A collection of tools for simulating the Patmos processor.
A tiny assembler (accepts Patmos instructions, empty lines, and comments in the form of lines starting with # -- no symbols/relocation yet)
Usage: paasm <input assembly> <binary stream>
A tiny disassembler (accepts a binary stream and prints the decoded instructions).
usage: padasm <binary stream> <output assembly>
The Patmos simulator (accepts a binary stream, loads the entire stream into the simulator's main memory and begins execution at address zero).
usage: pasim <binary stream> <trace output>
Supported Platforms:
- Ubuntu 18 or higher
- MacOs
Prebuilt binaries can be found here.
- Download the latest release tarball for your machine.
- Extact the tarball in your T-CREST installation folder as-is (
tar -xvf patmos-simulator.tar.gz
). - Ensure your T-CREST installation folder is on your PATH.
- If on MacOs, to use the tools you need to have Boost's
program_options
package installed using version 1.46 or above. - Done.
To build the simulator from source, see the developer instructions below.
The simulator provides the following memory models:
-
Ideal memory: All memory transfers are executed without any delays. Used for single-core if --gtime=0 and --tdelay=0
-
Fixed delay (default): Memory transfers are executed using burst of fixed length and time. Used for single-core if --psize=0 and either --gtime or --tdelay is non-zero.
-
Variable bursts: Memory transfers are executed using bursts of arbitrary length. Used for single-core if --psize and either --gtime or --tdelay is non-zero.
-
TDM access: Similar to fixed delay, but the memory is accessed using TDM. Used for multi-core configurations.
-
Ramulator (optional): Simulates the main memory using the DRAM emulator ramulator (single-core only). In order to enable ramulator the USE_RAMULATOR option has to be activated when invoking cmake (e.g., -DUSE_RAMULATOR=on). Subsequently, two additional command-line options become available --gkind and --ramul-config.
Memory read or write request are issued by the method cache or instruction cache, the stack cache, the data cache, bypass loads and stores. Requests may not be aligned nor limited in size (e.g., the method cache loads a whole function using one request).
The memory system is responsible for aligning requests and might transfer the data using multiple bursts.
The following options are common to all memory models (if applicable):
--gsize= The size of the global memory in bytes.
--gtime=<t_B> The number of cycles per burst. The exact meaning depends on the memory model as explained below.
--bsize= The number of bytes per burst. The exact meaning depends on the memory model as explained below.
--tdelay=<t_DLY> For every read or non-posted write request, n cycles are added once. This might include delays over the NOC and the memory controller (including both send and receive delays). It is assumed that for sequencial bursts, the delay for subsequent bursts can be hidden using pipelining.
--posted=
If p=0, posted writes are disabled. If p=1, posted writes are used. If p>1, posted writes are used, and at most n writes can be queued before the processor stalls.
Options: None
The time to complete any request is always zero, i.e.,
t_REQ(n) = 0
Options: --bsize burst_size Bytes transferred per burst --gtime t_B Cycles per burst (including activation and prefetch) --tdelay t_DLY Read delay per request in cycles
Request start and end addresses are aligned to burst_size, and only full bursts are transferred. A read or non-posted write request consisting of n bursts takes
t_REQ(n) = t_DLY + n * t_B
to complete. For posted writes, it takes
t_REQwp(n) = n * t_B
cycles to complete.
Options: --bsize burst_size Bytes transferred per page access --psize page_size Bytes per page --gtime t_B Minimum cycles per accessed page --tdelay t_DLY Read delay per request in cycles
The memory is segmented into pages of page_size bytes. For each page of a request, a minimum of burst_size bytes is transferred, taking t_B cycles. Every additional word is transferred in one cycle.
t_B models the maximum activation and precharge delay for accessing one row, and burst_size the number of read or written bytes that can overlap activation and precharge.
Request start and end addresses are aligned to burst_size (thus, at least burst_size bytes are transferred per page, taking t_B cycles per page).
Let start_page(s,b) and end_page(s,b) be the first and the last page number of a request of b bytes, starting at address s. Then, the number of page activations required is
Pages(s,b) = end_page(s,b) - start_page(s,b) + 1
and for reads and non-posted writes of b bytes length (after alignment), it takes
t_REQ(s,b) = t_DLY + Pages(s,b) * t_B + (b - Pages(s,b) * burst_size) / 4
and for posted writes
t_REQwp(b) = Pages(s,b) * t_B + (b - Pages(s,b) * burst_size) / 4
The first term describes the number of cycles required to activate each page, while the second term describes the number of cycles required to transfer the remaining bytes that have not been transferred during row activation.
