ATC '25 and OSDI '25 -Joint Keynote Address: Accelerating Software Development: The LLM (R)evolution
OSDI '25 - QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
OSDI '25 - WaferLLM: Large Language Model Inference at Wafer Scale
OSDI '25 - BlitzScale: Fast and Live Large Model Autoscaling with O(1) Host Caching
OSDI '25 - Bayesian Code Diffusion for Efficient Automatic Deep Learning Program Optimization
| //////////////////////////////// | |
| // NBHM - Non-blocking hashmap | |
| //////////////////////////////// | |
| // You wanna intern lots of things on lots of cores? this is for you. It's | |
| // inspired by Cliff's non-blocking hashmap. | |
| // | |
| // To use it, you'll need to define NBHM_FN and then include the header: | |
| // | |
| // #define NBHM_FN(n) XXX_hm_ ## n | |
| // #include <nbhm.h> |
| X(fn) | |
| X(do) | |
| X(let) | |
| X(new) | |
| X(while) | |
| X(return) | |
| #undef X |
| //////////////////////////////// | |
| // NBHS - Non-blocking hashset | |
| //////////////////////////////// | |
| // You wanna intern lots of things on lots of cores? this is for you. It's | |
| // inspired by Cliff's non-blocking hashmap. | |
| // | |
| // To use it, you'll need to define NBHS_FN and then include the header: | |
| // | |
| // #define NBHS_FN(n) XXX_hs_ ## n | |
| // #include <nbhs.h> |
| // combined "Peeps + ConstProp + GVN" solver (extended version of the one in Combining Analyses, Combining Optimizations 7.3.2) | |
| // | |
| // returns isomorphic node (might be the same node). | |
| Node* peephole_node(Optimizer* opt, Node* n) { | |
| Node* k; | |
| bool progress = false; | |
| // potential mutations from ideal() means we need to get rehash. | |
| hashset_remove(opt->gvn_nodes, n); | |
| import sun.misc.Unsafe; | |
| import java.lang.reflect.Field; | |
| import java.nio.file.*; | |
| import java.io.*; | |
| class mur { | |
| public static Unsafe getUnsafe() { | |
| try { | |
| final Field fld = Unsafe.class.getDeclaredField("theUnsafe"); |
| import java.nio.file.*; | |
| import java.io.*; | |
| import java.util.*; | |
| import java.nio.charset.StandardCharsets; | |
| class csv { | |
| // i don't want all my fucking int array elements boxed. | |
| static class IntArray { | |
| int cnt; |
| // This is a PoC for incremental compilation for Zig, the idea is that we | |
| // can break down all of type checking into a simple term rewriting system | |
| // which has one step Church-Rosser. | |
| // | |
| // Church-Rosser (i'll shorten to CR): | |
| // CR states that given our reductions terminate, the order of reduction is irrelevant, more | |
| // importantly we can say that if we reduce to our normal form in one step per term then our | |
| // total reductions are linear with the size of the IR. | |
| // | |
| // Term? |
| // Mostly portable example of creating an object file to embed a binary. | |
| // Requires stb_ds.h (https://github.com/nothings/stb/blob/master/stb_ds.h) | |
| // Arguments: | |
| // | |
| // coff_embed [name] [input] | |
| // | |
| // name must be a valid C-identifier if you wanna actually import into C, the | |
| // idea is that the 'input' file is going to be embedded into an object file | |
| // under the symbol name 'name', the content's length is a size_t with the name | |
| // _length appended to the name. |
because of the new APX stuff i got to thinking about different changes to x86 over time and how they affect code, like general code not just the manual intrinsics case...
VEX was part of AVX and introduced ternaries which shorted the encoding.
; pre-VEX
movaps xmm1, xmm2
subss xmm1, xmm3
; VEX
vsubss xmm1, xmm2, xmm3 ; less bytes mostly