William #267
William #267
Conversation
This change adds hardware support to the stdlib's RISCV board to simulate NUMA memory ranges. Change-Id: If3e371dcbd51b6eb3903d42809ae6579aef1b163 Signed-off-by: Kaustav Goswami <[email protected]>
Signed-off-by: Kaustav Goswami <[email protected]>
Signed-off-by: Kaustav Goswami <[email protected]>
… test config in same directory
There was a problem hiding this comment.
Ok, thanks for putting this together. I have lots of small comments. Please take the time to learn gem5's style guide and take the time to write code that is up to our standards. Use stack overflow and github copilot!! to find the "right way" to write code.
For the actual problem that you're facing... the code is missing some pieces. I can try to find some time to sit down with you and work on this. Maybe tomorrow afternoon?
src/mem/DRAMInterface.py
Outdated
| @@ -1806,3 +1806,101 @@ class LPDDR5_6400_1x16_8B_BL32(LPDDR5_6400_1x16_BG_BL32): | |||
| tCCD_L = "0ns" | |||
| tRRD_L = "0ns" | |||
| tWTR_L = "0ns" | |||
|
|
|||
|
|
|||
| class LLM(DRAMInterface): | |||
There was a problem hiding this comment.
Can you drop LLM changes from this PR?
| type = 'MessageQueue' | ||
| cxx_header = "mem/message_queue.hh" | ||
| cxx_class = "gem5::MessageQueue" | ||
| # can either have a vector of ports connecting to the cores, or have a vector of write ports and 1 read port |
There was a problem hiding this comment.
make sure you install pre-commit and follow the gem5 style guide (in this case lines should be < 80 chars)
src/mem/message_queue.hh
Outdated
|
|
||
| }; | ||
|
|
||
| std::deque<std::tuple<uint32_t, uint32_t, uint32_t, Tick>> queue; //Could be address or vertexID, vertexID might be easier |
There was a problem hiding this comment.
This seems like a very complex datastructure. Maybe it should be a map? A deque of a complex tuple doesn't seem like the right solution. In other words, this is a bad code smell :)
src/mem/message_queue.hh
Outdated
| std::deque<std::tuple<uint32_t, uint32_t, uint32_t, Tick>> queue; //Could be address or vertexID, vertexID might be easier | ||
|
|
||
| uint32_t queueSize; | ||
| AddrRange my_range; |
There was a problem hiding this comment.
Please follow the gem5 coding style. In this case, you need to use camelCase
| AddrRange my_range; | |
| AddrRange myRange; |
| uint32_t queueSize; | ||
| AddrRange my_range; | ||
| // std::vector<RespPort> corePorts; | ||
| RespPort corePorts;//= new RespPort(); |
There was a problem hiding this comment.
This can't be new since the RespPort is not a pointer.
| for chnl in range(num_chnls): | ||
| # size = addr_range.size() | ||
| interface = HBM_1000_4H_1x128() | ||
| interface.range = AddrRange(addr_range.start, size = addr_range.size(), | ||
| intlvHighBit = intlv_low_bit + intlv_bits - 1, | ||
| xorHighBit = 0, | ||
| intlvBits = intlv_bits, | ||
| intlvMatch = chnl) | ||
| ctrl = MemCtrl() | ||
| ctrl.dram = interface | ||
|
|
||
| #ctrl.dram.null = True | ||
| #ctrl.dram.addr_mapping = addr_map | ||
| #ctrl.dram.page_policy = page_policy | ||
| mem_ctrls.append(ctrl) |
There was a problem hiding this comment.
You should simplify this for now. You should simply have one memory channel.
| thispath = os.path.dirname(os.path.realpath(__file__)) | ||
|
|
||
| binpath = os.path.join(thispath, '../../', | ||
| 'tests/test-progs/hello/bin/x86/linux/mapped_queue_fixed_private') |
There was a problem hiding this comment.
This doesn't have to be in test-progs, and you should avoid checking in binaries.
| process.cmd = [binpath] | ||
| # Set the cpu to use the process as its workload and create thread contexts | ||
| system.cpu.workload = process | ||
| # system.cpu.workload.map(vaddr=0x1000000, paddr=0x3CF9BDC0, size=4096) |
There was a problem hiding this comment.
This paddr (0x3CF9BDC0) is what you need to set your addr_range for the message queue.
|
|
||
| # process.map(vaddr=0x10000, paddr=0x3CF9BDC0, size=4096, cacheable=False) | ||
|
|
||
| process.map(vaddr=0x1000000, paddr=0xFF00000, size=4096, cacheable=False) |
There was a problem hiding this comment.
Or maybe the above comment should be here?
There was a problem hiding this comment.
I set the Addr Range of Message Queue to be 255MB - 256MB which I believe 255 MB is 0xFF00000, unless I calculated that wrong. I know the size doesn't really match up to 1 MB but I figured the pointer should be pointing to 255 MB.
|
|
||
| print("Beginning simulation!") | ||
| exit_event = m5.simulate() | ||
| print('Exiting @ tick %i because %s' % (m5.curTick(), exit_event.getCause())) |
There was a problem hiding this comment.
Don't use % for formatting. Use f-strings
Another link: https://www.geeksforgeeks.org/formatted-string-literals-f-strings-python/#
Signed-off-by: Jason Lowe-Power <[email protected]>
Change-Id: Iec495f9997c165c211248148dd959d119e042298 Signed-off-by: Jason Lowe-Power <[email protected]>
Though in "ext" this directory is regularly modified. `pre-commit` should run on these files. This PR includes running `pre-commit run --files ext/testlib` to reformat the files in "ext/testlib" using Python Black.
For Jason to review, MessageQueue changes are in message_queue.cc, .hh, Msg_Map_process_test.py is my test config, mapped_queue.cpp is what was used to create the binary I am running in the config