[DRAFT] Dynamic allocation scheme by bowwwang · Pull Request #126 · pulp-platform/mempool

bowwwang · 2025-07-01T12:43:37Z

This PR enabled TeraPool with a dynamic allocation scheme with support for the DMA engine.

TODO: Adapt to the latest system CSR
FIX: the maximum number of memory rows per DAS scrambling is fixed to the number of tiles in the system.
FIX: the allocation can only start at specific addresses in memory. The only requirement should be that the address is aligned to the memory row. This has to be changed in the address_scrambler and in the software for allocation.
FIX: keep only the fastest DMA transfer option.
FIX: keep only the necessary defines in the mempool_pkg.sv.
CHANGE: change the naming of DAS registers ex. tiles_das, start_das, rows_das, and make it consistent across all the files.
VERIFY: that all the changes do not conflict with the parametrization on main.

Changelog

Added

Changed

address_scrambler.sv for dynamic address mapping
idma_midend_distributor.sv and idma_midend_splitter.sv for augmented transaction split
mempool_tile.sv, mempool_sub_group.sv, mempool_group.sv, and mempool_cluster.sv for passing scrambling configuration

Fixed

(Reference to issues, labels, and related merge requests)

Checklist

Automated tests pass
Changelog updated
Code style guideline is observed

Please check our contributing guidelines before opening a Pull Request.

[hardware] Correct address scrambler parametrization

[software] Add DMA with DAS test

… naming

mbertuletti

Thank you for your work,

Please keep naming scheme consistent between hardware and software. Add explanatory comments where needed to clarify the use of some signals. In particular the group_factor was replaced by tiles_das.
Please remove unused or commented out code blocks and format the .sv files.
Please verify the critical path of the design.

mbertuletti · 2026-04-14T11:59:16Z

 #define CONTROL_REGISTERS_PARAM_MAX_NUMGROUPS 8

+// Supported number of DAS partitions
+#define CONTROL_REGISTERS_PARAM_NUM_D_A_S_PARTITIONS 4


Please use consistent naming, you use "DAS" for other defines.

Register naming has been changed from D_A_S to DAS

mbertuletti · 2026-04-14T12:02:17Z

 // Wake Up Offst Register
 #define CONTROL_REGISTERS_WAKE_UP_OFFST_REG_OFFSET 0x30

+// Tile grouping for DAS partition (common parameters)


These registers shift the address of other previously existing registers. Couldn't you place them at address 0x64 and following?

I placed DAS related registers to the end of the existing regs

mbertuletti · 2026-04-14T12:05:05Z

 // Wake Up Offst Register
 #define CONTROL_REGISTERS_WAKE_UP_OFFST_REG_OFFSET 0x30

+// Tile grouping for DAS partition (common parameters)


Please add a Comment on what the DAS addresses are used for. The difference between the "TILES_DAS" and "START_DAS" registers is not clear.

Desc for the DAS related regs has been addded

mbertuletti · 2026-04-14T12:10:07Z

 ## Unreleased

 ### Added
+- Add Dynamic Address Scrambling (DAS) support with configurable partitioning, hardware address scrambler, and software runtime for dynamic heap allocation


Please move this to the bottom of the list.

Moved to the end of the list

mbertuletti · 2026-04-14T12:13:43Z

 [![ci](https://github.com/pulp-platform/mempool/actions/workflows/ci.yml/badge.svg)](https://github.com/pulp-platform/mempool/actions/workflows/ci.yml)
 [![lint](https://github.com/pulp-platform/mempool/actions/workflows/lint.yml/badge.svg)](https://github.com/pulp-platform/mempool/actions/workflows/lint.yml)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
+# MemPool Dynamic Allocation Scheme


Please add this to the bottom of ./config/README.md and provide more information about how the DAS registers should be configured.

This part has been moved to config/README.md

mbertuletti · 2026-04-14T13:23:39Z

          .tcdm_slave_resp_valid_o (tcdm_slave_resp_valid[g]                                        ),
          .tcdm_slave_resp_ready_i (tcdm_slave_resp_ready[g]                                        ),
+`ifdef DAS
+          .tiles_das_i         (tiles_das_i                                                 ),


FOrmat misaligned parentheses.

mbertuletti · 2026-04-14T13:26:04Z

Please confirm that this solution does not compromise the timing of the cores to L1 connection.

I narrowed the add/sub datapath from 32-bit to 8-bit.
I will schedule a backend run asap

mbertuletti · 2026-04-14T13:26:58Z

Can the address scrambler of the mempool/hw/src/ be reused? In this case the naming of the input signals is possibly wrong and the group_factor_i should be replaced by the more intuitive das_tile_i.

Good catch, i have aligned the naming accordingly.
However, the scrambler in DMA is not the same as the one in the Tile, because DAS does not consider the logic for the stack region. Therefore, the logic related to the stack region has been stripped.
I would keep these two module apart.

mbertuletti · 2026-04-14T13:35:11Z

+    end
+  end
+
+  // ------ Considering Partition Scheme ------ //


Can you add a comment on what shift_index, partition_mask, masked_start_addr, shift_index_sc, shift_row, shift_partition, masked_shift_row are used for? The naming is not very clear and should be possibly improved.

