Add first version of mglib library#46
Open
FrancescoConti wants to merge 16 commits into
Open
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
This PR introduces a new lightweight mglib static library that provides event-driven, non-blocking wrapper APIs for iDMA and RedMulE, and integrates it into the build and test system. It also adds per-tile-aliased scratch storage support (.tile_bss) in the linker script and ensures it is zero-initialized at startup for correct per-tile state in multi-tile execution.
Changes:
- Add
mglib(mg_event / mg_idma / mg_redmule) wrappers and build integration. - Add per-tile
.tile_bssreservation in the linker script and clear it incrt0. - Add a new
test_mm_os_mglibtest and integrate it into the mesh test suite.
Reviewed changes
Copilot reviewed 13 out of 14 changed files in this pull request and generated 6 comments.
Show a summary per file
| File | Description |
|---|---|
| CMakeLists.txt | Adds mglib as a top-level subdirectory for the build. |
| mglib/CMakeLists.txt | Defines the mglib static library target, includes, and LTO-related compile flags. |
| mglib/include/mg_event.h | Adds a small generic event record + 8-bit wrapping sequence comparison helper. |
| mglib/include/mg_idma.h | Declares event-driven iDMA memcpy wrappers and wait routine. |
| mglib/include/mg_redmule.h | Declares event-driven RedMulE GEMM wrapper and wait routine. |
| mglib/src/mg_idma.c | Implements per-tile (tile-aliased) iDMA issue/completion bookkeeping and wait logic. |
| mglib/src/mg_redmule.c | Implements RedMulE job issue/wait bookkeeping using per-tile state. |
| targets/magia_v2/link.ld | Pins entrypoint to _start and adds fixed-size .tile_bss reservation in per-tile memory. |
| targets/magia_v2/src/crt0.S | Clears .tile_bss on every hart during startup. |
| tests/magia/mesh/CMakeLists.txt | Adds mm_os_mglib to the mesh test suite. |
| tests/magia/mesh/mm_os_mglib/CMakeLists.txt | Adds the new test_mm_os_mglib executable, LTO link options, and links against mglib. |
| tests/magia/mesh/mm_os_mglib/src/test.c | New matrix-multiply test variant that drives iDMA/RedMulE through mglib. |
| scripts/gvsoc2perfetto-rs/.gitignore | Ignores Rust /target build artifacts. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+81
to
+83
| if (((tile_w_max * x_id) + tile_w_max) > N_SIZE) { | ||
| tile_w = N_SIZE - (tile_w_max * x_id); | ||
| } else { |
Comment on lines
+226
to
+228
| if (i != 0) { | ||
| mg_idma_wait(&eu_ctrl, 1, WAIT_MODE, &idma_evt_y); | ||
| } |
Comment on lines
+291
to
+302
| mg_idma_memcpy_2d(&idma_ctrl, | ||
| &eu_ctrl, | ||
| WAIT_MODE, | ||
| 1, | ||
| axi_addr_y, | ||
| obi_addr_y, | ||
| len_y, | ||
| std_y, | ||
| reps_y, | ||
| &idma_evt_y, | ||
| NULL); | ||
| } |
| // consumed while waiting on a later id - in which case we must not | ||
| // wait on the hardware at all. | ||
| while (!mg_seq_ge(mg_idma_completed[idx], target)) { | ||
| // dnot yet the right one: spin back into the hardware wait. |
Comment on lines
+24
to
+48
| static uint8_t mg_redmule_completed __attribute__((section(".tile_bss"))); | ||
|
|
||
| void mg_redmule_gemm(redmule_controller_t *ctrl, | ||
| eu_controller_t *eu, | ||
| eu_wait_mode_t mode, | ||
| uint32_t x, | ||
| uint32_t w, | ||
| uint32_t y, | ||
| uint16_t m, | ||
| uint16_t n, | ||
| uint16_t k, | ||
| mg_event_t *event, | ||
| mg_event_callback_t callback) | ||
| { | ||
| int32_t id; | ||
| while ((id = redmule_acquire(ctrl)) < 0) { | ||
| // Hardware queue (depth 2) is full: drain one completion pulse to | ||