Note that if page_size == burst_size, this memory is identical to the fixed delay memory.
Also note that aligning start and end addresses to burst_size is not strictly required for full-page bursts and leads to longer transfers than necessary, but simplifies modelling the timing of corner-cases when a request starts or ends very close to the beginning of a page.
Options: --cores N Number of cores --cpuid CPU-ID CPU ID, defines phase in TDM round --bsize burst_size Bytes transferred per slot --gtime t_TDM Cycles per slot --tdelay t_DLY Read delay per request in cycles --trefresh t_Refresh Cycles per round reserved for refresh
On a multi-core setup with N cores, the memory is accessed using TDM with rounds of length
t_Round = N * t_TDM + t_Refresh
where t_TDM is the length of one TDM slot (--gtime), and t_Refresh is the number of cycles required per round to perform memory refreshs (--trefresh). Within one TDM slot, burst_size (--bsize) bytes are transferred. The NOC and memory controller roundtrip time is t_DLY (--tdelay).
The TDM slot of a core depends on its CPU-ID (--cpuid). The phase of a core is
Phase(CPU-ID) = CPU-ID * t_TDM
Each request is aligned to burst_size. A core must wait t_Wait < t_Round cycles for its TDM slot, and then starts sending a burst every t_Round cycles. The response per round is received after t_TDM + t_DLY cycles, which can be pipelined.
Therefore a read or non-posted write request of n bursts takes
t_REQ(n) = t_Wait + (n-1) * t_Round + t_DLY + t_TDM
cycles, and posted writes of n bursts take
t_REQwp(n) = t_Wait + (n-1) * t_Round + t_TDM
cycles.
Ramulator support is optional and needs to be activated using the USE_RAMULATOR option when invoking cmake (e.g., pass -DUSE_RAMULATOR=on to cmake).
Once activated two additional options become available: --gkind arg Select the main memory kind. The default value is "simple", which activates one of the standard memory configurations (see above). Alternative options are "ddr3", "ddr4", "lpddr3", and "lpddr4", which activate the simulation of the appropriate DRAM model using ramulator.
--ramul-config file Supply the name of a ramulator configuration file, which allows to select specific device details (number of channels, ranks, speed, ...). Example configuration files are included in the directory ramulator/configs.
Note that several of the other memory options are ignored when ramulator is active, this includes (--gtime, --tdelay, --trefresh, --psize, --posted). However, the device configuration is (automatically) adapted by ramulator in order to match the selected burst size (--bsize). This may fail, though, when the device configuration (provided through --ramul-config) is incompatible.
Requirements:
- cmake 2.6 or above
- boost 1.46 or above
- C++14 GCC or clang
- libelf 0.8.13 or above
- expect 5.45 or above
To build the simulator, run the following commands in the root directory
(assuming build
is chosen as the build directory):
mkdir build
cd build
cmake ..
make -j
To test the simulator, run the following command in the build directory:
make -j test
Requirements:
- GNU-Tar utility available as
tar
on the command line.
To package binaries for release, run the following command in the build directory:
make -j Package
It will make a tarball with the name patmos-simulator-*.tar.gz
(where *
is the build target)
that includes the following:
- A
bin
folder containing the release binaries. - A YAML file with various metadata about the release and which files are inluded.
Github Actions (CI) manages deploying a new release of these tools to Github Releases.
To trigger a release, you simply have to push a tag with the version number of the new release:
Locally, create a new tag with a version number for the release (we will use 1.0.0
as an example):
git tag 1.0.0
Then push the tag to Github:
git push --tags
Travis-CI will then handle everything and the new release binaries can be found under the Releases tab on Github.
Copyright 2012 Technical University of Denmark, DTU Compute. All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
1. Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
The views and conclusions contained in the software and documentation are those of the authors and should not be interpreted as representing official policies, either expressed or implied, of the copyright holder.
The source code of ramulator is included in the Patmos repository and licensed under the GPL-compatible MIT license (aka. Expat). See ramulator/LICENSE for additional details.