scratch.txt

Introduction
    - "In a few years, only a fraction of the computing power will be used" (see streamcomputing)
    - trends until 2003: increase f_clk => faster program, same code
        since 2003: f_clk remains unchanged ; N_cores increases => need to adapt the code.
        There is no longer "Just wait 18 months, my program will be twice faster" (see "your free lunch is over")

GPU architecture


Basic concepts
---------------
Context, Queue, instructions. (a)synchronous.
    This encapsulation is OOP-friendly
Host, device, kernel. On-the-fly compilation.
Software: Grid, blocks, threads
Hardware: Stream Multiprocessors, Warps, cores

Differences between other programming paradigms
    wrt OMP : fine thread control, explicit communications, host/device distinction, ...


Memory
--------

GPU: no virtual memory => no "electric fence" => no segfault. The memory can be corrupted !
Most of the times, performances are memory-related. Thus, memory access are crucial !
Warm-up : which is fastest: y*Nc + x  or  x*Nr + y ?  (on CPU and GPU).

Latency and bandwidth
    Give figures for BW on recent GPU (DH, HD, DD).
    time = latency + bandwidth*nbits
    kernels should "make the memory transfers worth". No speed-up if most of the time is spent transfering the data !
    Nvidia visual profiler, OpenCL events and profiling mode

Element-wise
-------------
add
flat-field
# Block size is free to tune

Gather
-------
2x2 binning
convolution/gradient/Haar transform
# block size is often imposed

Reduction
----------
min/max/sum
histogram
# block size is application-dependent


Linear Algebra and Fourier Transforms
--------------------------------------
cublas, cufft => easy to call from CUDA C/C++ ; python wrappers (eg. pyfft)
clblas, clfft => ?



CUDA vs OpenCL
---------------
Performance portable
Open issues regarding OpenCL
    - Will OpenCL remain a serious competitor of CUDA ? (AMD is going CUDA, Nvidia does not maintain OpenCL anymore, bugs in Apple's implementation, ...)
    - OpenCL 2.x in Linux (eg. Intel's Beignet driver)