CUDA对数：nvprof以双精度检测单精度运算
[英]CUDA logarithm: nvprof detects single precision operations in double precision
问题描述
I'm computing "log(x)" in double precision in CUDA, but when I profile, it detects single precision operations using metric "flop_count_sp_special". 
我正在CUDA中以双精度计算“ log（x）”，但是当我分析时，它使用度量“ flop_count_sp_special”检测到单精度运算。
I'm compiling with "-arch=sm_30" to ensure compute capability 3.0 and double precision arithmetic, but I can't seem to find a way to ensure special functions are computed in double precision. 我正在使用“ -arch = sm_30”进行编译以确保具有3.0的计算能力和双精度算术，但是我似乎找不到一种方法来确保以双精度计算特殊函数。 Is this possible? 这可能吗？
1 个解决方案
解决方案1
4 已采纳 2018-08-31 18:14:37
It appears that the CUDA calculation of double-precision floating-point log does involve (a small number of) single precision floating point calculations. 似乎双精度浮点log的CUDA计算确实涉及（少量）单精度浮点计算。 I think this need not be cause for concern, in and of itself. 我认为这本身不必引起关注。
The GPU hardware does not support a native double-precision log calculation, therefore the compiler, upon encountering this, replaces the instruction with a sequence of operations (eg add, multiply, etc.) designed to perform the calculation. GPU硬件不支持本机双精度log计算，因此，编译器在遇到这种情况时，会使用旨在执行计算的一系列操作（例如加，乘等）替换指令。 We can confirm this with a simple example: 我们可以用一个简单的例子来确认这一点：
$ cat t288.cu
#include <math.h>
#include <stdio.h>
__global__ void kernel(double *y)
{
double x = *y;
*y = log(x);
}
int main(){
double *d_y, h_y = 10.0;
cudaMalloc(&d_y, sizeof(double));
cudaMemcpy(d_y, &h_y, sizeof(double), cudaMemcpyHostToDevice);
kernel<<<1,1>>>(d_y);
cudaDeviceSynchronize();
cudaMemcpy(&h_y, d_y, sizeof(double), cudaMemcpyDeviceToHost);
printf("val = %f\n", h_y);
}
$ nvcc -arch=sm_35 -o t288 t288.cu
$ cuda-memcheck ./t288
========= CUDA-MEMCHECK
val = 2.302585
========= ERROR SUMMARY: 0 errors
$ CUDA_VISIBLE_DEVICES="1" nvprof --metrics flop_count_dp,flop_count_sp,flop_count_sp_special ./t288
==14909== NVPROF is profiling process 14909, command: ./t288
==14909== Some kernel(s) will be replayed on device 0 in order to collect all events/metrics.
Replaying kernel "kernel(double*)" (done)
val = 2.302585rnal events
==14909== Profiling application: ./t288
==14909== Profiling result:
==14909== Metric result:
Invocations Metric Name Metric Description Min Max Avg
Device "Tesla K40m (0)"
Kernel: kernel(double*)
1 flop_count_dp Floating Point Operations(Double Precision) 44 44 44
1 flop_count_sp Floating Point Operations(Single Precision) 0 0 0
1 flop_count_sp_special Floating Point Operations(Single Precision Special) 1 1 1
$ cuobjdump -sass ./t288
Fatbin elf code:
================
arch = sm_35
code version = [1,7]
producer = <unknown>
host = linux
compile_size = 64bit
code for sm_35
Fatbin elf code:
================
arch = sm_35
code version = [1,7]
producer = cuda
host = linux
