《《Using_Inline_PTX_Assembly_In_CUDA》.pdf

USING INLINE PTX ASSEMBLY IN CUDA DA-05713-001_v01 | February 2011 Application Note DOCUMENT CHANGE HISTORY DA-05713-001_v01 Version Date Authors Description of Change 01 February 14, 2011 CUDA Initial release Using Inline PTX Assembly in CUDA DA-05713-001_v01 | ii TABLE OF CONTENTS Using Inline PTX Assembly in CUDA 4 Assembler (ASM) Statements 4 Parameters 4 Constraints 6 Pitfalls 7 Namespace Conflicts 7 Memory Space Conflicts 7 Incorrect Optimization 7 Incorrect PTX 8 Error Checking 8 Using Inline PTX Assembly in CUDA DA-05713-001_v01 | iii USING INLINE PTX ASSEMBLY IN CUDA The NVIDIA® CUDA™ programming environment provides a parallel thread execution (PTX) instruction set architecture (ISA) for using the GPU as a data-parallel computing device. For more information on the PTX ISA, refer to the latest version of the PTX ISA reference document ptx_isa_[version].pdf in the CUDA Toolkit doc folder. This application note describes how to inline PTX assembly language statements into CUDA code. ASSEMBLER (ASM) STATEMENTS Assembler statements, asm(), provide a way to insert arbitrary PTX code into your CUDA program. A simple example is: asm(membar.gl;); This inserts a PTX membar.gl into your generated PTX code at the point of the asm() statement. Parameters An asm() statement becomes more complicated, and more useful, when we pass values in and out of the asm. The basic syntax is as follows: asm(template-string : constraint(output) : constraint(input)); where you can have multiple input or output operands separated by commas. The template string contains PTX