產品介紹 > 中介軟體 > GPU平台轉換加速套件> HMPP

GPU 平台轉換加速套件 - HMPP

2012/01/30 - HMPP 3.0.0 now available
HMPP 3.0.0 is available only for Linux operating systems and NVIDIA GPU architecture (CUDA target clause only).
OpenCL target will be available in Q1 2012.
Windows platform will be available in Q2 2012.

What is New ?
Data mirroring
Data mirroring is a new mode of data management in HMPP. With data mirroring you can directly manage data from their host address instead of using named buffer.
See data mirroring directives for further details.

Parallel directive clause for multi-GPU
HMPP provides some mechanism to distribute data and dispatch computation on multiple HWAs according to the “owner compute rule”.
See the parallel HMPP directive for further details.

HMPP Runtime API for C++, C and Fortran languages
The HMPP Runtime API is provided to let end-user instrument directly their application. This can be an alternative to the usage of HMPP directives by providing HMPP functionalities at programmation level.

Data transfer modes
With HMPP 3 several transfer policies are available to manage data transfers between host and GPU. Atcall, atfirstcall, manual, automatic modes are available

Improved multi-file support
HMPP 3 support the introduction of directive set into different files of an application.

CUDA official support starts from version 4.0


2011/04/06 - HMPP 2.4.4 now available
What is New ?
Addition of the support of FORTRAN interface in HMPP
This functionality allows user to reference functions declared in FORTRAN interface into HMPP codelets.

Addition of a native support of the cuBLAS {s,d,c,z}gemm functions
This functionality allows user to easily replace their CPU computation by a direct call to a cuBLAS implementation. HMPP provides user with a native support of these functions just by adding codelet declaration on the prototype and by prefixing the subroutine call by a callsite HMPP directive.
Full examples can be provided on demand.

Support for Absoft’s Pro Fortran V11.1 Compiler
HMPP supports now Absoft Pro Fortran Compiler version 11.0 and 11.1 on Windows (already available for Linux platforms)

- Reduction of the delay between two HMPP CUDA kernel invocations
- More detailed messages for unsupported compilers
- Improvement of the HMPP compilation time

2010/11/15 - HMPP 2.4 is the latest release now available
This release new features and changes include:

- HMPP for Microsoft Windows
- Region programming simplification
- Automatic data transfer management
- HMPP Codelet Generator directives
 Introduction of the "global" and "none" clauses for loop nest gridification
 Introduction of the hmppcg gridify directive
- HMPP Plugin for Eclipse Galileo
- Integration with Vampir Profiling tool and Allinea DDT
- Support of ATI Stream 2.2
- Support of Absoft Pro Fortran Compiler version 11


2010/07/25 - HMPP Workbench Version 2.3.3

IMPORTANT: Please, note that due to the evolution of the license mechanism from HMPP 2.3.0, licenses for previous versions of HMPP need to be updated.

What is New?
- Possibility to allocate several codelets to the same GPU
- Support for NVIDIA CUDA Toolkit 3.1

Bug fix:
- Correction on memory leak in runtime messages and section transfers

- Increased performance for multi-dimensional arrays in CUDA backend

Possibility to allocate several codelets to the same GPU

GPU management may be disabled by using the environment variable HMPP_DEFAULT_DEVICE_OVERRIDE=true

This variable should be used in conjunction with the following environment variables: HMPP_CUDA_DEFAULT_DEVICE=#GPU: target CUDA, allows to specify which GPU to used.

HMPP_OPENCL_DEFAULT_DEVICE=#GPU: target OPENCL, allows to specify which GPU to used.

HMPP_CUDA_NO_THREAD_EXIt=1: disable the release of the GPU when several codelets are allocated in the same thread. In such situation, codelets or groups of codelets are allocated on a same device independently of the occupancy of the GPU.

Known issues: This list covers some of the known problems that will be fixed in future updates or releases:
- Using constantmemory pragma is currently not supported with CUDA SDK 3.1


2010/06/01 - NEW OPENCL BACK-END IN HMPP 2.3

CAPS is pleased to announce the availability of an OpenCL code generator within the just released 2.3 version of its HMPP directive-based hybrid compiler. Also, the CUDA back-end generator has been enhanced with Fermi capabilities and this new release brings support for more native compilers with Intel ifort/icc, GNU gcc/gfortran and PGI pgcc/pgfort compilers, enabling developers to freely use their favorite compiler with HMPP 2.3.

Based on GPU programming and tuning directives, HMPP offers an incremental programming model that allows developers with different levels of expertise to fully exploit GPU hardware accelerators in their legacy code.

As an emerging open programming standard, OpenCL back-end expands the portfolio of targets supported by HMPP to the AMD ATI GPUs. The OpenCL version of HMPP fully supports AMD and NVIDIA GPU compute processors, bringing to users a wider set of hybrid platforms they can execute their applications on. Recently released, the NVIDIA Tesla 200-series GPUs based on the ‘Fermi' codename new CUDA architecture is also supported by HMPP 2.3.

“The addition of this OpenCL back-end to our existing NVIDIA CUDA back-end is a major milestone in HMPP development that gives users another powerful standard programming option,” comments Stephane Bihan, Sales manager at CAPS, “We really look forward to demonstrating it at ISC'10 in Hamburg this week.”




   HMPP 產品介紹 
   為什麼要使用 HMPP
   HMPP 試用版申請