【DSP开发】【并行计算-CUDA开发】TI OpenCL v01.01.xx

TI OpenCL v01.01.xx

TI OpenCL™ Runtime Documentation Contents:

• Introduction
• OpenCL 1.1 Reference Material
• Compilation
  • Compile Host OpenCL Applications
  • Compiling OpenCL C Programs
    • Create an OpenCL program from source, with embedded source
    • Create an OpenCL program from source, with source in a file
    • Create an OpenCL program from binary, with binary in a file
    • Create an OpenCL program from binary, with embedded binary
  • Caching on-line compilation results
  • The TI off-line OpenCL C compiler: clocl
• Memory Usage
  • Device Memory
    • Caching
  • How DDR3 is Partitioned for Linux System and OpenCL
    • 66AK2x
    • AM57
    • Changing DDR3 Partition for OpenCL
  • Alternate Host malloc/free Extension for Zero Copy OpenCL Kernels
  • The OpenCL Memory Model
  • OpenCL Buffers
    • Global Buffers
    • Local Buffers
    • Sub-Buffers
  • Buffer Read/Write vs. Map/Unmap
  • Discovering OpenCL Memory Sizes and Limits
  • Cache Operations
  • Large OpenCL buffers and Memory Beyond the 32-bit DSP Address Space
    • Large Buffer Use Cases
  • User Defined DSP Heap Extension
    • User Defined DSP Heap Built-in Functions
    • Allocation of the Underlying Memory for User Defined DSP Heaps
    • Putting it all Together
• Execution Model
  • Terminology
  • Device Discovery
  • Understanding Kernels, Work-groups and Work-items
    • Enqueueing a Kernel
    • Mapping the OpenCL C work-item Built-in Functions
    • OpenCL C Kernel Code
    • NDRangeKernel Execution on DSP Devices
• Extensions
  • Calling Standard C Code From OpenCL C Code
  • Calling Standard C code with OpenMP from OpenCL C code
    • OpenMP dispatch from OpenCL
  • C66x standard C compiler intrinsic functions
  • OpenCL C code using printf
  • DMA Control Using EdmaMgr Functions
    • Single Transfer EdmaMgr APIs
    • Multiple Transfer EdmaMgr APIs
  • Using Extended Memory on the 66AK2x device
  • Fast Global buffers in on-chip MSMC memory
  • OpenCL C Builtin Function Extensions
  • Cache Operations
• Environment Variables
• Optimization Tips
  • Optimization Techniques for Host Code
    • Use Off-line, Embedded Compilation Model
    • Avoid the read/write Buffer model on shared memory SoC platforms
    • Use MSMC Buffers Whenever Possible
    • Dispatch Appropriate Compute Loads
    • Prefer Kernels with 1 work-item per work-group
  • Optimization Techniques for Device (DSP) Code
    • Prefer Kernels with 1 work-item per work-group
    • Use Local Buffers
    • Use async_work_group_copy and async_work_group_strided_copy
    • Avoid DSP writes directly to DDR
    • Use the reqd_work_group_size attribute on kernels
    • Use the TI OpenCL extension than allows Standard C code to be called from OpenCL C code
    • Avoid OpenCL C Barriers
    • Use the most efficient data type on the DSP
    • Do Not Use Large Vector Types
    • Consecutive memory accesses
    • Prefer the CPU style of writing OpenCL code over the GPU style
  • Typical Steps to Optimize Device Code
  • Optimizing 3x3 Gaussian smoothing filter
    • Overview of Gaussian Filter
    • Natural C Code
    • Optimizing for DSP
    • Performance Improvement
  • Performance Data
• Examples
  • Building and Running
  • Example Descriptions
    • platforms example
    • simple example
    • mandelbrot, mandelbrot_native examples
    • ccode example
    • matmpy example
    • offline example
    • vecadd_openmp example
    • vecadd_openmp_t example
    • vecadd example
    • vecadd_mpax example
    • vecadd_mpax_openmp example
    • dsplib_fft example
    • ooo, ooo_map examples
    • null example
    • sgemm example
    • dgemm example
    • edmamgr example
    • dspheap example
  • Float compute example
    • Host Code (main.cpp)
    • OpenCL C kernel code (dsp_compute.cl)
    • Sample Output
  • Monte Carlo example
    • Algorithm for Gaussian Random Number Generation
    • Executing the code
    • Sample Output
• Debug
  • Debug with printf
    • Host side OpenCL application code
    • DSP side OpenCL kernel code
  • Debug with gdb
    • Host side gdb
    • DSP side debug with host side client gdbc6x
  • Debug with CCS
    • Connect emulator to EVM and CCS
    • Debug DSP side code with CCS
  • Debug with dsptop
• Profiling
  • Host Side Profiling
  • DSP Side Profiling
• OpenCL on TI-RTOS
  • Overview
    • OpenCL on RTOS Package
    • Running Examples Shipped with OpenCL Package
  • Basic OpenCL RTOS Application Development
    • Building Application on Linux
    • Building Application on Windows
    • Creating an OpenCL RTOS Application
    • Limited Customization: Participating DSP Core(s)
    • Differences from OpenCL Linux (Host running Linux)
  • Advanced OpenCL RTOS Application Development
• Frequently Asked Questions
  • How do I get support for TI OpenCL products?
  • Which TI OpenCL Version is Installed?
  • Using Python OpenCL with the TI OpenCL implementation
  • Guidelines for porting Stand-alone DSP applications to OpenCL
    • Heap Memory Management
    • Stack Usage
    • Boot Routine Dependencies
    • Linker Command Files
  • OpenCL Interoperability with Host OpenMP
  • MCSDK-HPC to OpenCL Component Version Map
  • Does TI’s OpenCL support images and samplers?
  • Why does the OpenCL ICD installed on my platform not find the TI OpenCL implementation?
  • Why do I get messages about /var/lock/opencl when running OpenCL applications?
  • Why do I get DLOAD error messages when running OpenCL applications?
  • How do I limit log file sizes on EVM’s temporary file storage (tmpfs)?
    • 66AK2* EVMs
    • AM57* EVMs
• Readme
  • OpenCL v01.01.09.x Readme
    • Platforms supported
    • Release Notes
    • Compiler Versions
  • OpenCL v01.01.08.x Readme
    • Platforms supported
    • Release Notes
    • Compiler Versions
  • OpenCL v01.01.07.x Readme
    • Platforms supported
    • Release Notes
    • Compiler Versions
• Disclaimer
• Important Notice