University:
Boston University
Course:
GRS CS 910 | Directed Study: Systems
Academic year:

2024
Views:

73

Pages:

16
Author:

kulawaixgi

Easy, Effective, Efficient: GPU Programming in Python with PyOpenCL and PyCUDA Outline 1 Leftovers 2 Code writes Code 3 Case Study: Generic OpenCL Reduction 4 Reasoning about Generated Code 5 Automatic GPU Programming Outline 1 Leftovers OpenCL implementations 2 Code writes Code 3 Case Study: Generic OpenCL Reduction 4 Reasoning about Generated Code 5 Automatic GPU Programming Ahem. . . Show the spec! Well. . . Thank you! Can’t say this often enough If you are performing asynchronous transfers, . . . . . . beware of Python’s big yellow garbage truck. Kernel Attributes kernel void foo( attribute ((...)) global float4 ∗p ) { .... } Implicit ↔ explicit SIMD Example: kernel void foo( attribute (( vec type hint ( float4 ))) global float4 ∗p ) { .... } Autovectorize assuming float4 as the basic computation width. Enforcing work group sizes attribute (( reqd work group size (X, Y, Z))) Outline 1 Leftovers OpenCL implementations 2 Code writes Code 3 Case Study: Generic OpenCL Reduction 4 Reasoning about Generated Code 5 Automatic GPU Programming The Nvidia CL implementation Targets only GPUs Notes: Nearly identical to CUDA No native C-level JIT in CUDA (→ PyCUDA) Page-locked memory: Use CL MEM ALLOC HOST PTR. Careful: double meaning Need page-locked memory for genuinely overlapped transfers. No linear memory texturing CUDA device emulation mode deprecated → Use AMD CPU CL (faster, too!) The Apple CL implementation Targets CPUs and GPUs General notes: Different header name OpenCL/cl.h instead of CL/cl.h Use -framework OpenCL for C access. Beware of imperfect compiler cache implementation (ignores include files) CPU notes: One work item per processor GPU similar to hardware vendor implementation. (New: Intel w/ Sandy Bridge) The AMD CL implementation Targets CPUs and GPUs (from both AMD and Nvidia) GPU notes: Wide SIMD groups (64) Native 4/5-wide vectors But: very flop-heavy machine, may ignore vectors for memory-bound workloads → Both implicit and explicit SIMD CPU notes: Many work items per processor (emulated) General: cl amd printf Outline 1 Leftovers 2 Code writes Code The Idea RTCG in Action How can I do it? 3 Case Study: Generic OpenCL Reduction 4 Reasoning about Generated Code 5 Automatic GPU Programming Outline 1 Leftovers 2 Code writes Code The Idea RTCG in Action How can I do it? 3 Case Study: Generic OpenCL Reduction 4 Reasoning about Generated Code 5 Automatic GPU Programming The OpenCL Ecosystem: One Language, Many Devices OpenCL generalizes over many types of devices: Multicore CPUs Various GPU architectures Accelerator boards The OpenCL Ecosystem: One Language, Many Devices OpenCL generalizes over many types of devices: Multicore CPUs Various GPU architectures Accelerator boards Devices differ by Memory Types, Latencies, Bandwidths Vector Widths Units of Scheduling The OpenCL Ecosystem: One Language, Many Devices OpenCL generalizes over many types of devices: Multicore CPUs Various GPU architectures Accelerator boards Devices differ by Memory Types, Latencies, Bandwidths Vector Widths Units of Scheduling Optimally tuned code will (often) be different for each device