1-0-introduction.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "2b17e0f6",
   "metadata": {},
   "source": [
    "# Introduction\n",
    "\n",
    "Goal of this course: learn how to program GPUs with Julia\n",
    "\n",
    "Expected experience: minimal familiarity with Julia and GPU programming\n",
    "\n",
    "Day 1: programming the GPU\n",
    "- Introduction\n",
    "- Programming models\n",
    "- Profiling and optimization\n",
    "\n",
    "Day 2: advanced topics\n",
    "- Memory management\n",
    "- Concurrent computing\n",
    "- Low-level CUDA APIs"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4a58fd2b",
   "metadata": {},
   "source": [
    "## Why GPU programming in Julia?\n",
    "\n",
    "*Very briefly*: Julia is\n",
    "\n",
    "- a **high level** language: easy to write and read\n",
    "- **designed for performance**: by using careful abstractions, and a JIT compiler\n",
    "\n",
    "This relatively unique combination makes it a great language for GPU programming, offering both high-level abstractions that are more customizable than in other languages, while also making it possible to write your own kernels and do other low-level work."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e36f2946",
   "metadata": {},
   "source": [
    "## Why CUDA.jl?\n",
    "\n",
    "Essentially, because it's the most mature and best optimized GPU back-end for Julia.\n",
    "\n",
    "However, other back-ends are built on top of the same stack, and share a lot of the same infrastructure:\n",
    "\n",
    "- GPUArrays.jl: vendor neutral array operations\n",
    "- GPUCompiler.jl: shared infrastructure for kernel compilation\n",
    "- LLVM.jl, Adapt.jl, etc\n",
    "\n",
    "As a result, it should be possible to apply much of what you learn here to other back-ends as well. In order of maturity:\n",
    "\n",
    "- AMDGPU.jl: for [ROCM-supported](https://rocm.docs.amd.com/en/latest/release/gpu_os_support.html) GPUs on Linux\n",
    "- oneAPI.jl: for most [Intel IGPs and dedicated GPUs](https://github.com/intel/compute-runtime#supported-platforms) on Linux\n",
    "- Metal.jl: for Apple M-series GPUs on macOS\n",
    "\n",
    "There's some other experimental back-ends; for more details, see [juliagpu.org](https://juliagpu.org/)."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0e536876",
   "metadata": {},
   "source": [
    "## Set-up"
   ]
  },
  {
   "attachments": {
    "image.png": {
     "image/png": "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"
    }
   },
   "cell_type": "markdown",
   "id": "3ae97e61",
   "metadata": {},
   "source": [
    "Once connected to the CSCS JupyterLab instance, check-out the course repository at `https://github.com/maleadt/cscs2023` using the Jupyter UI:\n",
    "\n",
    "![image.png](attachment:image.png)\n",
    "\n",
    "Then, double-click on the `cscs2023` folder, and open the `1-0-introduction.ipynb` notebook to follow along.\n",
    "\n",
    "<div class=\"alert alert-block alert-warning\">\n",
    "    <strong>Warning</strong>: Because of how Piz Daint is set-up, you can only access the GPU from a single process. That means you may have to kill previous notebooks before starting a new one.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "59b64961",
   "metadata": {},
   "source": [
    "CUDA.jl is easy to install, just do `Pkg.add(\"CUDA\")`. For this course, I've provided a pre-configured environment with all relevant packages:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "29815c76",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\u001b[32m\u001b[1m  Activating\u001b[22m\u001b[39m project at `~/cscs2023`\n",
      "\u001b[32m\u001b[1mPrecompiling\u001b[22m\u001b[39m project...\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSIMDTypes\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRangeArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRealDot\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mUnPack\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIndirectArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCustomUnitRanges\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mManualMemory\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStatsAPI\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCompilerSupportLibraries_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIterTools\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTensorCore\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPkgVersion\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mDocStringExtensions\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIntervalSets\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mProgressMeter\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIfElse\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRatios\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mInflate\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLazyModules\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mNaNMath\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIrrationalConstants\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mMappedArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mComputationalResources\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCpuId\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mlibpng_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTranscodingStreams\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSimpleTraits\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIntelOpenMP_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSuiteSparse\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImath_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mWoodburyMatrices\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mJpegTurbo_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mZstd_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mFFTW_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLERC_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mQuaternions\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mParameters\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRecipesBase\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mOpenSpecFun_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIntervalSets → IntervalSetsStatisticsExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRatios → RatiosFixedPointNumbersExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mThreadingUtilities\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCoordinateTransformations\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRegionTrees\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mOffsetArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStatic\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mArnoldiMethod\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mDistances\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTranscodingStreams → TestExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLogExpFunctions\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mGraphics\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mMKL_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mAxisAlgorithms\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mOpenEXR_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mArrayInterface\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mlibsixel_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mFileIO\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStackViews\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCatIndices\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLibtiff_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPaddedViews\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mBitTwiddlingConvenienceFunctions\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mIntegralArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mAxisArrays\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mDistances → DistancesSparseArraysExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mArrayInterface → ArrayInterfaceGPUArraysCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCPUSummary\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39mBenchmarkTools\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mQOI\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mColorVectorSpace\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mMosaicViews\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLogExpFunctions → LogExpFunctionsChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mRotations\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mOpenEXR\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mDistances → DistancesChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mArrayInterface → ArrayInterfaceStaticArraysCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageMagick_jll\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mHostCPUFeatures\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPolyesterWeave\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mInterpolations\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStringDistances\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStatsBase\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mNearestNeighbors\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSpecialFunctions\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mColorVectorSpace → SpecialFunctionsExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mFFTW\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSpecialFunctions → SpecialFunctionsChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mGraphs\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mClustering\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mFFTViews\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mDiffRules\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSimpleWeightedGraphs\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mColorSchemes\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mArrayInterface → ArrayInterfaceCUDAExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCUDA → SpecialFunctionsExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStaticArrayInterface\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39mTullio\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTullio → TullioChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStaticArrayInterface → StaticArrayInterfaceStaticArraysExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mStaticArrayInterface → StaticArrayInterfaceOffsetArraysExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPolynomials\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCUDA → ChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mCloseOpenIntervals\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTiledIteration\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPolynomials → PolynomialsChainRulesCoreExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mJLD2\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLayoutPointers\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTullio → TullioCUDAExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mMetaGraphs\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageCore\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mVectorizationBase\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mJpegTurbo\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSLEEFPirates\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mSixel\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageBase\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mPNGFiles\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageMagick\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageShow\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mHistogramThresholding\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageTransformations\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageAxes\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageMetadata\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageBinarization\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageContrastAdjustment\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mNetpbm\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mTiffImages\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39mImageIO\n",
      "\u001b[32m  ✓ \u001b[39mTestImages\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLoopVectorization\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mLoopVectorization → SpecialFunctionsExt\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageMorphology\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageDistances\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageFiltering\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageSegmentation\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageCorners\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39m\u001b[90mImageQualityIndexes\u001b[39m\n",
      "\u001b[32m  ✓ \u001b[39mImages\n",
      "  137 dependencies successfully precompiled in 100 seconds. 76 already precompiled.\n"
     ]
    }
   ],
   "source": [
    "using Pkg\n",
    "Pkg.activate(@__DIR__)\n",
    "Pkg.instantiate()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3c33037c",
   "metadata": {},
   "source": [
    "This will take a couple of minutes. Afterwards, try `CUDA.versioninfo()` to see some details about the environment:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "1da9df78",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CUDA runtime 11.8, artifact installation\n",
      "CUDA driver 11.4\n",
      "NVIDIA driver 470.57.2\n",
      "\n",
      "CUDA libraries: \n",
      "- CUBLAS: 11.11.3\n",
      "- CURAND: 10.3.0\n",
      "- CUFFT: 10.9.0\n",
      "- CUSOLVER: 11.4.1\n",
      "- CUSPARSE: 11.7.5\n",
      "- CUPTI: 18.0.0\n",
      "- NVML: 11.0.0+470.57.2\n",
      "\n",
      "Julia packages: \n",
      "- CUDA: 5.1.0\n",
      "- CUDA_Driver_jll: 0.7.0+0\n",
      "- CUDA_Runtime_jll: 0.10.0+1\n",
      "\n",
      "Toolchain:\n",
      "- Julia: 1.10.0-rc1\n",
      "- LLVM: 15.0.7\n",
      "\n",
      "Preferences:\n",
      "- CUDA_Driver_jll.compat: false\n",
      "\n",
      "1 device:\n",
      "  0: Tesla P100-PCIE-16GB (sm_60, 15.897 GiB / 15.899 GiB available)\n"
