CUDA C: race conditions and atomics
Day: Monday, October 21Time: 2:10 PM - 3:00 PM
Place: Snedecor Hall 2113
This lecture will explore what happens when a CUDA C user must launch several threads that compete to access the same data, violating the Single Instruction Multiple Data (SIMD) paradigm. Elusive defects called race conditions occur unless the programmer uses atomic operations to force certain instruction sets to be executed sequentially.
Materials
SlidesVideo
Example Code
dot_product_atomic_builtin.cu
dot_product_atomic_lock.cu
race_condition.cu
race_condition_fixed.cu
blockCounter.cu
Makefiledot_product_atomic_lock.cu
race_condition.cu
race_condition_fixed.cu
blockCounter.cu
Makefile
Makefile
Makefile
Makefile