compile_size = 64bit
code for sm_35
Function : _Z6kernelPd
.headerflags @"EF_CUDA_SM35 EF_CUDA_PTX_SM(EF_CUDA_SM35)"
/* 0x088010fc10a010ac */
/*0008*/ MOV R1, c[0x0][0x44]; /* 0x64c03c00089c0006 */
/*0010*/ MOV R2, c[0x0][0x140]; /* 0x64c03c00281c000a */
/*0018*/ MOV R3, c[0x0][0x144]; /* 0x64c03c00289c000e */
/*0020*/ LD.E.64 R12, [R2]; /* 0xc5800000001c0830 */
/*0028*/ MOV32I R0, 0xfffffc01; /* 0x747ffffe009fc002 */
/*0030*/ ISETP.GT.AND P0, PT, R13, c[0x2][0x0], PT; /* 0x5b481c40001c341e */
/*0038*/ MOV R4, R12; /* 0xe4c03c00061c0012 */
/* 0x08ac1080a0a0a4ac */
/*0048*/ MOV R5, R13; /* 0xe4c03c00069c0016 */
/*0050*/ @!P0 DMUL R4, R4, 1.80143985094819840000e+16; /* 0xc400021a80201011 */
/*0058*/ @!P0 MOV R13, R5; /* 0xe4c03c0002a00036 */
/*0060*/ IADD32I R6, R13, -0x1; /* 0x407fffffff9c3419 */
/*0068*/ ISETP.LT.U32.AND P1, PT, R6, c[0x2][0x4], PT; /* 0x5b101c40009c183e */
/*0070*/ @!P0 MOV R12, R4; /* 0xe4c03c0002200032 */
/*0078*/ @!P0 MOV32I R0, 0xfffffbcb; /* 0x747ffffde5a3c002 */
/* 0x08b88010a4a010ac */
/*0088*/ @P1 BRA 0xc0; /* 0x120000001804003c */
/*0090*/ MOV32I R8, 0x0; /* 0x74000000001fc022 */
/*0098*/ MOV32I R9, 0x7ff00000; /* 0x743ff800001fc026 */
/*00a0*/ DFMA R8, R4, +INF , R8; /* 0xb38023ff801c1021 */
/*00a8*/ FCMP.NEU R12, R8, RZ, R5; /* 0xdd6814007f9c2032 */
/*00b0*/ FCMP.NEU R13, R9, -QNAN , R5; /* 0xbd6817ff801c2435 */
/*00b8*/ BRA 0x240; /* 0x12000000c01c003c */
/* 0x08a010b010a0a010 */
/*00c8*/ LOP32I.AND R4, R13, 0x800fffff; /* 0x204007ffff9c3410 */
/*00d0*/ IMAD.U32.U32.HI R0, R13, 0x1000, R0; /* 0xa2000008001c3401 */
/*00d8*/ LOP32I.OR R5, R4, 0x3ff00000; /* 0x211ff800001c1014 */
/*00e0*/ ISETP.LT.AND P0, PT, R5, c[0x2][0x8], PT; /* 0x5b181c40011c141e */
/*00e8*/ MOV R4, R12; /* 0xe4c03c00061c0012 */
/*00f0*/ @!P0 IADD32I R7, R5, -0x100000; /* 0x407ff8000020141d */
/*00f8*/ @!P0 IADD32I R0, R0, 0x1; /* 0x4000000000a00001 */
/* 0x08a010a01080a010 */
/*0108*/ @!P0 MOV R5, R7; /* 0xe4c03c0003a00016 */
/*0110*/ LOP32I.XOR R12, R0, 0x80000000; /* 0x22400000001c0030 */
/*0118*/ DADD R6, R4, 1; /* 0xc38001ff801c1019 */
/*0120*/ MOV R8, RZ; /* 0xe4c03c007f9c0022 */
/*0128*/ DADD R4, R4, -1; /* 0xcb8001ff801c1011 */
/*0130*/ MUFU.RCP64H R9, R7; /* 0x84000000031c1c26 */
/*0138*/ MOV32I R14, 0x8b7a8b04; /* 0x7445bd45821fc03a */
/* 0x08a01080a4a4a4a4 */
/*0148*/ DFMA R6, -R6, R8, c[0x2][0x10]; /* 0x9b882040021c181a */
/*0150*/ DFMA R6, R6, R6, R6; /* 0xdb801800031c181a */
/*0158*/ DFMA R6, R8, R6, R8; /* 0xdb802000031c201a */
/*0160*/ DMUL R8, R6, R4; /* 0xe4000000021c1822 */
/*0168*/ DFMA R8, R6, R4, R8; /* 0xdb802000021c1822 */
/*0170*/ MOV32I R15, 0x3ed0ee25; /* 0x741f6877129fc03e */
/*0178*/ MOV32I R13, 0x43300000; /* 0x74219800001fc036 */
/* 0x08a080a080a4a4a4 */
/*0188*/ DMUL R10, R8, R8; /* 0xe4000000041c202a */
