Posted on December 7, 2020 by admin
Tag: gpu programming
Comparative performance analysis of Vulkan and CUDA programming model implementations for GPUs
https://core.ac.uk/reader/323473500
Posted on December 2, 2020 by admin
Posted on November 15, 2020 by admin
Posted on January 5, 2020 by admin
Vulkan vs. OpenGL API
A brief discussion of Vulkan vs. OpenGL and why I believe you should learn Vulkan over OpenGL going forward. I also talk briefly about OpenGL CUDA interop and the computer graphics industry as a whole. Enjoy!
Posted on January 3, 2020 by admin
NVIDIA Nsight Compute Tutorial #1 | Video Walkthrough (57+ minutes) | Cuda Education
A very basic overview of how to use NVIDIA Nsight Compute to optimize your CUDA kernel.
Posted on December 23, 2019 by admin
NVIDIA Nsight Systems Tutorial #1 | Video Walkthrough (48+ minutes)
A very basic video walkthrough of how to use Nsight Systems to help in optimizing your application. Nsight Systems is software from NVIDIA and is mainly intended to work with the NVIDIA graphics cards and the CUDA programming language. This […]
Posted on December 10, 2019 by admin
CUDA Graphs Tutorial | How to launch CUDA Graph by Stream Capture & Explicit API Method | Video Walkthrough (57+ min.) | Includes Source Code
Learn how to use CUDA Graphs to make your application run faster and more efficiently. This video walkthrough shows you how to create CUDA Graphs by the Stream Capture Method and the Explicit API method. It also includes source code.
Posted on July 18, 2019 by admin
CUDA Streams: Video Walkthrough (40 minutes) + Example CUDA C Code
A 40 minute video discussion about CUDA Streams, including example code.
Posted on July 6, 2019 by admin
CUDA Synchronization | syncthreads | cudaDeviceSynchronize | cudaEventRecord | Cuda Education
Donate
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 |
// // // // // //CUDA Synchronization | A quick overview of synchronization in CUDA | Cuda Education | Cuda Tutorial //Company: CUDA Education | cudaeducation.com | cudaeducation@gmail.com | Please donate at cudaeducation.com //YouTube Channel (please subscribe): https://www.youtube.com/channel/UCzpwNg0Ai8zCzbsEtozkfFQ //Twitter: @cudaeducation //Have questions? Comment on the YouTube channel https://www.youtube.com/channel/UCzpwNg0Ai8zCzbsEtozkfFQ or email cudaeducation@gmail.com //Donate: Visit cudaeducation.com to donate to the cause //DISCLAIMER: This code is for teaching purposes only! CUDA Education does not guarantee the accuracy of this code in any way. This code should not be used in a production or commercial environment. Any liabilities or loss resulting from the use of this code, in whole or in part, will not be the responsibility of CUDA Education. //All rights reserved. This code is the property of CUDA Education. Please contact CUDA Education at cudaeducation@gmail.com if you would like to use this code in any way, shape or form. //CODING ENVIRONMENT: //CUDA Toolkit 10.0 //Windows environment //Visual Studio 2017 Community Edition //nVidia GeForce 1050 ti Graphics Card //Compute Capability 6.5 //Install CUDA Toolkit cudaeducation.com/howtoprogramcuda //make sure to include all relevant libraries #include "cuda_runtime.h" #include "device_launch_parameters.h" #include <windows.h> #include <stdio.h> __global__ void warp_divergence_on_CUDA_GPU(unsigned int gpudummy) { //calculate the linear address of the thread. //this is the thread's index among all threads available in the x dimension. //the GPU only sees linear addresses. //for example, threadIdx.x might equal 0, but we are not sure for which block this threadIdx.x is in, so we have to calculate the linear address to know exactly which thread in which block we are talking about. //in other words, the first thread in block 1 will have a different linear address than the first thread in block 7. both threads will have threadIdx.x = 0, but they are completely different so we use the linear address to know where we REALLY are in the bigger scheme of things. //for more information on why this calculation is done, check out https://www.pelagos-consulting.com/wp-content/uploads/2017/08/CUDA_description.pdf unsigned int linear_id_of_thread = blockIdx.x * blockDim.x + threadIdx.x; //START thread warp divergence code //when you comment out the "IF ELSE" statement, you will see the performance increase