Missing CUDA inline PTX constraint letter for 8 bit variables in order to disable L1 cache for 8 bit variable (bool)

StackOverflow https://stackoverflow.com/questions/14261881

Вопрос

INTRODUCTION

In this question we can learn how to disable L1 cache for one single variable. Here is the accepted answer:

As mentioned above you can use inline PTX, here is an example:

__device__ __inline__ double ld_gbl_cg(const double *addr) {
  double return_value;
  asm("ld.global.cg.f64 %0, [%1];" : "=d"(return_value) : "l"(addr));
  return return_value;
}

You can easily vary this by swapping .f64 for .f32 (float) or .s32 (int) etc., the constraint of return_value "=d" for "=d" (float) or "=r" (int) etc. Note that the last constraint before (addr) - "l" - denotes 64 bit addressing, if you are using 32 bit addressing, it should be "r".

Now, I however, I want to load a boolean (1 byte) not a floating point. So, I thought that I could do something like this (for architecture >=sm_20):

__device__ inline bool ld_gbl_cg(const bool* addr){
  bool return_value;
  asm("ld.global.cg.u8 %0, [%1];" : "=???"(return_value) : "l"(addr));
  return return_value;
}

, where "???" should be the appropriate constraint letter for a boolean, respectively for an 8 bit unsinged integer (From this question, i deduced this, since it is noted that for >=sm_20, "u8" is used for a boolean). Howevever, I cannot find an appropriate constraint letter in nvidias document "Using inline PTX Assembly in CUDA" (On page 6 are listed some constraint letters). So my question is:

QUESTION

  1. Is there any CUDA inline PTX constraint letter for any of the types:

    • boolean
    • unsigned 8 bit integer
    • or evtl 8 bit binary variable
  2. If not, what can I do in my case (explained in the introduction)? - Can the parameters "b0", "b1", etc shortly discussed here, be of help?

Thank you very much in advance for any help or comments!

UPDATE

I also need a store function reading from L2 cache instead of global memory - i.e. the store function that is complementary to the above ld_gbl_cg function (only once I have this function, I can completely verify that njuffa's answer works). My best guess based on njuffa's answer below would be:

__device__ __forceinline__ void st_gbl_cg (const bool *addr, bool t)
{
#if defined(__LP64__) || defined(_WIN64)
    asm ("st.global.cg.u8 [%0], %1;" : "=l"(addr) : "h"((short)t));
#else
    asm ("st.global.cg.u8 [%0], %1;" : "=r"(addr) : "h"((short)t));
#endif
}

However, the compiler gives the warning "parameter "addr" was set but never used" and the programm fails at runtime with an "unspecified launch failure". I also tried with .u16 instead of .u8, as I do not know to what exactly it refers. Yet the result is the same.

(Additional information) The following paragraph in the PTX 3.1 documentation, seems to be important for this question:

5.2.2 Restricted Use of Sub-Word Sizes The .u8, .s8, and .b8 instruction types are restricted to ld, st, and cvt instructions. The .f16 floating-point type is allowed only in conversions to and from .f32 and .f64 types. All floating-point instructions operate only on .f32 and .f64 types. For convenience, ld, st, and cvt instructions permit source and destination data operands to be wider than the instruction-type size, so that narrow values may be loaded, stored, and converted using regular-width registers. For example, 8-bit or 16-bit values may be held directly in 32-bit or 64-bit registers when being loaded, stored, or converted to other types and sizes.

Это было полезно?

Решение

According to the document "Using Inline PTX in CUDA", there is no constraint for byte-sized operands. Best I can tell, the closest you can get to the desired functionality is to move the data through an intermediate 'short'. This results in one additional SASS instruction for the conversion from 'short' to 'bool'.

__device__ __forceinline__ bool ld_gbl_cg (const bool *addr)
{
    short t;
#if defined(__LP64__) || defined(_WIN64)
    asm ("ld.global.cg.u8 %0, [%1];" : "=h"(t) : "l"(addr));
#else
    asm ("ld.global.cg.u8 %0, [%1];" : "=h"(t) : "r"(addr));
#endif
    return (bool)t;
}
Лицензировано под: CC-BY-SA с атрибуция
Не связан с StackOverflow
scroll top