ref: 35ff0931afeac236868423ae29a8dde69892f64b
dir: /slides/slides.ms/
.so tmac
.TL
.rs
.sp |3i
GPU Filesystem for Plan 9
.LP
.ce 100
Joel Fridolin Meyer
.CW \s-3joel@sirjofri.de\s0
.sp 1i
10th International Workshop on Plan 9
.br
\s-3April 12-14 2024\s0
.ce 0
.
.bp
.ds LF "IWP9 2024
.ds CF "GPU Filesystem for Plan 9
.ds RF "%
.
.TL
What is a GPU?
.DS
.PS <what.p
.PE
.DE
.IP •
Coprocessor
.IP •
``Separate Computer''
.IP •
``Graphics'' Processing Unit
.IP •
Single Instruction, Multiple Data (SIMD)
.
.bp
.TL
Application Architecture
.DS
.PS <arch.p
.PE
.DE
.IP •
Resource Abstraction (Plug and Play)
.IP •
9P Performance?
.IP •
Scalability on the application level
.
.bp
.TL
Inspired by Vulkan
.sp 1i
.IP •
Lowlevel API
.IP •
Abstract and cross-platform
.IP •
Shaders and buffers as objects
.
.bp +0
.TL
Inspired by Vulkan
.sp 1i
.IP •
Lowlevel API
.IP •
Abstract and cross-platform
.IP •
Shaders and buffers as objects
.DS
.PS <comp.p
.PE
.DE
.
.bp
.TL
Filesystem Interface
.DS
.PS
down
move 2
box "\fB\s+3Work in progress!\s0\fR" wid 4 ht 1.2
.PE
.DE
.
.bp
.TL
Filesystem Interface
.sp 1i
.IP •
Simple to understand
.IP •
Generic enough to be portable
.IP •
Plug and Play
.IP •
GPGPU first (compute shader)
.
.bp +0
.TL
Filesystem Interface
.sp .5i
.DS
.ft CW
; echo n b 4 > ctl # object 0, buffer 4 bytes
; echo n s > ctl # object 1, shader
; cp buffer.bin 0/buffer
; cp shader.spv 1/shader
.DE
.sp 1i
.DS
.ft CW
/dev/gpu
/ctl \fR\f(CW
/desc \fRdescriptor table\f(CW
/0/ctl \fRobject 0 ctl file\f(CW
/0/buffer \fRbuffer data\f(CW
/1/ctl \fRobject 1 ctl file\f(CW
/1/shader \fRshader program\f(CW
.DE
.
.bp
.TL
Filesystem Interface
.sp .5i
.DS
.ft CW
DescPool 0
Set 0
0 -1 \fR(unbound)\f(CW
1 0 \fR(Buffer 0)\f(CW
2 -1 \fR(unbound)\f(CW)
3 -1 \fR(unbound)\f(CW)
Set 1
.DE
.IP •
Shader is bound to a descriptor pool
.IP •
Buffer is bound to a specific binding
.IP •
Shader program knows which set+binding to access
.
.bp
.TL
Example Shader Program
.sp .2i
.DS
.ft CW
#version 450 core
layout(std430, set = 0, binding = 1) buffer pos
{
vec4 positions[];
};
layout(local_size_x = 1000,
local_size_y = 1, local_size_z = 1) in;
void main() {
vec4 position =
positions[gl_GlobalInvocationID.x];
// do something beautiful
}
.DE
.
.bp
.TL
Workgroups
.sp 1i
.IP •
Number of workgroups (set by application)
.IP •
Workgroup size (compiled in the shader; \f(CWlocal_size_x\fR, …)
.IP •
Both are 3D vectors
.IP •
#(Invocations) = #(workgroups) · (Workgroup size)
.
.bp
.TL
Example Shader Program
.DS
.PS <glsl.p
.PE
.DE
.IP •
Lots of magic happens in the driver
.IP •
SPIR-V program is cross-platform
.
.bp
.TL
Implementation
.sp 1i
.IP •
Draft Implementation (CPU)
.IP •
Drawterm (GPU, Vulkan SDK)
.IP •
Possible others? (SSE/CPU, network, hardware, …)
.
.bp
.LP
.rs
.sp |3.5i
.ce 100
\s+5\fBThank you\fR\s0
Questions?
.ce 0
.ds LF "
.ds RF "
.ds CF "\s-8\f(CWpic slides.ms | troff -ms\fR\s0
.