rocBLAS User Guide

Contents:

• 1. Getting Started Guide for Linux
  • 1.1. Introduction
    • 1.1.1. Documentation Roadmap
  • 1.2. Prerequisites
  • 1.3. Installing pre-built packages
  • 1.4. Building and Installing rocBLAS
    • 1.4.1. Requirements
    • 1.4.2. Download rocBLAS
    • 1.4.3. Library Dependencies
    • 1.4.4. Build library dependencies + library
    • 1.4.5. Build library dependencies + client dependencies + library + client
    • 1.4.6. Build clients without library
    • 1.4.7. Use of Tensile
• 2. API Reference Guide
  • 2.1. Introduction
  • 2.2. rocBLAS API and Legacy BLAS functions
    • 2.2.1. Rules for obtaining rocBLAS API from Legacy BLAS
    • 2.2.2. Example Code
    • 2.2.3. LP64 interface
    • 2.2.4. Column-major storage and 1 based indexing
    • 2.2.5. Pointer mode
    • 2.2.6. Asynchronous API
    • 2.2.7. Complex Number Data Types
    • 2.2.8. MI100 (gfx908) Considerations
    • 2.2.9. MI200 (gfx90a) Considerations
  • 2.3. Using rocBLAS API
    • 2.3.1. rocBLAS Datatypes
      • 2.3.1.1. rocblas_handle
      • 2.3.1.2. rocblas_int
      • 2.3.1.3. rocblas_stride
      • 2.3.1.4. rocblas_half
      • 2.3.1.5. rocblas_bfloat16
      • 2.3.1.6. rocblas_float_complex
      • 2.3.1.7. rocblas_double_complex
    • 2.3.2. rocBLAS Enumeration
      • 2.3.2.1. rocblas_operation
      • 2.3.2.2. rocblas_fill
      • 2.3.2.3. rocblas_diagonal
      • 2.3.2.4. rocblas_side
      • 2.3.2.5. rocblas_status
      • 2.3.2.6. rocblas_datatype
      • 2.3.2.7. rocblas_pointer_mode
      • 2.3.2.8. rocblas_atomics_mode
      • 2.3.2.9. rocblas_layer_mode
      • 2.3.2.10. rocblas_gemm_algo
      • 2.3.2.11. rocblas_gemm_flags
    • 2.3.3. rocBLAS Helper functions
      • 2.3.3.1. Auxiliary Functions
      • 2.3.3.2. Device Memory Allocation Functions
      • 2.3.3.3. Build Information Functions
    • 2.3.4. rocBLAS Level-1 functions
      • 2.3.4.1. rocblas_iXamax + batched, strided_batched
      • 2.3.4.2. rocblas_iXamin + batched, strided_batched
      • 2.3.4.3. rocblas_Xasum + batched, strided_batched
      • 2.3.4.4. rocblas_Xaxpy + batched, strided_batched
      • 2.3.4.5. rocblas_Xcopy + batched, strided_batched
      • 2.3.4.6. rocblas_Xdot + batched, strided_batched
      • 2.3.4.7. rocblas_Xnrm2 + batched, strided_batched
      • 2.3.4.8. rocblas_Xrot + batched, strided_batched
      • 2.3.4.9. rocblas_Xrotg + batched, strided_batched
      • 2.3.4.10. rocblas_Xrotm + batched, strided_batched
      • 2.3.4.11. rocblas_Xrotmg + batched, strided_batched
      • 2.3.4.12. rocblas_Xscal + batched, strided_batched
      • 2.3.4.13. rocblas_Xswap + batched, strided_batched
    • 2.3.5. rocBLAS Level-2 functions
      • 2.3.5.1. rocblas_Xgbmv + batched, strided_batched
      • 2.3.5.2. rocblas_Xgemv + batched, strided_batched
      • 2.3.5.3. rocblas_Xger + batched, strided_batched
      • 2.3.5.4. rocblas_Xsbmv + batched, strided_batched
      • 2.3.5.5. rocblas_Xspmv + batched, strided_batched
      • 2.3.5.6. rocblas_Xspr + batched, strided_batched
      • 2.3.5.7. rocblas_Xspr2 + batched, strided_batched
      • 2.3.5.8. rocblas_Xsymv + batched, strided_batched
      • 2.3.5.9. rocblas_Xsyr + batched, strided_batched
      • 2.3.5.10. rocblas_Xsyr2 + batched, strided_batched
      • 2.3.5.11. rocblas_Xtbmv + batched, strided_batched
      • 2.3.5.12. rocblas_Xtbsv + batched, strided_batched
      • 2.3.5.13. rocblas_Xtpmv + batched, strided_batched
      • 2.3.5.14. rocblas_Xtpsv + batched, strided_batched
      • 2.3.5.15. rocblas_Xtrmv + batched, strided_batched
      • 2.3.5.16. rocblas_Xtrsv + batched, strided_batched
      • 2.3.5.17. rocblas_Xhemv + batched, strided_batched
      • 2.3.5.18. rocblas_Xhbmv + batched, strided_batched
      • 2.3.5.19. rocblas_Xhpmv + batched, strided_batched
      • 2.3.5.20. rocblas_Xher + batched, strided_batched
      • 2.3.5.21. rocblas_Xher2 + batched, strided_batched
      • 2.3.5.22. rocblas_Xhpr + batched, strided_batched
      • 2.3.5.23. rocblas_Xhpr2 + batched, strided_batched
