构建ROCm,有对这个话题感兴趣的吗？

moosewoler

龙芯3A5000可用RX580显卡，理论上可在上面运行ROCm框架。
目前参考xuhuisheng公布于github上的rocm-build项目，一共有42个脚本。

moosewoler

目前进度：

12.roct-thunk-interface.sh 通过
13.rocm-cmake.sh 通过
14.rocm-device-libs.sh 通过
15.rocr-runtime.sh 通过
16.rocminfo.sh 通过
17.rocm-compilersupport.sh 通过
18.hip.sh 错误 Failed to embed PCH

Xinmudotmoe

近况如何

我经过一些暴力魔改构建，使得rocm-5.6.x分支完成了clr(对应旧版本的18.hip)，在rocFFT(编号21前进了，但只前进了一点儿)构建时发生错误。

从结果看我认为可能出在rocm/llvm-project/clang上面。

FAILED: library/src/CMakeFiles/rocfft-rtc-launch.dir/rtc_kernel.cpp.o
/opt/rocm-5.6.x//bin/hipcc -I/opt/rocm-5.6.x-build/src/rocFFT/library/src/include -I/opt/rocm-5.6.x-build/src/rocFFT/library/include -I/opt/rocm-5.6.x-build
/build/rocFFT/_deps/sqlite_local-src -O3 -DNDEBUG -std=gnu++17 -fPIC -fvisibility=hidden -fvisibility-inlines-hidden -Wall -Wno-unused-function -Wimplicit-fa
llthrough -Wunreachable-code -Wsign-compare -DROCFFT_RUNTIME_COMPILE -MD -MT library/src/CMakeFiles/rocfft-rtc-launch.dir/rtc_kernel.cpp.o -MF library/src/CM
akeFiles/rocfft-rtc-launch.dir/rtc_kernel.cpp.o.d -o library/src/CMakeFiles/rocfft-rtc-launch.dir/rtc_kernel.cpp.o -c /opt/rocm-5.6.x-build/src/rocFFT/librar
y/src/rtc_kernel.cpp
fatal error: error in backend: Cannot select: t63: f32 = fp16_to_fp t57
t57: i64 = fp_to_fp16 t2
t2: f64,ch = CopyFromReg t0, Register:f64 %1
t1: f64 = Register %1
In function: _ZN9RTCKernel6launchER12DeviceCallIn

Xinmudotmoe

新状态

目前已经打包出来部分子项目，详情参见。

http://nas.xinmu.moe:8086/xinmu/loongarch64-rocm-arch/README.html

如果自己的性能比较强悍，可以选择从github中获取构建脚本。

构建脚本项目 https://github.com/loongarch-moe/rocm-loongarch 。

大概只完成了不到一半的构建移植，移植前景神秘莫测。

Xinmudotmoe

有了一些新东西需要改，以下的小补丁可以去掉构建amdkfd驱动限制。

目前执行大多数rocm指令，均会有内核级报错。贴在最下方。

diff --git a/drivers/gpu/drm/amd/amdkfd/Kconfig b/drivers/gpu/drm/amd/amdkfd/Kconfig
index d3c3d3ab7225..1b8ebc97e0dc 100644
--- a/drivers/gpu/drm/amd/amdkfd/Kconfig
+++ b/drivers/gpu/drm/amd/amdkfd/Kconfig
@@ -5,7 +5,7 @@
 
 config HSA_AMD
        bool "HSA kernel driver for AMD GPU devices"
-       depends on DRM_AMDGPU && (X86_64 || ARM64 || PPC64)
+       depends on DRM_AMDGPU && (X86_64 || ARM64 || PPC64 || LOONGARCH)
        select HMM_MIRROR
        select MMU_NOTIFIER
        select DRM_AMDGPU_USERPTR

[ 36.738846] ------------[ cut here ]------------

[ 36.738870] WARNING: CPU: 3 PID: 30 at drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c:615 amdgpu_irq_put+0x64/0x80 [amdgpu]