     ]
    }
   ],
   "source": [
    "using CUDA\n",
    "CUDA.versioninfo()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "de438801",
   "metadata": {},
   "source": [
    "CUDA.jl bundles many CUDA libraries (from the core toolkit as well as external libraries), making them available as artifacts packaged using BinaryBuilder.jl. Selection of these libraries is tricky, as several compatibility requirements need to be taken into account:\n",
    "\n",
    "- the local NVIDIA driver has a CUDA Toolkit compatibility level\n",
    "- starting with CUDA 11.0, the driver is forwards-compatible (\"Enhanced Compatibility\")\n",
    "- it may be possible to load a more recent driver library to raise the compatibility level (\"Forward Compatibility\")\n",
    "- external libraries (CUDNN, CUTENSOR) may have builds for each CUDA Toolkit"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c4ca9cec",
   "metadata": {},
   "source": [
    "For this course, this should be using CUDA 11.8 from artifacts (instead of the old CUDA 11.0 toolkit provided by the module system)."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7ddd8c79",
   "metadata": {},
   "source": [
    "If you would ever need to disable use of artifacts, you can call `CUDA.set_runtime_version!(version; local_toolkit=true)` where `version` is the version of the local CUDA toolkit. In this configuration, you are responsible for ensuring compatibility between the driver, toolkit, and any libraries! `CUDA.versioninfo()` will then look as follows:"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a8f28a85",
   "metadata": {},
   "source": [
    "```julia-repl\n",
    "julia> CUDA.versioninfo()\n",
    "CUDA toolkit 11.0, local installation\n",
    "CUDA driver 11.4\n",
    "NVIDIA driver 470.57.2\n",
    "\n",
    "...\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0f1dfed7",
   "metadata": {},
   "source": [
    "If CUDA.jl doesn't manage to discover your local CUDA installation, you can try launching Julia with the environment variable `JULIA_DEBUG` set to `CUDA_Runtime_Discovery`. This will reveal the locations CUDA.jl searches in for toolkit libraries and binaries:"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d03d79f3",
   "metadata": {},
   "source": [
    "```\n",
    "$ JULIA_DEBUG=CUDA_Runtime_Discovery \\\n",
    "  julia -e 'using CUDA'\n",
    "┌ Debug: Looking for CUDA toolkit via environment variables CUDA_PATH\n",
    "└ @ CUDA_Runtime_Discovery ~/.julia/packages/CUDA_Runtime_Discovery/VQ7FU/src/CUDA_Runtime_Discovery.jl:294\n",
    "┌ Debug: Looking for binary ptxas in /opt/cuda\n",
    "│   all_locations =\n",
    "│    2-element Vector{String}:\n",
    "│     \"/opt/cuda\"\n",
    "│     \"/opt/cuda/bin\"\n",
    "└ @ CUDA_Runtime_Discovery ~/.julia/packages/CUDA_Runtime_Discovery/VQ7FU/src/CUDA_Runtime_Discovery.jl:164\n",
    "┌ Debug: Found ptxas at /opt/cuda/bin/ptxas\n",
    "└ @ CUDA_Runtime_Discovery ~/.julia/packages/CUDA_Runtime_Discovery/VQ7FU/src/CUDA_Runtime_Discovery.jl:170\n",
    "┌ Debug: Looking for binary nvdisasm in /opt/cuda\n",
    "⋮\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0de9e2f6",
   "metadata": {},
   "source": [
    "## A quick tour\n",
    "\n",
    "CUDA.jl is a large package, containing a lot of functionality:\n",
    "\n",
    "- wrappers for the CUDA driver, to manage devices, streams, etc\n",
    "- wrappers for CUDA's libraries, such as cuBLAS, cuFFT, cuRAND, etc\n",
    "- native kernel programming support: `@cuda`, intrinsics wrappers, a compiler, etc\n",
    "- high-level functionality that builds on all of the above: `CuArray`, array operations, stdlib integrations, etc\n",
    "\n",
    "As a quick introduction, I'll use a simple parallel operation that can be used to demonstrate all of the above: AXPY, or $z = \\alpha x + y$."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "209926b1",
   "metadata": {},
   "source": [
    "A simple CPU implementation of AXPY using array operations would look as follows:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "83425962",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2-element Vector{Int64}:\n",
       "  6\n",
       " 11"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "function axpy!(z, a, x, y)\n",
    "    z .= a .* x .+ y\n",
    "end\n",
    "\n",
    "x = [1, 2]\n",
    "y = [2, 3]\n",
    "alpha = 4\n",
    "z = similar(x)\n",
    "\n",
    "axpy!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7bf220a2",
   "metadata": {},
   "source": [
    "At the highest level of abstraction, we can use exactly the same array operations with CUDA.jl to run this on the GPU:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "cffb930a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2-element CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}:\n",
