D-PP-Lec03-Modern_Multicore_Processor_Efficient_Execution

Thread-Level Parallelism and SIMD Parallelism on OpenMP

Thread-Level Parallelism

• Utilize multiple CPU cores
• Suitable for task-level parallelism (every thread can run different instruction)

SIMD Parallelism

• Utilize SIMD vector units within a CPU core
• Ideal for data-level parallelism (can only do the same instruction)
• Best for simple, repetitive operations on arrays

---

Conditional Execution

Conditional Execution

D-PP-Lec03a-Conditional-Execution

Terminology

Coherent Execution

• Same instructions are applied to all elements
• Necessary for SIMD

Divergent Execution

• have if else or different process might execute different part of the code
• Not efficient use of SIMD or SIMT

---

Efficient Execution

Memory Constraints

D-PP-Lec03b-Memory-Constraints

Hiding Memory Latency

D-PP-Lec03c-Prefetching D-PP-Lec03d-Multi-Threading