| // free a slot before retrying the acquire. | ||
| if (eu_redmule_wait(eu, mode)) { | ||
| mg_redmule_completed++; | ||
| } | ||
| } | ||
| mg_event_init(event, id, callback); | ||
| redmule_gemm(ctrl, x, w, y, m, n, k); | ||
| } |
Comment on lines
+18
to
+22
| * The event id is the real hardware-issued job id, obtained by reading the | ||
| * ACQUIRE register (redmule_api.acquire()). If the hardware's job queue | ||
| * (depth 2) is already full, this call blocks - draining completion pulses | ||
| * through `eu`/`mode`, exactly like mg_redmule_wait() does - until a slot | ||
| * frees and a fresh id can be acquired. |
FrancescoConti
force-pushed
the
fc/mglib-rebase
branch
from
d7854ce to
11dcb35
Compare
FrancescoConti
force-pushed
the
fc/mglib-rebase
branch
from
11dcb35 to
cba1088
Compare
Member
Author
|
RTL breaks because the |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This pull request introduces a new lightweight library,
mglib, providing thin, event-driven wrappers for iDMA and RedMulE, and integrates it into the build and test system. It also adds robust per-tile memory allocation and initialization support in the linker script and startup code, ensuring correct operation in multi-tile environments. Additionally, a new test is added to validate the new functionality.These changes provide a foundation for event-driven accelerator programming with per-tile state, targeting high utilization.
Compared to the HAL-based
test_mm_os_2, which was used as a reference, themglib-driventest_mm_os_mglibprovides a significant performance improvement by reducing the amount of time the tile spends without computing with RedMulE.With a sufficiently large workload,
mglibenables full RedMulE utilization, e.g. (test_mm_os_mglibwithM=N=K=256, 8 timeslots, 2x2 mesh size):Note: correct GVSOC behavior relies on this PR for
gvsoc-pulp: gvsoc/gvsoc-pulp#91 , which has been merged in upstreamgvsocin gvsoc/gvsoc#276Library and Event-Driven Accelerator Wrappers
mglibstatic library, which provides event-based, non-blocking wrappers for iDMA and RedMulE accelerators, including per-tile completion bookkeeping and callback support. (CMakeLists.txt,mglib/CMakeLists.txt,mglib/include/mg_event.h,mglib/include/mg_idma.h,mglib/include/mg_redmule.h,mglib/src/mg_idma.c,mglib/src/mg_redmule.c) [1] [2] [3] [4] [5] [6] [7]Per-Tile Memory Management
.tile_bsssection for per-tile-aliased scratch memory, reserved above the stack, and ensured it is zeroed by every hart in the startup code. This supports safe per-tile static data for the new library. (targets/magia_v2/link.ld,targets/magia_v2/src/crt0.S) [1] [2] [3]Testing and Build Integration
test_mm_os_mglib, undertests/magia/mesh, which exercises the new mglib wrappers and links against the new library with LTO and appropriate options for symbol resolution. (tests/magia/mesh/CMakeLists.txt,tests/magia/mesh/mm_os_mglib/CMakeLists.txt) [1] [2]Build and Tooling Quality-of-Life
/targetto the.gitignorefor the Rust-basedgvsoc2perfettotool to prevent build artifacts from being committed. (scripts/gvsoc2perfetto-rs/.gitignore)Other Linker Improvements
_startin the linker script to ensure correct boot behavior when using LTO. (targets/magia_v2/link.ld)Limitations
idmamodel currently used in GVSOC does not support any kind of hardware job queue. As a consequence, queueing is implemented entirely in software, which means that DMA jobs cannot be offloaded back-to-back.mglibAPI is in draft state, and there are likely some misunderstandings in how to divide between thehal,drivers, andmgliblayers.