       "  6\n",
       " 11"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x = CuArray([1, 2])\n",
    "y = CuArray([2, 3])\n",
    "alpha = 4\n",
    "z = similar(x)\n",
    "\n",
    "axpy!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "49e50f38",
   "metadata": {},
   "source": [
    "This is one of the strenghts of Julia's GPU ecosystem: Generic array operations make it possible to write code that works on all kinds of inputs. Just changing the input type is sufficient to \"port\" an application to the GPU. Of course, this doesn't always work out perfectly, and sometimes you still need custom operations, but it's a great starting point."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "22f319ca",
   "metadata": {},
   "source": [
    "Let's peel back one of the layers, and implement a custom kernel instead of re-using the `broadcast` implementation from CUDA.jl"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "df4f8635",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2-element CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}:\n",
       "  6\n",
       " 11"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "function axpy_kernel!(z, a, x, y)\n",
    "    function kernel(z, a, x, y)\n",
    "        i = threadIdx().x\n",
    "        if i ≤ length(z)\n",
    "            @inbounds z[i] = a * x[i] + y[i]\n",
    "        end\n",
    "        return\n",
    "    end\n",
    "\n",
    "    @cuda threads=length(z) kernel(z, a, x, y)\n",
    "\n",
    "    return z\n",
    "end\n",
    "\n",
    "x = CuArray([1, 2])\n",
    "y = CuArray([2, 3])\n",
    "alpha = 4\n",
    "z = similar(x)\n",
    "axpy_kernel!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b916f0f7",
   "metadata": {},
   "source": [
    "The kernel above is too simple though, as it ignores a crucial limitation: **you cannot launch an unbounded number of threads**, but need to respect the device limit and, if needed, launch multiple blocks. Let's use at most 256 threads, which should work on any GPU:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "e5db8ff1",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2-element CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}:\n",
       "  6\n",
       " 11"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "function axpy_kernel!(z, a, x, y)\n",
    "    function kernel(z, a, x, y)\n",
    "        i = (blockIdx().x - 1) * blockDim().x + threadIdx().x\n",
    "        if i ≤ length(z)\n",
    "            @inbounds z[i] = a * x[i] + y[i]\n",
    "        end\n",
    "        return\n",
    "    end\n",
    "\n",
    "    kernel = @cuda launch=false kernel(z, a, x, y)\n",
    "    threads = min(length(z), 256)\n",
    "    blocks = cld(length(z), threads)\n",
    "    kernel(z, a, x, y; threads, blocks)\n",
    "\n",
    "    return z\n",
    "end\n",
    "\n",
    "x = CuArray([1, 2])\n",
    "y = CuArray([2, 3])\n",
    "alpha = 4\n",
    "z = similar(x)\n",
    "axpy_kernel!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "812f6544",
   "metadata": {},
   "source": [
    "In fact, there's already a good reason to use this `saxpy_kernel!` version, as it's generates slightly better code (requiring fewer registers) than the fully generic `broadcast` implementation:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "2c04df47",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Profiler ran for 467.3 µs, capturing 4 events.\n",
       "\n",
       "Host-side activity: calling CUDA APIs took 49.83 µs (10.66% of the trace)\n",
       "┌────┬──────────┬──────────┬────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m    Start \u001b[0m│\u001b[1m     Time \u001b[0m│\u001b[1m Name           \u001b[0m│\n",
       "├────┼──────────┼──────────┼────────────────┤\n",
       "│  2 │ 41.96 µs │\u001b[31m 49.59 µs \u001b[0m│\u001b[1m cuLaunchKernel \u001b[0m│\n",
       "└────┴──────────┴──────────┴────────────────┘\n",
       "\n",
       "Device-side activity: GPU was busy for 370.5 µs (79.29% of the trace)\n",
       "┌────┬──────────┬──────────┬─────────┬────────┬──────┬─────────────────────────────────────────────────────────────────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m    Start \u001b[0m│\u001b[1m     Time \u001b[0m│\u001b[1m Threads \u001b[0m│\u001b[1m Blocks \u001b[0m│\u001b[1m Regs \u001b[0m│\u001b[1m Name                                                                        \u001b[0m│\n",
       "├────┼──────────┼──────────┼─────────┼────────┼──────┼─────────────────────────────────────────────────────────────────────────────┤\n",
       "│  2 │ 84.88 µs │\u001b[31m 370.5 µs \u001b[0m│     256 │  65536 │    8 │\u001b[1m _Z6kernel13CuDeviceArrayI7Float32Li2ELi1EES0_S_IS0_Li2ELi1EES_IS0_Li2ELi1EE \u001b[0m│\n",