You are very right, it was very confusing. i hope now it is clearer

mbertuletti · 2026-04-14T13:38:44Z

+  logic [AddrWidth-1:0]      partition_mask;
+  addr_t                     masked_start_addr;
+
+  always_comb begin


Similar case blocks were replaced by lzcs in the address scramblers. I believe such implementation and a consistent name replacement for group_factor would improve clarity.

Thank you for the advice, this has been adapted

…S naming

…tions

…gnals with proper comments

bowwwang · 2026-04-17T13:27:51Z

+          // L2 --> L1
+          if (burst_req_i.num_bytes<=DmaRegionWidth )begin
+            burst_req_o[i].src = burst_req_i.src+i*rows_das_i*DmaRegionWidth;
+          end else if (i==2) begin


need to change

mbertuletti force-pushed the asap25-release/das branch 5 times, most recently from 08a9bb1 to 6eef71c Compare October 23, 2025 09:11

bowwang and others added 22 commits October 23, 2025 11:40

[das] init open source

fb21be6

[config] add terapool-das configurations

1989369

[traffic generator] bandwidth benchmark in selected number of tiles

2b5cc29

[hw] add DAS support in idma

320854a

[hw] add DAS control logic to idma

afad72f

[hw] extend scrambler for das

4087bf4

[hw] add das to terapool

6cfabd8

[sw] add das runtime support

bd0386b

[hardware] Add DAS registers and keep only one DMA transfer option

6ef5845

[hardware] Correct address scrambler copy-pasted code

9199936

[config] Parametrize feature with meaningful define

c155c4d

[software] Rename "group_factor" with meaningful tile grouping

a92a71f

[software] Add partition test

833ae78

[hardware] Group DAS registers and assign external reset

699eb47

[software] Streamline allocations in dynamic address regions

bbf4280

[hardware] Correct waves display

56654a5

[hardware] Streamline address scrambler

9c02172

[hardware] Correct address scrambler parametrization

[software] Allocation stress-test over multiple partitions & Tile-groups

d26b44a

[hardware] Trash redundant file

92242b7

[software] Correct format

543027b

[software] Remove DMA with mode selection

473bd0e

[hardware] Correct scrambler parametrization

737d106

mbertuletti force-pushed the asap25-release/das branch from 6eef71c to 7702798 Compare October 23, 2025 10:04

[hardware] Parametrize DMA for DAS

ee0ff54

[software] Add DMA with DAS test

mbertuletti force-pushed the asap25-release/das branch from 7702798 to ee0ff54 Compare October 24, 2025 14:35

rerun ci

0a7a90c

bowwwang force-pushed the asap25-release/das branch from d70ea47 to 0a7a90c Compare April 13, 2026 08:24

bowwang added 9 commits April 13, 2026 10:38

[merge] main to das

c6a8fd8

[sw] add #ifdef guards for DAS-specific code

5c11838

[sw] fix clang-format violations in DAS code

13f9335

[hw] fix trailing whitespace

d738153

[sw] remove unused variable in das_dma test

47a0425

[hw] remove unused DASMemSize localparam from mempool_pkg

9777579

[hw/sw] rename partition_sel to tiles_das for consistent DAS register…

6252225

… naming

[docs] add DAS feature to changelog

b5b6c7d

[hw/sw] regenerate control registers after tiles_das rename

997392e

bowwwang marked this pull request as ready for review April 13, 2026 10:21

bowwwang requested review from mbertuletti, sermazz and yichao-zh as code owners April 13, 2026 10:21

[sw] add das_fp32 kernel

e982df0

mbertuletti requested changes Apr 14, 2026

View reviewed changes

bowwang added 12 commits April 15, 2026 10:35

[hw/sw] rename NumDASPartitions to NumDasPartitions for consistent DA…

310d75c

…S naming

[hw/sw] move DAS registers to end of register map and clarify descrip…

2423007

…tions

[docs] move DAS entry to bottom of changelog list

b7cc499

[docs] move DAS documentation to config/README.md with usage guide

7504598

[sw] simplify DAS test filtering in Makefile

c2bac05

[sw] remove duplicate partition_free declaration in alloc.h

0cde68d

[sw] remove commented-out code and debug prints from alloc.c

c7163dc

[hw] fix misaligned DAS port parentheses

7a000bc

[hw] move DAS port declarations to bottom of module port lists

bd5e0a7

[hw] rename iDMA scrambler ports to consistent DAS naming

e51b268

[hw] replace case-based log2 with lzc in DMA split midend, renamed si…

40d2db2

…gnals with proper comments

[hw] narrow address scrambler adders from 32-bit to row-index width

e1fd77d

bowwwang commented Apr 17, 2026

View reviewed changes

Conversation

bowwwang commented Jul 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changelog

Added

Changed

Fixed

Checklist

Uh oh!

mbertuletti left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bowwwang commented Jul 1, 2025 •

edited

Loading