elfloader changes for shim patching for aie4 #8355

HimanshuChoudhary-Xilinx · 2024-08-16T08:18:27Z

Problem solved by the commit

Host address patching on shimBD for aie4.

Bug / issue (if any) fixed, which PR introduced the bug, how it was discovered

aie4 have different shimbd configuration, Patching is done on 0th and 1st word of BD.

How problem was solved, alternative solutions (if any) and why they were rejected

added new patching schema, shim_dma_aie4_base_addr_symbol_kind

Risks (if any) associated the changes in the commit

What has been tested and how, request additional testing if necessary

Documentation impact (if any)

gbuildx · 2024-08-16T08:20:03Z

Can one of the admins verify this patch?

stsoe

Please fill out the PR description.

xdavidz · 2024-08-16T15:34:24Z

src/runtime_src/core/common/api/xrt_module.cpp

@@ -796,7 +814,7 @@ class module_elf : public module_impl
  ert_cmd_opcode
  get_ert_opcode() const override
  {
-    if (m_os_abi == Elf_Amd_Aie2ps)
+    if (m_os_abi == Elf_Amd_Aie2ps || m_os_abi == Elf_Amd_npu3)


I am not an expert of this area of the code, but it looks to me that there are lots of code like
"if aie2ps or aie4(npu3) then ..." can we make it more clear if you don't want to use more comment to explain what we are doing here.

can we use aie2ps and aie4 instead of aie2ps and npu3, those seems not a pair in same level

can we use a inline function or macro to make those dup code less?

larry9523 · 2024-08-16T15:55:48Z

src/runtime_src/core/common/api/xrt_module.cpp

+        if (m_os_abi == Elf_Amd_npu3)
+          patch57_npu3(bd_data_ptr, new_value + item.offset_to_base_bo_addr);
+        else
+          patch57(bd_data_ptr, new_value + item.offset_to_base_bo_addr);


This OK. But I think differentiating this by a new symbol type is cleaner than os_abi. Conceptly, one symobl type should map to one patching schema

I think @HimanshuChoudhary-Xilinx is trying to avoid adding more patching schema because currently we only use 3bits for encoding schema-type.

I think @HimanshuChoudhary-Xilinx is trying to avoid adding more patching schema because currently we only use 3bits for encoding schema-type.

not 4 bits (0-3)?

If that is the case, we should definitely add a new patching scheme instead of adding a new os_abi. @larry9523

its not a different patching... shimbd in aie2ps and npu3 have different configuration
in aie2ps, host address get patched on 1st,2nd, 8th word of BD while in npu3 its on 0th and 1st word of BD.
its the same 57bit patching

its not a different patching... shimbd in aie2ps and npu3 have different configuration in aie2ps, host address get patched on 1st,2nd, 8th word of BD while in npu3 its on 0th and 1st word of BD. its the same 57bit patching

I would argue shim_dma_48 is not a different patching either. It just patch to second and third of 8 BD WORDs.

OS_ABI is used to determine ELF's architecture. What ELF spec we should follow.

do we have enough symbol type to hold all potential patching method?
if so, then we need something like,
patch57_aie2ps
patch57_aie4
if not, we may need combination of symbol and os_abi, like what Himanshu did.

Yes. We do have enough symbol type to hold all potential patching method for the moment.

A Open question, if os_abi is about ELF spec, then how we differentiate between a elf with aie2ps controlcode and elf with aie4 controlcode because an elf with aie2ps controlcode will not work on aie4 device or vice-versa.
XRT should have a check on this.

A Open question, if os_abi is about ELF spec, then how we differentiate between a elf with aie2ps controlcode and elf with aie4 controlcode because an elf with aie2ps controlcode will not work on aie4 device or vice-versa. XRT should have a check on this.

I am totally fine if we define a different os_abi between aie2ps and aie4 if we need to. I am just saying we should use a new patching schema at this place instead of using os_abi.

dezhiAmd · 2024-08-16T16:14:07Z

Please fill out the PR description.

I keep saying this in almost all PRs from the link below:
https://gitenterprise.xilinx.com/XRT/aiebu/pulls?q=is%3Apr+author%3Achoudhar+

Please, please fill out PR description especially

Problem solved by this commit
How problem is solved

HimanshuChoudhary-Xilinx · 2024-08-28T08:33:35Z

Added new patching schema
Made same os_abi for aie2ps and aie4

xuhz

i think the os_abi is still required, but not to be used to choose patching method
e.g to check an elf compiled with -t aie4 loading to aie4 hw

HimanshuChoudhary-Xilinx · 2024-08-28T18:47:18Z

i think the os_abi is still required, but not to be used to choose patching method e.g to check an elf compiled with -t aie4 loading to aie4 hw

we dont need os_abi because as @larry9523 said it represent elf spec and elf spec for aie2ps and aie4 is same

larry9523 · 2024-08-28T18:47:47Z

i think the os_abi is still required, but not to be used to choose patching method e.g to check an elf compiled with -t aie4 loading to aie4 hw

Let's clean this up by using e_machine

HimanshuChoudhary-Xilinx · 2024-08-28T18:50:00Z

i think the os_abi is still required, but not to be used to choose patching method e.g to check an elf compiled with -t aie4 loading to aie4 hw