       "└────┴──────────┴──────────┴─────────┴────────┴──────┴─────────────────────────────────────────────────────────────────────────────┘\n"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x = CUDA.rand(Float32, 4096, 4096)\n",
    "y = CUDA.rand(Float32, 4096, 4096)\n",
    "z = similar(x)\n",
    "alpha = rand(Float32)\n",
    "axpy_kernel!(z, alpha, x, y)   # warm-up\n",
    "CUDA.@profile trace=true axpy_kernel!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "2452e9f4-9e83-4117-a097-fdaed064a36c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(30, 80)"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Base.displaysize()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "9d963774",
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Profiler ran for 1.06 ms, capturing 6 events.\n",
       "\n",
       "Host-side activity: calling CUDA APIs took 461.58 µs (43.39% of the trace)\n",
       "┌────┬──────────┬───────────┬────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m    Start \u001b[0m│\u001b[1m      Time \u001b[0m│\u001b[1m Name           \u001b[0m│\n",
       "├────┼──────────┼───────────┼────────────────┤\n",
       "│  4 │ 52.93 µs │\u001b[31m 458.48 µs \u001b[0m│\u001b[1m cuLaunchKernel \u001b[0m│\n",
       "└────┴──────────┴───────────┴────────────────┘\n",
       "\n",
       "Device-side activity: GPU was busy for 552.42 µs (51.93% of the trace)\n",
       "┌────┬───────────┬───────────┬─────────┬────────┬──────┬───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m     Start \u001b[0m│\u001b[1m      Time \u001b[0m│\u001b[1m Threads \u001b[0m│\u001b[1m Blocks \u001b[0m│\u001b[1m Regs \u001b[0m│\u001b[1m Name                                                                                                                                                                                                                                                                      \u001b[0m│\n",
       "├────┼───────────┼───────────┼─────────┼────────┼──────┼───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┤\n",
       "│  4 │ 508.55 µs │\u001b[31m 552.42 µs \u001b[0m│    1024 │    112 │   32 │\u001b[1m _Z16broadcast_kernel15CuKernelContext13CuDeviceArrayI7Float32Li2ELi1EE11BroadcastedI12CuArrayStyleILi2EE5TupleI5OneToI5Int64ES5_IS6_EE1_S4_IS2_IS3_ILi2EEvS7_S4_IS1_8ExtrudedIS0_IS1_Li2ELi1EES4_I4BoolS9_ES4_IS6_S6_EEEES8_IS0_IS1_Li2ELi1EES4_IS9_S9_ES4_IS6_S6_EEEES6_ \u001b[0m│\n",
       "└────┴───────────┴───────────┴─────────┴────────┴──────┴───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘\n"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "axpy!(z, alpha, x, y)\n",
    "CUDA.@profile trace=true axpy!(z, alpha, x, y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "a2961e94-c408-441c-9e65-f642fdffe20d",
   "metadata": {},
   "outputs": [],
   "source": [
    "# work around a Jupyter bug\n",
    "HTML(\"\"\"\n",
    "    <style>\n",
    "        /*Disable code output line wrapping*/\n",
    "        div.jp-RenderedText pre {\n",
    "            white-space: pre;\n",
    "        }\n",
    "    </style>\"\"\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2e7e493b",
   "metadata": {},
   "source": [
    "Another way to compute AXPY is to use the cuBLAS library, which provides the `cublasXaxpy` functions. We've got those conveniently wrapped in CUDA.jl as `CUBLAS.axpy!`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "478b6372",
   "metadata": {},
   "outputs": [
    {
     "ename": "LoadError",
     "evalue": "MethodError: no method matching axpy!(::Int64, ::Int64, ::CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}, ::CuArray{Int64, 1, CUDA.Mem.DeviceBuffer})\n\n\u001b[0mClosest candidates are:\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{ComplexF16}\u001b[39m, \u001b[91m::StridedCuArray{ComplexF16}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:226\u001b[24m\u001b[39m\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{Float16}\u001b[39m, \u001b[91m::StridedCuArray{Float16}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:221\u001b[24m\u001b[39m\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{ComplexF32}\u001b[39m, \u001b[91m::StridedCuArray{ComplexF32}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:211\u001b[24m\u001b[39m\n\u001b[0m  ...\n",
     "output_type": "error",
     "traceback": [
