:45:30 engineer wrote something specific and project manager changed the words
00:47:10 selected pipe, queue does not matter
00:48:10 different queues
00:50:15 PM4, PGM_RSRC2
00:56:10 EVENT_WRITE.partial_flush comment
00:58:40 more testing of KFD, moving away from AQL
01:05:30 does someone want's to write PM4?
01:09:45 when I make the GPU
01:10:20 nvidia and amd very similar
01:10:50 firmware absolutely simple as possible, how easy is updating different tier of software
01:11:20 why AQL exists?
01:19:45 we got the crash
01:29:00 double banned, umr, smc bits read_smc
01:33:00 30% chance PM4 does not have bugs
01:33:40 rosetta stone style translation F32 and RS64
01:38:10 radeon_tools marcan
01:38:20 is AMD going to open source anything useful
01:40:40 ghost locomotion, argo ai
01:42:40 argo the next thing, can everyone stop, cleaning solution
01:44:30 can we all just be more honest
01:45:35 AMD NDA, infringing patents,
01:52:15 being afraid of getting sued
01:55:30 DISPATCH_DIRECT packet
02:00:40 amd gpus codenames, fiji
02:02:55 linux-firmware license, not a lawyer
02:06:10 nvidia gsp license, cuda license no datacenter use
02:07:55 risk averseness of people
02:10:40 linux-firmware
02:20:40 rosetta stone, F32, RS64
02:37:00 launching specific kernel
02:40:40 first GPU to support AQL
02:46:50 NSA, Ghidra, IDA Pro, Russian tools
02:48:30 respect NSA github
02:48:50 NSA better at open source then AMD
02:49:14 NSA more open source than AMD
02:49:30 backdoors
02:54:20 drake rosetta stone song
02:57:30 credit marcan for writing the disassembler
03:15:45 binvis
03:19:30 fail0verflow radeon talk
03:39:25 F32
03:48:30 gfx1101
03:53:20 amd rs64
04:10:15 gfx rs64
04:17:55 python os lseek
04:23:15 amdgpu_regs, mmio
04:34:35 decompiling now
04:38:50 all the different addresses
04:42:30 dump_ip.c, upset writing C
04:48:25 c example mmap fd
04:51:18 read without advancing file pointer
05:03:20 bugs in the dispatch
05:20:25 frequency analysis on code
05:31:45 next 4 bytes are register
05:47:00 amd_emu_mode
05:57:35 we broke the gpu, what did we learn today?
06:00:05 I have given you the tools you need to solve the problem
06:05:40 switch to nvidia if you have AMD crashes
06:09:05 risc-v opcode groups
06:12:30 capstone risc-v disassembler
06:18:20 skip instructions, CS_ARCH_RISCV mode
06:22:20 bad at ghidra
06:23:30 don't spoil the risc-v
06:33:00 if labeled easier to read, launch instructions, almost just compile
06:33:55 figuring out the first instruction
06:38:30 load this register map
06:52:00 ghidra load map file
07:35:20 jump table that goes there, some CP function
07:54:25 cmd 3 packets pm4
08:05:00 it's cool that is risc-v
08:23:40 now we can know if AMD open sources anything real
08:29:50 I bet this did not support AQL
08:33:50 we need to stup
08:42:45 I have found why there is so many bugs in this
08:49:15 back in the days of bonaire
08:56:50 HSA was a marketing campaign
08:57:30 internet issues
Official George Hotz communication channels:
Follow for notifications:
Thank you for reading and using the SHOW MORE button.
We hope you enjoy watching George's videos as much as we do.
See you at the next video.
