attic/i5-roofline
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity

Init

Browse source
This commit is contained in:
Armin Friedl 2020-09-03 18:49:50 +02:00
commit c61491e713
5 changed files with 1237 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

120
Makefile Normal file
View file

@ -0,0 +1,120 @@
all: clean bin lib
# Roofline Binary
## This is the least demanding target, use it if nothing else works
nofancy: roofline roofline_o3
mkdir bin
mv $^ bin
## Your processor needs an FMA unit for this target to work
fmacap: roofline roofline_o3 roofline_fma roofline_fma_o3 roofline_fma_fast_o3 roofline_fma_fast_o2 roofline_fma_fast_fastmath_o3
mkdir bin
mv $^ bin
## This will compile just everything
bin: roofline roofline_o3 roofline_fma roofline_fma_o3 roofline_fma_fast_o3 roofline_fma_fast_o2 roofline_fma_fast_fastmath_o3 roofline_full roofline_profile roofline_full_clang roofline_full_manpack
mkdir bin
mv $^ bin
roofline: roofline.c aikern.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_o3: roofline.c aikern_o3.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_fma: roofline.c aikern_fma.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_fma_o3: roofline.c aikern_fma_o3.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_fma_fast_o3: roofline.c aikern_fma_fast_o3.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_fma_fast_o2: roofline.c aikern_fma_fast_o2.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_fma_fast_fastmath_o3: roofline.c aikern_fma_fast_fastmath_o3.a
gcc -Wall -Wextra -std=c99 -fopenmp $^ -o $@
roofline_full: roofline.c aikern_full.a
gcc -Wall -Wextra -std=c99 -fopenmp -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -malign-double -march=native $^ -o $@
roofline_full_clang: roofline.c aikern_full_clang.a
clang -Wall -Wextra -std=c99 -fopenmp -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -march=native $^ -o $@
roofline_full_manpack: roofline.c aikern_full_manpack.a
gcc -Wall -Wextra -std=c99 -fopenmp -DINTRINS -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -malign-double -march=native $^ -o $@
roofline_profile: roofline.c aikern_profile.a
gcc -Wall -Wextra -std=c99 -fopenmp -O0 -g -pg $^ -o $@
# Static Libraries
lib: aikern.a aikern_o3.a aikern_fma.a aikern_fma_o3.a aikern_fma_fast_o2.a aikern_fma_fast_o3.a aikern_fma_fast_fastmath_o3.a aikern_full.a aikern_full_clang.a aikern_profile.a aikern_full_manpack.a
mkdir lib
mv $^ lib
aikern.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -fopenmp -c -o aikern.o $<
ar rcs aikern.a aikern.o
rm aikern.o
aikern_o3.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -fopenmp -c -o aikern_o3.o $<
ar rcs $@ aikern_o3.o
rm aikern_o3.o
aikern_fma.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O2 -mavx -mfma -fopenmp -c -o aikern_fma.o $<
ar rcs $@ aikern_fma.o
rm aikern_fma.o
aikern_fma_o3.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -mavx -mfma -fopenmp -c -o aikern_fma_o3.o $<
ar rcs $@ aikern_fma_o3.o
rm aikern_fma_o3.o
aikern_fma_fast_o2.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O2 -mavx -mfma -fopenmp -Ofast -c -o aikern_fma_fast_o2.o $<
ar rcs $@ aikern_fma_fast_o2.o
rm aikern_fma_fast_o2.o
aikern_fma_fast_o3.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -mavx -mfma -fopenmp -Ofast -c -o aikern_fma_fast_o3.o $<
ar rcs $@ aikern_fma_fast_o3.o
rm aikern_fma_fast_o3.o
aikern_fma_fast_fastmath_o3.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -c -o aikern_fma_fast_fastmath_o3.o $<
ar rcs $@ aikern_fma_fast_fastmath_o3.o
rm aikern_fma_fast_fastmath_o3.o
aikern_full.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -malign-double -march=native -c -o aikern_full.o $<
ar rcs $@ aikern_full.o
rm aikern_full.o
aikern_full_clang.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -march=native -c -o aikern_full_clang.o $<
ar rcs $@ aikern_full_clang.o
rm aikern_full_clang.o
aikern_profile.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -fopenmp -O0 -g -pg -c -o aikern_profile.o $<
ar rcs $@ aikern_profile.o
rm aikern_profile.o
aikern_full_manpack.a: aikern.c aikern.h
gcc -Wall -Wextra -Wno-unused -DINTRINS -O3 -mavx -mfma -fopenmp -Ofast -ffast-math -malign-double -march=native -c -o aikern_full_manpack.o $<
ar rcs $@ aikern_full_manpack.o
rm aikern_full_manpack.o
# Cleanup
clean:
rm -f *.a
rm -f *.o
rm -f roofline roofline_o3 roofline_fma roofline_fma_o3 roofline_fma_fast_o3 roofline_fma_fast_o2 roofline_fma_fast_fastmath_o3 roofline_fma_fast_fastmath_aligned_o3 roofline_profile roofline_full_clang
rm -fR bin
rm -fR lib

