Bug 140 - Implement AMDVLK / RADV Mesa Vulkan Driver NLNet 2019-10-042
Summary: Implement AMDVLK / RADV Mesa Vulkan Driver NLNet 2019-10-042
Alias: None
Product: Libre-SOC's first SoC
Classification: Unclassified
Component: Milestones (show other bugs)
Version: unspecified
Hardware: Other Linux
: --- enhancement
Assignee: Luke Kenneth Casson Leighton
: 247 (view as bug list)
Depends on: 53 260 386 405
Blocks: 251 252 253 254 255
  Show dependency treegraph
Reported: 2019-10-03 08:25 BST by Luke Kenneth Casson Leighton
Modified: 2020-06-26 18:42 BST (History)
5 users (show)

See Also:
NLnet milestone: NLNet.2019.Vulkan
total budget (EUR) for completion of task and all subtasks: 37500
budget (EUR) for completion of task (excludes budget allocated to subtasks): 0
parent task for budget allocation:
child tasks for budget allocation: 251 252 253 254 255


Note You need to log in before you can comment on or make changes to this bug.
Description Luke Kenneth Casson Leighton 2019-10-03 08:25:57 BST
An additional Vulkan driver is to be developed, alongside Kazan, starting instead from preexisting code, AMDVLK or RADV. Decision still to be made. AMDGPU assembly to be replaced with Libre RISCV SV Vectorisation and hardware accelerated opcodes for Texture interpolation, YUV2RGB, Z buffering etc.

LLVM IR assembly portion extremely likely to be shared with Kazan.

* bug #251 - Initial software-only 3D MESA driver
* bug #252 - 3D opcodes to be added to ISA simulator
* bug #253 - 3D opcodes hardware implementations
* bug #254 - 3D second iteration, simulation to hardware.
* bug #255 - 3D Formal Standard for OpenPOWER
Comment 1 Luke Kenneth Casson Leighton 2020-02-23 22:19:48 GMT
* create a 3D driver that "works with LLVM-IR and is effectively a portable software-only driver", very similar to SwiftShader.  this would be estimated (as a major milestone) around 3-4 months.
* (working closely with Jacob) identify the instructions which, if added, would accelerate 3D rendering, and arrange to use them in both the Mesa NIR *and* make sure that the LLVM-IR compiler recognises them.
Comment 2 Luke Kenneth Casson Leighton 2020-02-23 22:38:07 GMT
Schedule A for MoU 2019-10-042

# Initial Software-only 3D MESA Driver
software only (SwiftShader style) 3D MESA Driver, portable (x86 etc) using non-accelerated LLVM

Budget: EUR 12000
URL: http://bugs.libre-riscv.org/show_bug.cgi?id=251

# 3D accelerated opcodes need to be added to the POWER ISA simulator 
adding iterated support for 3D opcodes to ISA Simulator (and unit tests)

Budget: EUR 7000
URL: http://bugs.libre-riscv.org/show_bug.cgi?id=252

# Add hardware implementations of 3D accelerated opcodes 
first iterative effort developing and adding hardware accelerated 3D opcodes to 3D MESA driver

Budget: EUR 12500
URL: http://bugs.libre-riscv.org/show_bug.cgi?id=253

# 2nd iteration of opcodes in 3D MESA driver
A 2nd iteration of opcodes in 3D MESA driver is needed.  This involves assessing and reporting on the level of success of the first round of simulations and ease of implementability in hardware, as well as the effectiveness in number of gates, power consumption and (learning from Jeff Bush's Nyuzi work) the number of pixels processed per clock.

Budget: EUR 12500
URL: http://bugs.libre-riscv.org/show_bug.cgi?id=254

# Second iteration round for opcodes, simulation and hardware for 3D MESA
Documentation to the level of an ISA Standards Proposal of all hardware opcodes

Budget: 6000
URL: http://bugs.libre-riscv.org/show_bug.cgi?id=255
Comment 3 Luke Kenneth Casson Leighton 2020-02-23 23:03:22 GMT
note that bug #53 has budget from NLNet 2019-02-012 which should not duplicate here.  #53 is primarily for Kazan however the HW can be used here as well obvioudly
Comment 4 Luke Kenneth Casson Leighton 2020-02-24 16:14:41 GMT
Software Dependencies:

* All software build dependencies of the MESA library (yes this itself is a large list)
* latex (and, depending on Alain's preferences for GUI editors, potentially texstudio)
* LibreSOC HDL and all LibreSoc build dependencies, which include yosys and nmigen 
* the gem5 POWER ISA simulator and its software build dependencies.

everything that we need to modify will be on http://git.libre-riscv.org.
Comment 5 Luke Kenneth Casson Leighton 2020-02-24 16:39:01 GMT

We have the unusual goal of developing not just a new 3D MESA driver:
the hardware itself is being simultaneously developed in conjunction with
the software driver.

However before spending the time implementing full hardware (even in HDL),
which will be emulated significantly slower than actual hardware, we need
to implement - and test and evaluate - the planned instructions using an
emulator, first.  This significantly reduces both development time and
cost, allowing fast iteration and evaluation.

Once we are happy with the chosen hardware instructions at the simulation
level, then we may move to implementing them in hardware.  We will need
to write full unit tests of both the hardware and also run software unit
tests as best as practical on emulated hardware (knowing that it will be
operating thousands to tens of thousands of times slower) and, if practical,
running on an FPGA board.

After the hardware-level simulation, if the gate count and estimated
gate count is acceptable when compared to commercial GPUs, *and* the number
of pixels generated per clock is at an acceptable commercial level,
then we can consider the hardware and software part of the project to be

Until then, we iterate round the above, and, once completed, finish off
by documenting the hardware instructions as an "ISA Extension" for POWER.
Comment 6 Alain D D Williams 2020-03-13 15:10:27 GMT
Comment 7 Luke Kenneth Casson Leighton 2020-03-13 18:12:18 GMT
*** Bug 247 has been marked as a duplicate of this bug. ***