      "MethodError: no method matching axpy!(::Int64, ::Int64, ::CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}, ::CuArray{Int64, 1, CUDA.Mem.DeviceBuffer})\n\n\u001b[0mClosest candidates are:\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{ComplexF16}\u001b[39m, \u001b[91m::StridedCuArray{ComplexF16}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:226\u001b[24m\u001b[39m\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{Float16}\u001b[39m, \u001b[91m::StridedCuArray{Float16}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:221\u001b[24m\u001b[39m\n\u001b[0m  axpy!(::Integer, ::Number, \u001b[91m::StridedCuArray{ComplexF32}\u001b[39m, \u001b[91m::StridedCuArray{ComplexF32}\u001b[39m)\n\u001b[0m\u001b[90m   @\u001b[39m \u001b[35mCUDA\u001b[39m \u001b[90m/scratch/snx3000/julia/tbesard/daint-gpu/packages/CUDA/hfp7r/lib/cublas/\u001b[39m\u001b[90m\u001b[4mwrappers.jl:211\u001b[24m\u001b[39m\n\u001b[0m  ...\n",
      "",
      "Stacktrace:",
      " [1] top-level scope",
      "   @ In[24]:4"
     ]
    }
   ],
   "source": [
    "x = CuArray([1, 2])\n",
    "y = CuArray([2, 3])\n",
    "alpha = 4\n",
    "CUBLAS.axpy!(length(y), alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9fb12c71",
   "metadata": {},
   "source": [
    "Well, that immediately demonstrates one of the limitations of NVIDIA's vendor libraries: they only support a handful of types. This is one of the reasons that a native compiler is so valuable, as it allows us to provide both a library-backed implementation and a native implementation."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5782607f-3bc5-4c87-86ea-3a617244f5fc",
   "metadata": {},
   "source": [
    "To test CUBLAS, let's use Float32 inputs:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "f9fca372",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2-element CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}:\n",
       "  6.0\n",
       " 11.0"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x = CuArray([1f0, 2f0])\n",
    "y = CuArray([2f0, 3f0])\n",
    "alpha = 4f0\n",
    "CUBLAS.axpy!(length(y), alpha, x, y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "7d71ca3b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Profiler ran for 957.01 µs, capturing 6 events.\n",
       "\n",
       "Host-side activity: calling CUDA APIs took 536.44 µs (56.05% of the trace)\n",
       "┌────┬───────────┬───────────┬──────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m     Start \u001b[0m│\u001b[1m      Time \u001b[0m│\u001b[1m Name             \u001b[0m│\n",
       "├────┼───────────┼───────────┼──────────────────┤\n",
       "│  2 │  15.02 µs │   1.67 µs │ cudaGetLastError │\n",
       "│  3 │  19.79 µs │\u001b[31m 534.06 µs \u001b[0m│\u001b[1m cudaLaunchKernel \u001b[0m│\n",
       "│  4 │ 554.32 µs │ 238.42 ns │ cudaGetLastError │\n",
       "└────┴───────────┴───────────┴──────────────────┘\n",
       "\n",
       "Device-side activity: GPU was busy for 402.69 µs (42.08% of the trace)\n",
       "┌────┬───────────┬───────────┬─────────┬────────┬──────┬─────────────────────────────────────────────────────────┐\n",
       "│\u001b[1m ID \u001b[0m│\u001b[1m     Start \u001b[0m│\u001b[1m      Time \u001b[0m│\u001b[1m Threads \u001b[0m│\u001b[1m Blocks \u001b[0m│\u001b[1m Regs \u001b[0m│\u001b[1m Name                                                    \u001b[0m│\n",
       "├────┼───────────┼───────────┼─────────┼────────┼──────┼─────────────────────────────────────────────────────────┤\n",
       "│  3 │ 551.22 µs │\u001b[31m 402.69 µs \u001b[0m│     256 │  65536 │   32 │\u001b[1m _Z15axpy_kernel_valIffEv19cublasAxpyParamsValIT_S1_T0_E \u001b[0m│\n",
       "└────┴───────────┴───────────┴─────────┴────────┴──────┴─────────────────────────────────────────────────────────┘\n"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x = CUDA.rand(Float32, 4096, 4096)\n",
    "y = CUDA.rand(Float32, 4096, 4096)\n",
    "alpha = rand(Float32)\n",
    "CUBLAS.axpy!(length(y), alpha, x, y)   # warm-up\n",
    "CUDA.@profile trace=true CUBLAS.axpy!(length(y), alpha, x, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "664fa9f0",
   "metadata": {},
   "source": [
    "Note that despite using the most registers, CUBLAS is likely to be faster than both our naive `axpy!` kernel, and the `broadcast` implementation. However, the benchmark results here are not robust, and we'll look into properly benchmarking GPU code as part of a later codebook."
   ]
  },
  {
   "attachments": {
    "0426d175-5682-491d-a1f0-f112254628ce.png": {
     "image/png": "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"
    }
   },
   "cell_type": "markdown",
   "id": "70b46c59",
   "metadata": {},
   "source": [
    "Now, before moving on to the next notebook, be sure the kernel attached to this notebook, as Piz Daint only allows a single process to use the GPU:\n",
    "\n",
    "![image.png](attachment:0426d175-5682-491d-a1f0-f112254628ce.png)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Julia 1.10.0-rc1",
   "language": "julia",
   "name": "julia-1.10"
  },
  "language_info": {
   "file_extension": ".jl",
   "mimetype": "application/julia",
   "name": "julia",
   "version": "1.10.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}