Let's clean this up by using e_machine

lets clean that in new pr as it will need aiebu changes also

xuhz

i didn't read the previous comments carefully
we have e_machine to tell aie2ps or aie4
os_abi is not for that purpose.
but elf_amd_aie2ps is not a good name which caused the confusion. It should be changed in future PR with a better name, something like elf_amd_aie_uc

larry9523 · 2024-08-28T21:29:49Z

src/runtime_src/core/common/api/xrt_module.cpp

@@ -115,6 +115,7 @@ struct patcher
    scalar_32bit_kind = 3,
    control_packet_48 = 4,              // patching scheme needed by firmware to patch control packet
    shim_dma_48 = 5,                    // patching scheme needed by firmware to patch instruction buffer
+    shim_dma_aie4_base_addr_symbol_kind = 6, // patching scheme needed by AIE4 firmware


Please update the ELF spce accordingly.

Signed-off-by: HimanshuChoudhary-Xilinx <[email protected]>

HimanshuChoudhary-Xilinx requested a review from stsoe as a code owner August 16, 2024 08:18

stsoe requested changes Aug 16, 2024

View reviewed changes

xdavidz reviewed Aug 16, 2024

View reviewed changes

larry9523 requested a review from dezhiAmd August 16, 2024 15:45

larry9523 reviewed Aug 16, 2024

View reviewed changes

HimanshuChoudhary-Xilinx force-pushed the npu3 branch from 1df4fd7 to daebfea Compare August 28, 2024 08:31

HimanshuChoudhary-Xilinx changed the title ~~elfloader changes for shim patching for npu3~~ elfloader changes for shim patching for aie4 Aug 28, 2024

xuhz reviewed Aug 28, 2024

View reviewed changes

xuhz approved these changes Aug 28, 2024

View reviewed changes

larry9523 reviewed Aug 28, 2024

View reviewed changes

larry9523 approved these changes Aug 28, 2024

View reviewed changes

dezhiAmd approved these changes Aug 28, 2024

View reviewed changes

maxzhen approved these changes Aug 29, 2024

View reviewed changes

HimanshuChoudhary-Xilinx added 2 commits September 3, 2024 22:01

elfloader changes for shim patching for npu3

b17f3af

Signed-off-by: HimanshuChoudhary-Xilinx <[email protected]>

review comments

0d033ba

Signed-off-by: HimanshuChoudhary-Xilinx <[email protected]>

HimanshuChoudhary-Xilinx force-pushed the npu3 branch from daebfea to 0d033ba Compare September 3, 2024 16:32

xdavidz approved these changes Sep 3, 2024

View reviewed changes

maxzhen merged commit ea90b89 into Xilinx:master Sep 4, 2024
17 checks passed

vipangul pushed a commit to vipangul/XRT that referenced this pull request Sep 11, 2024

elfloader changes for shim patching for aie4 (Xilinx#8355)

67d7e75

Signed-off-by: HimanshuChoudhary-Xilinx <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

elfloader changes for shim patching for aie4 #8355

elfloader changes for shim patching for aie4 #8355

HimanshuChoudhary-Xilinx commented Aug 16, 2024 •

edited

Loading

gbuildx commented Aug 16, 2024

stsoe left a comment

xdavidz Aug 16, 2024

larry9523 Aug 16, 2024

dezhiAmd Aug 16, 2024

larry9523 Aug 16, 2024

dezhiAmd Aug 16, 2024

HimanshuChoudhary-Xilinx Aug 19, 2024 •

edited

Loading

larry9523 Aug 19, 2024

xuhz Aug 19, 2024

dezhiAmd Aug 19, 2024

HimanshuChoudhary-Xilinx Aug 20, 2024

larry9523 Aug 20, 2024

dezhiAmd commented Aug 16, 2024

HimanshuChoudhary-Xilinx commented Aug 28, 2024 •

edited

Loading

xuhz left a comment

HimanshuChoudhary-Xilinx commented Aug 28, 2024

larry9523 commented Aug 28, 2024

HimanshuChoudhary-Xilinx commented Aug 28, 2024

xuhz left a comment

larry9523 Aug 28, 2024

elfloader changes for shim patching for aie4 #8355

elfloader changes for shim patching for aie4 #8355

Conversation

HimanshuChoudhary-Xilinx commented Aug 16, 2024 • edited Loading

Problem solved by the commit

Bug / issue (if any) fixed, which PR introduced the bug, how it was discovered

How problem was solved, alternative solutions (if any) and why they were rejected

Risks (if any) associated the changes in the commit

What has been tested and how, request additional testing if necessary

Documentation impact (if any)

gbuildx commented Aug 16, 2024

stsoe left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HimanshuChoudhary-Xilinx Aug 19, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dezhiAmd commented Aug 16, 2024

HimanshuChoudhary-Xilinx commented Aug 28, 2024 • edited Loading

xuhz left a comment

Choose a reason for hiding this comment

HimanshuChoudhary-Xilinx commented Aug 28, 2024

larry9523 commented Aug 28, 2024

HimanshuChoudhary-Xilinx commented Aug 28, 2024

xuhz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HimanshuChoudhary-Xilinx commented Aug 16, 2024 •

edited

Loading

HimanshuChoudhary-Xilinx Aug 19, 2024 •

edited

Loading

HimanshuChoudhary-Xilinx commented Aug 28, 2024 •

edited

Loading