Parallel Thread Execution

Parallel Thread Execution (PTX, or NVPTX[1]) is a pseudo-assembly language used in Nvidia's CUDA programming environment. The nvcc compiler translates code written in CUDA, a C++-like language, into PTX, and the graphics driver contains a compiler which translates the PTX into a binary code[2] which can be run on the processing cores. The GNU Compiler Collection also has basic ability for PTX generation in the context of OpenMP offloading.[3] Inline PTX assembly can be used in CUDA.[4]

Registers

PTX uses an arbitrarily large register set; the output from the compiler is almost pure single-assignment form, with consecutive lines generally referring to consecutive registers. Programs start with declarations of the form

.reg .u32 %r<335>;            // declare 335 registers %r0, %r1, ..., %r334 of type unsigned 32-bit integer

It is a three-argument assembly language, and almost all instructions explicitly list the data type (in terms of sign and width) on which they operate. Register names are preceded with a % character and constants are literal, e.g.:

shr.u64 %rd14, %rd12, 32;     // shift right an unsigned 64-bit integer from %rd12 by 32 positions, result in %rd14
cvt.u64.u32 %rd142, %r112;    // convert an unsigned 32-bit integer to 64-bit

There are predicate registers, but compiled code in shader model 1.0 uses these only in conjunction with branch commands; the conditional branch is

@%p14 bra $label;             // branch to $label

The setp.cc.type instruction sets a predicate register to the result of comparing two registers of appropriate type, there is also a set instruction, where set.le.u32.u64 %r101, %rd12, %rd28 sets the 32-bit register %r101 to 0xffffffff if the 64-bit register %rd12 is less than or equal to the 64-bit register %rd28. Otherwise %r101 is set to 0x00000000.

There are a few predefined identifiers that denote pseudoregisters. Among others, %tid, %ntid, %ctaid, and %nctaid contain, respectively, thread indices, block dimensions, block indices, and grid dimensions.[5]

State spaces

Load (ld) and store (st) commands refer to one of several distinct state spaces (memory banks), e.g. ld.param. There are eight state spaces:[5]

  • .reg : registers
  • .sreg : special, read-only, platform-specific registers
  • .const : shared, read-only memory
  • .global : global memory, shared by all threads
  • .local : local memory, private to each thread
  • .param : parameters passed to the kernel
  • .shared : memory shared between threads in a block
  • .tex : global texture memory (deprecated)

Shared memory is declared in the PTX file via lines at the start of the form:

.shared .align 8 .b8 pbatch_cache[15744]; // define 15,744 bytes, aligned to an 8-byte boundary

Writing kernels in PTX requires explicitly registering PTX modules via the CUDA Driver API, typically more cumbersome than using the CUDA Runtime API and NVIDIA's CUDA compiler, nvcc. The GPU Ocelot project provided an API to register PTX modules alongside CUDA Runtime API kernel invocations, though the GPU Ocelot is no longer actively maintained.[6]

gollark: * response, not requesrt
gollark: There is a limit for request length of 15MB or so.
gollark: CC will let you read HTTP responses in chunks, I believe, and you could write in chunks too, although I don't think most programs actually *use* that.
gollark: You could rewrite it as a streaming thing if you *cared*, but this way was easier.
gollark: As much as the server has, that is.

See also

References
  1. "User Guide for NVPTX Back-end — LLVM 7 documentation". llvm.org.
  2. "CUDA Binary Utilities". docs.nvidia.com. Retrieved 2019-10-19.
  3. "nvptx". GCC Wiki.
  4. "Inline PTX Assembly in CUDA". docs.nvidia.com. Retrieved 2019-11-03.
  5. "PTX ISA Version 2.3" (PDF).
  6. "GPUOCelot: A dynamic compilation framework for PTX". github.com.
This article is issued from Wikipedia. The text is licensed under Creative Commons - Attribution - Sharealike. Additional terms may apply for the media files.