Lecture 24: High Performance Image Processing & Halide (10)
arjunsrinivasan1997
With the gradual increases in memory speed we are seeing, would there ever be a point where locality is no longer a critical factor in improving performance/efficiency
tristanburke
Probably a bit far away (hopefully not too far), but I wonder what sort of advancements quantum computing could bring to computer graphics?
julialuo
I think when we have talked about performance in previous classes it has always been about the algorithm being used (CS 61B/170) or parallelism/hardware (CS 61C). The concept of locality being able to improve performance drastically is super simple, yet I haven't really thought about it given the context of the other CS classes here. I think it's really cool that this class is touching on it :)
venkatvarsha
For those who may have missed it in this lecture or the previous one - remember that the shader programming model is restricted such that computation is intrinsically data parallel! Since the programming model doesn't let different pixel shader instances communicate with each other, it's trivial to run the shader in memory.
fywu85
Is there a case in graphics computing where CPUs are the more suitable platform?
woojinko
though I can't say for sure, I would think that almost all graphics related computing would run better on a GPU due to all of the optimization modifications we saw in the previous lecture about going from a CPU to a GPU (no need for single instruction optimization, amortizing cost, etc.)
With the gradual increases in memory speed we are seeing, would there ever be a point where locality is no longer a critical factor in improving performance/efficiency
Probably a bit far away (hopefully not too far), but I wonder what sort of advancements quantum computing could bring to computer graphics?
I think when we have talked about performance in previous classes it has always been about the algorithm being used (CS 61B/170) or parallelism/hardware (CS 61C). The concept of locality being able to improve performance drastically is super simple, yet I haven't really thought about it given the context of the other CS classes here. I think it's really cool that this class is touching on it :)
For those who may have missed it in this lecture or the previous one - remember that the shader programming model is restricted such that computation is intrinsically data parallel! Since the programming model doesn't let different pixel shader instances communicate with each other, it's trivial to run the shader in memory.
Is there a case in graphics computing where CPUs are the more suitable platform?
though I can't say for sure, I would think that almost all graphics related computing would run better on a GPU due to all of the optimization modifications we saw in the previous lecture about going from a CPU to a GPU (no need for single instruction optimization, amortizing cost, etc.)