Burst support #15

mbertuletti · 2025-05-23T07:28:28Z

This PR adds burst support to the variable latency interconnect.

In systems where multiple interconnects are instantiated hierarchically, the arbitration across multiple initiators at the boundary between hierarchies causes a bottleneck in the request-per-cycle injected into the interconnect. The burst functionality solves the problem. We provide:

A burst_req_grouper module, that groups a parallel valid request in a burst request,
A burst_cutter module, that issues two bursts when the parallel request crosses the boundary between hierarchies and targets different arbitrators,
A burst_manager module, that receives a burst request and translates it into a parallel request,
A burst_rsp_grouper module, that increases the bandwidth of the response channel to send multiple words from a parallel response with the same resp_valid, thus reducing congestion to the arbiters also on the response path.

More details on the burst protocol applied to a system with hierarchical crossbars and wide memory requests initiators are in the following publication:

TCDM Burst Access: Breaking the Bandwidth Barrier in Shared-L1 RVV Clusters Beyond 1000 FPUs

@misc{shen2025tcdmburstaccessbreaking,
      title={TCDM Burst Access: Breaking the Bandwidth Barrier in Shared-L1 RVV Clusters Beyond 1000 FPUs}, 
      author={Diyou Shen and Yichao Zhang and Marco Bertuletti and Luca Benini},
      year={2025},
      eprint={2501.14370},
      archivePrefix={arXiv},
      primaryClass={cs.AR},
      url={https://arxiv.org/abs/2501.14370}, 
}

This paper is available on arXiv:2501.14370 [cs.AR].

DiyouS · 2025-05-23T10:25:13Z

rtl/variable_latency_interconnect/variable_latency_interconnect.sv

  localparam int unsigned NumOutLog2      = $clog2(NumOut);
-  localparam int unsigned IniAggDataWidth = 1 + BeWidth + AddrMemWidth + DataWidth;
+  localparam int unsigned ReqAggDataWidth = 1 + BeWidth + AddrMemWidth + DataWidth + BurstWidth;
+  localparam int unsigned RespAggDataWidth = DataWidth + 32;


We should define the width based on RspGF from the burst package. Also, the default width should be DataWidth if burst is not used

Thanks, we added a wrapper to ensure modular instantiation of the burst function. This should be handled there.

DiyouS · 2025-05-23T10:25:51Z

rtl/variable_latency_interconnect/variable_latency_interconnect.sv

+`else
+    assign req_agg_in[j] = {req_wen_i[j], req_be_i[j], req_tgt_addr_i[j][ByteOffWidth + NumOutLog2 +: AddrMemWidth], req_wdata_i[j]};
+`endif
+    assign {resp_rdata_o[j], resp_burst_o[j]} = resp_agg_out[j];


Do not include resp_burst_o when Burst is not used => signal not found issue

Thanks, we added a wrapper to ensure modular instantiation of the burst function.

DiyouS · 2025-05-23T10:27:46Z

rtl/variable_latency_interconnect/burst_manager.sv

+  parameter int unsigned NumInLog2 = (NumIn == 1) ? 1 : $clog2(NumIn),
+  // Burst response type can be overwritten for DataWidth > 32b
+  // This can happen when the DataWidth includes transaction metadata
+  parameter type burst_resp_t = tcdm_burst_pkg::burst_gresp_t


use the package definition from the burst_pkg.sv

Thanks, done.

DiyouS · 2025-05-23T10:28:05Z

rtl/variable_latency_interconnect/burst_rsp_grouper.sv

+  parameter int unsigned NumInLog2 = (NumIn == 1) ? 1 : $clog2(NumIn),
+  // Burst response type can be overwritten for DataWidth > 32b
+  // This can happen when the DataWidth includes transaction metadata
+  parameter type burst_resp_t = tcdm_burst_pkg::burst_gresp_t


use the package definition from the burst_pkg.sv

Thanks, done.

micprog · 2025-05-23T12:14:33Z

rtl/variable_latency_interconnect/variable_latency_interconnect.sv

 // Note that only the full crossbar allows NumIn/NumOut configurations that are not
 // aligned to a power of 2.

+`ifdef USE_BURST


I'm a bit concerned about the use of a define here. If using the define, it prevents SoCs from using both the bursty and the non-bursty implementations of this interconnect block (e.g., when integrated into different subsystems). My suggestion would be to convert this to a parameter or to create two different modules, each implementing different functions (one could also call the other, tying off some inputs).

Thank you, I added a wrapper.

An "isburst" signal is needed to redistribute grouped responses at the initiator.

DiyouS · 2025-06-05T14:42:53Z

rtl/variable_latency_interconnect/variable_latency_interconnect.sv

  parameter int unsigned AddrWidth         = 32,                    // Address Width on the Initiator Side
-  parameter int unsigned DataWidth         = 32,                    // Data Word Width
-  parameter int unsigned BeWidth           = DataWidth/8,           // Byte Strobe Width
+  parameter int unsigned ReqDataWidth      = 32,                    // Data Word Width on the Request path


Here the original DataWidth is replaced by ReqDataWidth and RespDataWidth, which will require all repo using this module to adapt the connection. I am not sure if it is a good idea. Maybe we can keep the parameter int unsigned DataWidth = 32 and let the default values of ReqDataWidth and RespDataWidth equals to the DataWidth. Then the burst version can still overwrite these two values without needs of modification

Absolutely, good point.

mbertuletti added 3 commits April 23, 2025 17:39

Add support for burst transactions in variable-latency-interconnect

4ef422c

Add support for wide burst responses in variable-latency-interconnect

1113fd1

Add burst request and response support

c2e88e3

mbertuletti assigned DiyouS May 23, 2025

DiyouS reviewed May 23, 2025

View reviewed changes

micprog reviewed May 23, 2025

View reviewed changes

mbertuletti added 3 commits May 27, 2025 14:24

Add "isburst" signal on the response path

6ee2d7f

An "isburst" signal is needed to redistribute grouped responses at the initiator.

Add burst_variable_latency_interconnect wrapper

f90d4c9

Add check for consecutive addresses before bursting

f656897

mbertuletti requested a review from DiyouS June 5, 2025 06:25

DiyouS reviewed Jun 5, 2025

View reviewed changes

mbertuletti force-pushed the burst_support branch 2 times, most recently from a76e1a7 to 486a10b Compare September 10, 2025 14:11

mbertuletti added 2 commits September 19, 2025 09:28

Grouped response support for outstanding transactions

3aaf4db

Propagate correct id on grouped response

f4708a6

mbertuletti force-pushed the burst_support branch 2 times, most recently from 1837b50 to 1598ff7 Compare September 24, 2025 14:14

mbertuletti added 3 commits September 25, 2025 08:45

Fix handshake in burst manager

e6e284f

Fix response collection when grouping-factor is 1

278fe27

Fix reminder computation in cutter

8071742

mbertuletti force-pushed the burst_support branch from 1598ff7 to 8071742 Compare September 25, 2025 06:45

mbertuletti added 3 commits October 7, 2025 17:37

Fix naming of signals

b87dd18

Fix for number of inputs different from number of outputs

e903c49

Add grouped request

8529a1a

mbertuletti force-pushed the burst_support branch from bc6fa6f to 8529a1a Compare October 27, 2025 06:28

Correct burst parametrization

89d030e

Burst support #15

Are you sure you want to change the base?

Burst support #15

Uh oh!

Conversation

mbertuletti commented May 23, 2025

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!

4 participants