/*0190*/ DFMA R14, R10, c[0x2][0x18], R14; /* 0x5b803840031c283a */
/*0198*/ DFMA R14, R10, R14, c[0x2][0x20]; /* 0x9b803840041c283a */
/*01a0*/ DFMA R14, R10, R14, c[0x2][0x28]; /* 0x9b803840051c283a */
/*01a8*/ DADD R16, R4, -R8; /* 0xe3810000041c1042 */
/*01b0*/ DFMA R14, R10, R14, c[0x2][0x30]; /* 0x9b803840061c283a */
/*01b8*/ DADD R18, R16, R16; /* 0xe3800000081c404a */
/* 0x088880948880a080 */
/*01c8*/ DFMA R14, R10, R14, c[0x2][0x38]; /* 0x9b803840071c283a */
/*01d0*/ DADD R12, R12, c[0x2][0x50]; /* 0x638000400a1c3032 */
/*01d8*/ DFMA R16, R10, R14, c[0x2][0x40]; /* 0x9b803840081c2842 */
/*01e0*/ DFMA R14, R12, c[0x2][0x58], R8; /* 0x5b8020400b1c303a */
/*01e8*/ DFMA R4, -R4, R8, R18; /* 0xdb884800041c1012 */
/*01f0*/ DFMA R16, R10, R16, c[0x2][0x48]; /* 0x9b804040091c2842 */
/*01f8*/ DFMA R18, -R12, c[0x2][0x58], R14; /* 0x5b8838400b1c304a */
/* 0x08aca4a4a4a08094 */
/*0208*/ DMUL R4, R6, R4; /* 0xe4000000021c1812 */
/*0210*/ DMUL R10, R10, R16; /* 0xe4000000081c282a */
/*0218*/ DADD R18, -R8, R18; /* 0xe3880000091c204a */
/*0220*/ DFMA R4, R8, R10, R4; /* 0xdb801000051c2012 */
/*0228*/ DADD R18, R4, -R18; /* 0xe3810000091c104a */
/*0230*/ DFMA R12, R12, c[0x2][0x60], R18; /* 0x5b8048400c1c3032 */
/*0238*/ DADD R12, R14, R12; /* 0xe3800000061c3832 */
/* 0x080000000000b810 */
/*0248*/ ST.E.64 [R2], R12; /* 0xe5800000001c0830 */
/*0250*/ EXIT; /* 0x18000000001c003c */
/*0258*/ BRA 0x258; /* 0x12007ffffc1c003c */
/*0260*/ NOP; /* 0x85800000001c3c02 */
/*0268*/ NOP; /* 0x85800000001c3c02 */
/*0270*/ NOP; /* 0x85800000001c3c02 */
/*0278*/ NOP; /* 0x85800000001c3c02 */
......................
Fatbin ptx code:
================
arch = sm_35
code version = [6,2]
producer = cuda
host = linux
compile_size = 64bit
compressed
$
First of all we see that the profiler reports this double precision log op using mostly double-precision calculation but also one single-precision "special" op. 首先，我们看到事件探查器报告此双精度log运算主要使用双精度计算，但也使用一个单精度“特殊”运算。
Looking at the SASS dump, we observe this single-precision instruction: 查看SASS转储，我们观察到以下单精度指令：
/*0130*/ MUFU.RCP64H R9, R7; /* 0x84000000031c1c26 */
(we can confirm these are single-precision by referring to the documentation ) （我们可以通过参考文档确认这些是单精度的）
Most of the algorithm appears to use double-precision floating point. 大多数算法似乎使用双精度浮点。 This particular step may be involved in some estimation that only requires single-precision. 此特定步骤可能涉及仅需要单精度的某些估计。
So it would appear to be "expected" that use of log may report some (nonzero) flop_count_sp_special metric. 因此，似乎“预期”使用log可能会报告某些（非零） flop_count_sp_special指标。
(Yes, the FCMP instructions are also single-precision, but they appear to be in a code path not taken, for this case.) （是的，FCMP指令也是单精度的，但在这种情况下，它们似乎位于未采用的代码路径中。）
声明:本站的技术帖子网页，遵循CC BY-SA 4.0协议，如果您需要转载，请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.