Open
Description
What happens
Measured on RX 580 (no ray queries, so the software tracer), amdvlk 31.0.21924.61, Windows 10, i5-7400, 16 GB:
- Each stage of the software tracer takes 1.5 to 5 GB of private memory to compile. After The coverage of a program in a traversal records no texture feedback #84,
reflect.comptakes 5.4 GB and 144 s. - More threads is slower.
create_pipelineswith 4 threads instead of 2 took 637 s instead of 620 s: private memory rose to 10 GB, free memory fell to 0.17 GB and Windows paged. - Too few threads is slow too.
Pipeline_Threadhasclamp(cores/2-1,1,16)workers, which is 1 on 4 cores. A coldmirai captureofdemowaits for its stages about 17 min, beyond the 600 s limit of an RPC.
Both counts come from the cores, and neither from what a compile costs in memory or what the machine has free.
Expected
The number of compiles running at once is bounded by memory as well as cores, so a cold start neither pages nor leaves cores idle. A fixed ratio from one machine is not the answer (see the review of #15). The bound should come from what the compiles of this run actually took, or from what the system reports free.
🤖 Generated with Claude Code
Metadata
Metadata
Assignees
Labels
No labels
Activity