drastically. this is because the warps are not split apart into various segments and it can process the full 32 threads in each cycle with no delays. //if you do not comment it out, performance is slow. this is because warps can only take a single path in each cycle, so if the "IF" path is taken on a given cycle, the threads that are supposed to take the "ELSE" path are left idle until the next cycle. obviously this wastes cycles, processing power and productivity. and that is why the processing time is longer. //the moral of the story is that conditional statements slow up processing for a given warp if there are different paths taken within the warp. __syncthreads(); if (linear_id_of_thread % 2 == 0) { __syncthreads(); gpudummy = gpudummy + 1; __syncthreads(); } else { __syncthreads(); gpudummy = gpudummy - 1; __syncthreads(); } __syncthreads(); //END thread warp divergence code //CUDA EDUCATION //Website: cudaeducation.com //Twitter: @cudaeducation //Email: cudaeducation@gmail.com //YouTube: Cuda Education | Please subscribe //Donate: Visit cudaeducation.com to donate to the cause } int main(int argc, char **argv) { //get the nVidia GPU running CUDA ready int cuda_education_device = 0; //set the device to be used for CUDA execution cudaSetDevice(cuda_education_device); //how many integers are going to be in your array? int my_array_size = 80000000; //our dummy variable just to keep the show going int dummy = 0; //print out the number of integers in your array printf("array size %d \n", my_array_size); //how many threads in each BLOCK will be launched in the CUDA kernel? int number_of_threads_in_each_block = 256; //set up the CUDA timer for tracking how long it takes the kernel on the device (GPU) to do its job //for more information on this, check out https://devblogs.nvidia.com/how-implement-performance-metrics-cuda-cc/ cudaEvent_t start, stop; cudaEventCreate(&start); cudaEventCreate(&stop); //START create variables for the CUDA kernel launch //create BLOCK variable and declare the number of threads that should be in it. //we DO NOT declare the number of blocks here, just the number of threads in each block //variable type is dim3, which basically means there are 3 dimensions (no need to discuss further here) dim3 BLOCK(number_of_threads_in_each_block, 1); //create grid variable and declare the number of BLOCKS in the grid. //the number of BLOCKS in the grid is usually a calculated number (based on the size of the array and the number of threads in each block -> you need enough threads to tackle all the elements in the array). //variable type is dim3, which basically means there are 3 dimensions (no need to discuss further here) dim3 cuda_education_grid((my_array_size + BLOCK.x - 1) / BLOCK.x, 1); //print out the values used for block and grid printf("number of blocks %d | number of threads in each block %d\n", cuda_education_grid.x, BLOCK.x); //END create variables for the CUDA kernel launch //make sure all preceding tasks on the device (GPU) has been completed before proceeding with new business. wouldn't want to mix ingredients from two completely different recipes now would you?!?! cudaDeviceSynchronize(); //start recording event cudaEventRecord(start); //launch the CUDA kernel //cuda_education_grid and BLOCK are used as launch parameters warp_divergence_on_CUDA_GPU << <cuda_education_grid, BLOCK >> > (dummy); //stop recording event cudaEventRecord(stop); //another check to make sure all preceding business on the device (GPU) is complete before proceeding //remember that by default the CPU will launch the kernel and immediately continue processing other instructions. //we use cudaDeviceSynchronize() to tell the CPU to halt further processing until the CUDA kernel has finished doing its business cudaDeviceSynchronize(); cudaEventSynchronize(stop); //calculate and print the amount of time it took the CUDA kernel to process float milliseconds = 0; cudaEventElapsedTime(&milliseconds, start, stop); printf("DEVICE elapsed time: %f milliseconds\n", milliseconds); //practice good housekeeping by resetting the device when you are done cudaDeviceReset(); //CUDA EDUCATION //Website: cudaeducation.com //Twitter: @cudaeducation //Email: cudaeducation@gmail.com //YouTube: Cuda Education //Donate: Visit cudaeducation.com to donate to the cause } |
Posted on January 24, 2019 by admin