Live-stream chat added as Subtitles/CC - English (Twitch Chat) - at the bottom - Show Transcript
Source:
Follow for notifications:
Support George:
Pre-order tinybox:
Chapters:
00:00:00 intro
00:01:50 meditation
00:04:10 questions, AQL packets, radeon southern islands
00:06:00 two code paths
00:10:10 amd vulkan, dispatch_indirect
00:18:30 packet queue = AQL packets, amdgpu eop
00:22:10 AQL_CONTROL
00:30:55 blog part 3
00:37:00 resnet training job to crash
00:40:00 DEBUG=2
00:43:00 getting rid of AQL
00:45:30 engineer wrote something specific and project manager changed the words
00:47:10 selected pipe, queue does not matter
00:48:10 different queues
00:50:15 PM4, PGM_RSRC2
00:56:10 EVENT_WRITE.partial_flush comment
00:58:40 more testing of KFD, moving away from AQL
01:05:30 does someone want's to write PM4?
01:09:45 when I make the GPU
01:10:20 nvidia and amd very similar
01:10:50 firmware absolutely simple as possible, how easy is updating different tier of software
01:11:20 why AQL exists?
01:19:45 we got the crash
01:29:00 double banned, umr, smc bits read_smc
01:33:00 30% chance PM4 does not have bugs
01:33:40 rosetta stone style translation F32 and RS64
01:38:10 radeon_tools marcan
01:38:20 is AMD going to open source anything useful
01:40:40 ghost locomotion, argo ai
01:42:40 argo the next thing, can everyone stop, cleaning solution
01:44:30 can we all just be more honest
01:45:35 AMD NDA, infringing patents,
01:52:15 being afraid of getting sued
01:55:30 DISPATCH_DIRECT packet
02:00:40 amd gpus codenames, fiji
02:02:55 linux-firmware license, not a lawyer
02:06:10 nvidia gsp license, cuda license no datacenter use
02:07:55 risk averseness of people
02:10:40 linux-firmware
02:20:40 rosetta stone, F32, RS64
02:37:00 launching specific kernel
02:40:40 first GPU to support AQL
02:46:50 NSA, Ghidra, IDA Pro, Russian tools
02:48:30 respect NSA github
02:48:50 NSA better at open source then AMD
02:49:14 NSA more open source than AMD
02:49:30 backdoors
02:54:20 drake rosetta stone song
02:57:30 credit marcan for writing the disassembler
03:15:45 binvis
03:19:30 fail0verflow radeon talk
03:39:25 F32
03:48:30 gfx1101
03:53:20 amd rs64
04:10:15 gfx rs64
04:17:55 python os lseek
04:23:15 amdgpu_regs, mmio
04:34:35 decompiling now
04:38:50 all the different addresses
04:42:30 dump_ip.c, upset writing C
04:48:25 c example mmap fd
04:51:18 read without advancing file pointer
05:03:20 bugs in the dispatch
05:20:25 frequency analysis on code
05:31:45 next 4 bytes are register
05:47:00 amd_emu_mode
05:57:35 we broke the gpu, what did we learn today?
06:00:05 I have given you the tools you need to solve the problem
06:05:40 switch to nvidia if you have AMD crashes
06:09:05 risc-v opcode groups
06:12:30 capstone risc-v disassembler
06:18:20 skip instructions, CS_ARCH_RISCV mode
06:22:20 bad at ghidra
06:23:30 don't spoil the risc-v
06:33:00 if labeled easier to read, launch instructions, almost just compile
06:33:55 figuring out the first instruction
06:38:30 load this register map
06:52:00 ghidra load map file
07:35:20 jump table that goes there, some CP function
07:54:25 cmd 3 packets pm4
08:05:00 it's cool that is risc-v
08:23:40 now we can know if AMD open sources anything real
08:29:50 I bet this did not support AQL
08:33:50 we need to stup
08:42:45 I have found why there is so many bugs in this
08:49:15 back in the days of bonaire
08:56:50 HSA was a marketing campaign
08:57:30 internet issues
Official George Hotz communication channels:
Follow for notifications:
Thank you for reading and using the SHOW MORE button.
We hope you enjoy watching George's videos as much as we do.
See you at the next video.
Комментарии