[ 36.739445] Modules linked in: amdgpu xt_CHECKSUM xt_MASQUERADE xt_conntrack ipt_REJECT nf_reject_ipv4 dm_thin_pool dm_persistent_data dm_bio_prison dm_bufio ip6table_mangle ip6table_nat iptable_mangle iptable_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 iptable_filter ip_tables bpfilter bridge nf_tables nfnetlink 8021q garp stp mrp llc binfmt_misc cfg80211 rfkill acpi_ipmi amdxcp drm_exec gpu_sched uas drm_buddy drm_suballoc_helper drm_ttm_helper spi_loongson_pci spi_loongson_core drm_display_helper ipmi_si usb_storage uio_pdrv_genirq uio ipmi_devintf ipmi_msghandler efi_pstore pstore efivarfs

[ 36.739528] CPU: 3 PID: 30 Comm: kworker/3:0 Not tainted 6.6.0-rc4+ #11

[ 36.739532] Hardware name: To be filled by O.E.M.To be fill To be filled by O.E.M.To be fill/To be filled by O.E.M.To be fill, BIOS Loongson UEFI V4.0.5-s

[ 36.739536] Workqueue: pm pm_runtime_work

[ 36.739547] pc ffff800002834ba4 ra ffff8000029f75e4 tp 9000000100494000 sp 9000000100497b60

[ 36.739551] a0 9000000133000000 a1 900000012f704808 a2 0000000000000000 a3 0000000000000000

[ 36.739554] a4 9000000133016968 a5 00000000007bae52 a6 00000000007bae91 a7 0000000000000000

[ 36.739557] t0 0000000000000000 t1 900000010af0d530 t2 0000000000000000 t3 9000000133015000

[ 36.739560] t4 346240faa478c355 t5 0000000020000000 t6 ffff800000000000 t7 0000800000000000

[ 36.739563] t8 000000000000002f u0 9000000003082cec s9 90000001000ffe60 s0 900000012f704800

[ 36.739566] s1 900000012f704800 s2 ffff800002ecc000 s3 9000000133000000 s4 ffff800003027cc0

[ 36.739569] s5 9000000133000000 s6 9000000133041530 s7 9000000133046000 s8 9000000100141040

[ 36.739572] ra: ffff8000029f75e4 smu_smc_hw_cleanup+0x44/0x3c0 [amdgpu]

[ 36.739961] ERA: ffff800002834ba4 amdgpu_irq_put+0x64/0x80 [amdgpu]

[ 36.740343] CRMD: 000000b0 (PLV0 -IE -DA +PG DACF=CC DACM=CC -WE)

[ 36.740356] PRMD: 00000004 (PPLV0 +PIE -PWE)

[ 36.740362] EUEN: 00000000 (-FPE -SXE -ASXE -BTE)

[ 36.740370] ECFG: 00071c1c (LIE=2-4,10-12 VS=7)

[ 36.740376] ESTAT: 000c0000 [BRK] (IS= ECode=12 EsubCode=0)

[ 36.740382] PRID: 0014c011 (Loongson-64bit, Loongson-3C5000)

[ 36.740386] CPU: 3 PID: 30 Comm: kworker/3:0 Not tainted 6.6.0-rc4+ #11

[ 36.740389] Hardware name: To be filled by O.E.M.To be fill To be filled by O.E.M.To be fill/To be filled by O.E.M.To be fill, BIOS Loongson UEFI V4.0.5-s

[ 36.740392] Workqueue: pm pm_runtime_work

[ 36.740396] Stack : 900000000496e000 9000000004083cf4 900000000409f1b0 9000000100494000

[ 36.740405] 90000001004977e0 0000000000000000 90000001004977e8 900000000480b2a0

[ 36.740412] 9000000004a57d78 9000000004a57d70 90000001004976a0 0000000000000001

[ 36.740420] 0000000000000001 346240faa478c355 00000000064d0000 900000010044b2c0

[ 36.740427] 0000000000000000 0000000000000030 0000000000000003 0000000000000003

[ 36.740435] 0000000000001ce8 00000000000b4c28 00000000064d0000 90000001000ffe60

[ 36.740442] 900000000496e000 900000000480b2a0 0000000000000004 0000000000000000

[ 36.740450] 0000000000000001 0000000000000000 00000000003962ef 9000000133046000

[ 36.740457] 9000000100141040 0000000000000000 90000000030335d4 00007fffdbec8000

[ 36.740465] 00000000000000b0 0000000000000004 0000000000000000 0000000000071c1c

[ 36.740472] ...

