使用PTX在C ++ / CUDA程序中计算浮点运算的方法 [英] A Method of counting Floating Point Operations in a C++/CUDA Program using PTX

问题描述

我有一个有点大的CUDA应用程序，我需要计算获得的GFLOPs。
我正在寻找一种简单而且通用的计算浮点运算数的方法。

I have a somewhat large CUDA application and I need to calculate the attained GFLOPs. I'm looking for an easy and perhaps generic way of counting the number of floating point operations.

可以从生成的浮点运算计数PTX代码（如下所示），使用汇编语言中的预定义fpo的列表？基于代码，计数可以通用吗？例如， add.s32％r58，％r8，-2; 计为一个浮点运算？

Is it possible to count floating point operations from the generated PTX code (as shown below), using a list of predefined fpo in assembly language? Based on the code, can the counting be made generic? For example, does add.s32 %r58, %r8, -2; count as one floating point operation?

EXAMPLE：

BB3_2:
.loc 2 108 1
mov.u32         %r8, %r79;
setp.ge.s32     %p1, %r78, %r16;
setp.lt.s32     %p2, %r78, 0;
or.pred         %p3, %p2, %p1;
@%p3 bra        BB3_5;

add.s32         %r58, %r8, -2;
setp.lt.s32     %p4, %r58, 0;
setp.ge.s32     %p5, %r58, %r15;
or.pred         %p6, %p4, %p5;
@%p6 bra        BB3_5;

.loc 2 112 1
ld.global.u8    %rc1, [%rd17];
cvt.rn.f32.u8   %f11, %rc1;
mul.wide.u32    %rd12, %r80, 4;
add.s64         %rd13, %rd7, %rd12;
ld.local.f32    %f12, [%rd13];
fma.rn.f32      %f14, %f11, %f12, %f14;
.loc 2 113 1
add.f32         %f15, %f15, %f12;

或者有更简单的计数FPO的方法，这是浪费时间吗？

Or are there far simpler ways of counting FPOs and this is a waste of time?

推荐答案

计算FLOPS的最简单方法是使用CUDA profiler为您完成。通过选择已实现FLOPS 实验，您可以获得以下图表：

The easiest way to count FLOPS would be to have the CUDA profiler do it for you. By selecting the Achieved FLOPS experiment, you can get charts like this:

浮点操作图表显示您的内核执行的每种类型的浮点操作的计数。

The Floating Point Operations chart displays a count of each type of floating point operation executed by your kernel.

这篇关于使用PTX在C ++ / CUDA程序中计算浮点运算的方法的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！