    • 2.3.6. rocBLAS Level-3 functions
      • 2.3.6.1. rocblas_Xgemm + batched, strided_batched
      • 2.3.6.2. rocblas_Xsymm + batched, strided_batched
      • 2.3.6.3. rocblas_Xsyrk + batched, strided_batched
      • 2.3.6.4. rocblas_Xsyr2k + batched, strided_batched
      • 2.3.6.5. rocblas_Xsyrkx + batched, strided_batched
      • 2.3.6.6. rocblas_Xtrmm + batched, strided_batched
      • 2.3.6.7. rocblas_Xtrsm + batched, strided_batched
      • 2.3.6.8. rocblas_Xhemm + batched, strided_batched
      • 2.3.6.9. rocblas_Xherk + batched, strided_batched
      • 2.3.6.10. rocblas_Xher2k + batched, strided_batched
      • 2.3.6.11. rocblas_Xherkx + batched, strided_batched
      • 2.3.6.12. rocblas_Xtrtri + batched, strided_batched
    • 2.3.7. rocBLAS Extension
      • 2.3.7.1. rocblas_axpy_ex + batched, strided_batched
      • 2.3.7.2. rocblas_dot_ex + batched, strided_batched
      • 2.3.7.3. rocblas_dotc_ex + batched, strided_batched
      • 2.3.7.4. rocblas_nrm2_ex + batched, strided_batched
      • 2.3.7.5. rocblas_rot_ex + batched, strided_batched
      • 2.3.7.6. rocblas_scal_ex + batched, strided_batched
      • 2.3.7.7. rocblas_gemm_ex + batched, strided_batched
      • 2.3.7.8. rocblas_gemm_ext2
      • 2.3.7.9. rocblas_trsm_ex + batched, strided_batched
      • 2.3.7.10. rocblas_Xgeam + batched, strided_batched
      • 2.3.7.11. rocblas_Xdgmm + batched, strided_batched
    • 2.3.8. Device Memory Allocation in rocBLAS
      • 2.3.8.1. Environment Variable for Preallocating
      • 2.3.8.2. Functions for manually setting memory size
      • 2.3.8.3. Function for setting user owned workspace
      • 2.3.8.4. Functions for finding how much memory is required
      • 2.3.8.5. rocBLAS Function Return Values for insufficient device memory
    • 2.3.9. Logging in rocBLAS
    • 2.3.10. References
• 3. Programmer’s Guide
  • 3.1. Library Source Code Organization
    • 3.1.1. The library directory
      • 3.1.1.1. library/include
      • 3.1.1.2. library/src/blas[1,2,3]
      • 3.1.1.3. library/src/blas3/Tensile
      • 3.1.1.4. library/src/blas_ex
      • 3.1.1.5. library/src/include
    • 3.1.2. The clients directory
      • 3.1.2.1. clients/gtest
      • 3.1.2.2. clients/benchmarks
      • 3.1.2.3. clients/include
      • 3.1.2.4. clients/common
      • 3.1.2.5. clients/samples
    • 3.1.3. Infrastructure
  • 3.2. Handle, Stream, and Device Management
    • 3.2.1. Handle
    • 3.2.2. Stream and Device Management
    • 3.2.3. Multiple streams and Multiple devices
  • 3.3. Device Memory Allocation
    • 3.3.1. Requirements
    • 3.3.2. Design
    • 3.3.3. Answering Device Memory Size Queries In Function That Needs Memory
      • 3.3.3.1. Example
      • 3.3.3.2. Function
      • 3.3.3.3. Function
      • 3.3.3.4. Function
    • 3.3.4. Answering Device Memory Size Queries In Function That Does Not Needs Memory
      • 3.3.4.1. Example
      • 3.3.4.2. Macro
    • 3.3.5. rocBLAS Kernel Device Memory Allocation
      • 3.3.5.1. Example
      • 3.3.5.2. Example
      • 3.3.5.3. Function
    • 3.3.6. Performance degrade
      • 3.3.6.1. Example
  • 3.4. Thread Safe Logging
  • 3.5. rocBLAS Numerical Checking
  • 3.6. rocBLAS order of argument checking and logging
    • 3.6.1. Legacy BLAS
    • 3.6.2. rocBLAS
    • 3.6.3. rocBLAS has the following differences when compared to Legacy BLAS
    • 3.6.4. To accommodate the additions
    • 3.6.5. Device Memory Size Queries
      • 3.6.5.1. Logging
    • 3.6.6. rocBLAS control flow
    • 3.6.7. Legacy L1 BLAS “single vector”
    • 3.6.8. Legacy L1 BLAS “two vector”
    • 3.6.9. Legacy L2 BLAS
    • 3.6.10. Legacy L3 BLAS
  • 3.7. rocBLAS Benchmarking & Testing
    • 3.7.1. rocblas-bench
    • 3.7.2. rocblas-test
    • 3.7.3. Add new rocBLAS unit test
• 4. Contributor’s Guide
  • 4.1. License Agreement
  • 4.2. Pull-request guidelines
  • 4.3. StyleGuide
    • 4.3.1. Interface
    • 4.3.2. Philosophy
    • 4.3.3. Implementation
    • 4.3.4. Format
  • 4.4. Coding Guidelines
  • 4.5. Static Code Analysis
• 5. Acknowledgement
• 6. Disclaimer