[ 36.740476] Call Trace:

[ 36.740479] [<90000000030335d4>] show_stack+0x34/0x160

[ 36.740488] [<900000000409f1b0>] dump_stack_lvl+0x60/0x88

[ 36.740493] [<9000000004082310>] __warn+0x84/0xc8

[ 36.740498] [<9000000004051900>] report_bug+0x1a0/0x240

[ 36.740504] [<90000000040a04d8>] do_bp+0x278/0x2e0

[ 36.740508] [<9000000100511924>] 0x9000000100511924

[ 36.740512] [<ffff800002834ba4>] amdgpu_irq_put+0x64/0x80 [amdgpu]

[ 36.741084] [<ffff8000029f75e4>] smu_smc_hw_cleanup+0x44/0x3c0 [amdgpu]

[ 36.741663] [<ffff8000029f9504>] smu_suspend+0xa4/0x120 [amdgpu]

[ 36.742243] [<ffff80000274da40>] amdgpu_device_ip_suspend_phase2+0x140/0x280 [amdgpu]

[ 36.742808] [<ffff800002750fe4>] amdgpu_device_suspend+0x144/0x200 [amdgpu]

[ 36.743373] [<ffff80000274d26c>] amdgpu_pmops_runtime_suspend+0x10c/0x1e0 [amdgpu]

[ 36.743937] [<9000000003886394>] pci_pm_runtime_suspend+0x54/0x220

[ 36.743945] [<9000000003a54370>] __rpm_callback+0x50/0x240

[ 36.743948] [<9000000003a545ec>] rpm_callback+0x8c/0xa0

[ 36.743952] [<9000000003a54df8>] rpm_suspend+0x118/0x5c0

[ 36.743955] [<9000000003a55eec>] pm_runtime_work+0xcc/0x100

[ 36.743959] [<9000000003080f4c>] process_one_work+0x14c/0x2a0

[ 36.743965] [<90000000030813ec>] worker_thread+0x34c/0x480

[ 36.743970] [<900000000308bfc4>] kthread+0x124/0x140

[ 36.743975] [<9000000003031308>] ret_from_kernel_thread+0xc/0xa4

[ 36.743991] ---[ end trace 0000000000000000 ]---

[ 36.743997] amdgpu 0000:0a:00.0: amdgpu: Fail to disable thermal alert!

[ 36.744002] [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <smu> failed -22

Robin Lu

（可能有一些 off-topic ，但现在 Radeon VII / Radeon Instinct MI50 并不很贵的情况下其实推荐以其作为 target ，HiFive Unmatched 上能走到很后面 https://zhuanlan.zhihu.com/p/621314263

可以出 rocm-smi 的结果，但是 bandwidth test会seg fault，rocclr似乎会assertion failed，https://bugs.gentoo.org/895286 和 Gentoo 的这个 issue 状况一样

Xinmudotmoe

rocm-bandwidth-test 构建成功并成功运行(但得分远低于正常值)

大部分项目不会触发seg fault。但PCIE Atomics缺失，相当一部分示例项目仅通过编译，运行会报错(不是seg fault)。

yetist

loongarchlinux上面，缺了这一大块包。

Xinmudotmoe

ROCm移植进度

https://www.bilibili.com/video/BV1az421y75z/

已可以跑pytorch(rocm)

https://github.com/loongarch-moe/rocm-loongarch/tree/rocm-5.7.x

factfinding

Xinmudotmoe 您好，请问3a6000不是最高只支持rx580吗，请问您是怎么驱动成功rx 6800的，我最近想给3a6000添加一张独立显卡，请问有兼容性列表吗？

Xinmudotmoe

factfinding

参见 https://github.com/loongson/Firmware/pull/79#issuecomment-1831702740

https://github.com/loongson/Firmware/issues/80#issuecomment-1903111698

本站文章除其作者特殊声明外，一律采用CC BY-NC-SA 4.0许可协议进行授权。
进行转载或二次创作时务必以相同协议进行共享，严禁用于商业用途。