Chih-Yi Lin
0 followers · 265 views
on the atlas · 8
- (21 封私信 / 80 条消息) 熟悉计算机体系结构是种什么样的体验？ - 知乎1 savers
- Execution Model - SLING user documentation1 savers
- A friendly introduction to machine learning compilers and optimizers10 savers
- cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlass1 savers
- cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHub2 savers
- [聘書] Offer 請益(瑞昱/達發) - 看板 Tech_Job - 批踢踢實業坊1 savers
- How To Think Real Good | Meta-rationality6 savers
- Curius / Onboarding2704 savers
highlights · 19
这些问题的答案，汇聚到最终，也就是对冯诺依曼的理解了，看明白这点，面对火热的TPU、寒武纪什么的，很大程度就能理解他们本质上是在冯诺依曼的基础上，在数据的时间性和空间性取了什么巧了。
(21 封私信 / 80 条消息) 熟悉计算机体系结构是种什么样的体验？ - 知乎The two warp schedulers select two warps and issue one instruction from each warp to a group of 16 CUDA cores, 16 load/store units, or four special function units.
Execution Model - SLING user documentationWhat if we want to generate some code that can run on just any hardware backends?
A friendly introduction to machine learning compilers and optimizersOptimizing can occur at all stages, from high-level IRs to low-level IRs.
A friendly introduction to machine learning compilers and optimizersoptimizing compiler consists of two components: lowering and optimizing
A friendly introduction to machine learning compilers and optimizerscompilers bridge ML models and the hardware they run on.
A friendly introduction to machine learning compilers and optimizersthreadblock-scoped matrix multiply operation
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassintended to be used in host-side .cu code
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassstreamline instantiation and execution of the standard GEMM computation across the GPU.
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassThe outer-most loops correspond to CTA-level hardware concurrency
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassone iteration of this loop is one "k Group"
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassone iteration of this loop is one "stage"
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlasswarp-synchronous matrix multiply instruction like mma.sync
cutlass/media/docs/gemm_api.md at main · NVIDIA/cutlassThe above code focuses only on the matrix multiply computation C = AB whose result is held in the registers of each thread within the threadblock
cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHubt enables reuse of values held in registers for multiple math instructions.
cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHubeach thread is responsible for processing a certain number of elements
cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHubo maximize data reuse within the warp, a large warp-level GEMM tile should be chosen.
cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHubHowever, large threadblock tiles may not match the dimensions of the problem well. If either the GEMM M or N dimension is small, some threads within the threadblock may not perform meaningful work, as the threadblock may be partially outside the bounds of the problem. If both M and N are small while K is large, this scheme may launch relatively few threadblocks and fail to make full use of all multiprocessors within the GPU
cutlass/media/docs/efficient_gemm.md at main · NVIDIA/cutlass · GitHubarch只能說在分工上是屬於上層，但不代表在組織上 03/20 09:44 → cta : 是屬於上層，這兩個要分清楚。如果你是想要在組織 03/20 09:44 → cta : 上往上走，其實是愈接近市場的職務愈好。
[聘書] Offer 請益(瑞昱/達發) - 看板 Tech_Job - 批踢踢實業坊