288
aikern.c Normal file
View file

@ -0,0 +1,288 @@
# include <stdlib.h>
# include <stdio.h>
# include <unistd.h>
# include <stdarg.h>
# include <errno.h>
# include <string.h>
# include <sys/time.h>
# include "aikern.h"
/* === Macros === */
#ifdef ENDEBUG
#define DEBUG(...) do { fprintf(stderr, __VA_ARGS__); fprintf(stderr, "\n"); } while(0)
#else
#define DEBUG(...)
#endif
/**
* @brief terminate program on program error
* @param msg additional message to print
* @param ret exit value
*/
static void bail_out(char* fmt, ...);
/**
* @brief microseconds since epoch
*/
static double pin_time(void);
kern_result kernel_dispatch(kernel_t kernel,
double* a, double* b, double* c,
size_t size, size_t runs)
{
kern_result result = {0};
result.runs = runs;
result.starts = malloc(sizeof(double)*(runs));
result.ends = malloc(sizeof(double)*(runs));
result.size = size;
if(result.starts==NULL || result.ends==NULL)
{
bail_out("One of the mallocs failed\n. starts = %p, ends=%p", result.starts, result.ends);
}
switch(kernel)
{
case SIMPLE_1_16:
result.flops = 1;
result.kern_name = "Simple 1/16";
for(size_t r=0; r<runs; r++)
{
result.starts[r] = pin_time();
kernel_1_16_simple(a, size);
result.ends[r] = pin_time();
}
break;
case FMA_1_16:
result.flops = 2;
result.kern_name = "FMA aware 1/16";
for(size_t r=0; r<runs; r++)
{
result.starts[r] = pin_time();
kernel_1_16_fuseaware(a, b, c, size);
result.ends[r] = pin_time();
}
break;
case SIMPLE_8_1:
result.flops = 128;
result.kern_name = "Simple 8";
for(size_t r=0; r<runs; r++)
{
result.starts[r] = pin_time();
kernel_8_1_simple(a, size);
result.ends[r] = pin_time();
}
break;
case FMA_8_1:
result.flops = 128;
result.kern_name = "FMA aware 8";
for(size_t r=0; r<runs; r++)
{
result.starts[r] = pin_time();
kernel_8_1_fuseaware(a, size);
result.ends[r] = pin_time();
}
break;
case SIMPLE_8_1_FASTMATH:
DEBUG("AIKERN MANPACK");
result.flops = 128;
result.kern_name = "Simple 8 undermining fastmath";
for(size_t r=0; r<runs; r++)
{
result.starts[r] = pin_time();
kernel_8_1_simple_fastmath(a, size);
result.ends[r] = pin_time();
}
break;
case FMA_8_1_MANPACK:
DEBUG("AIKERN MANPACK");
#ifdef INTRINS
if(size%4 != 0)
{
bail_out("Must use multiple of 4 size for manpack");
}
result.flops = 128;
result.kern_name = "FMA aware 8 with manual packing";
for(size_t r=0; r<runs; r++)
{
DEBUG("running manpack run %zu",r);
result.starts[r] = pin_time();
kernel_8_1_fuseaware_manpack(a, size);
result.ends[r] = pin_time();
}
#endif
break;
default:
bail_out("No such kernel %s", kernel);
break;
}
return result;
}
inline void kernel_1_16_simple(double* a, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
a[i] = a[i] * a[i];
}
}
inline void kernel_1_16_fuseaware(double* a, double* b, double* c, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
a[i] = a[i] * b[i] + c[i];
}
}
inline void kernel_8_1_simple(double* a, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
a[i] = REP100(a[i]*)
REP20(a[i]*)
REP8(a[i]*)
REP1(a[i]);
}
}
inline void kernel_8_1_simple_fastmath(double* a, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
REP100(a[i]=a[i]*a[i];);
REP20(a[i]=a[i]*a[i];);
REP8(a[i]=a[i]*a[i];);
}
}
inline void kernel_8_1_fuseaware(double* a, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
REP60(a[i] = a[i] * a[i] + a[i];)
REP4(a[i] = a[i] * a[i] + a[i];)
}
}
#ifdef INTRINS
#include <immintrin.h>
inline void kernel_8_1_fuseaware_manpack(double* a, size_t size)
{
#pragma omp parallel for
for(size_t i=0; i<(size-4); i+=4)
{
// pack doubles
__m256d packvec = _mm256_set_pd(a[i], a[i+1], a[i+2], a[i+3]);
REP60(packvec = _mm256_fmadd_pd(packvec, packvec, packvec););
REP4(packvec = _mm256_fmadd_pd(packvec, packvec, packvec););
a[i] = packvec[0];
a[i+1] = packvec[1];
a[i+2] = packvec[2];
a[i+3] = packvec[3];
}
}
#endif /* INTRINS */
/********************************************
* Kernels which potentially compile to *
* different operational intensities than *
* specified *
********************************************/
void kernel_1_16_simple_dangerous(double* a, double* b, size_t size)
{
register volatile double tmp = 0.1;
#pragma omp parallel for
for(size_t i=0; i<size; i++){
tmp = a[i] * b[i];
}
}
void kernel_8_1_simple_dangerous(double* a, size_t size)
{
register volatile double tmp = 0.1;
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
tmp = a[i] * a[i] * a[i] * a[i] *
a[i] * a[i] * a[i] * a[i];
}
}
void kernel_1_8_vo_dangerous(double* a, size_t size)
{
register volatile double tmp=0.0;
#pragma omp parallel for
for(size_t i=0; i<size; i++)
{
tmp = a[i] * a[i];
}
}
/* === Helper Functions === */
static double pin_time(void)
{
struct timeval tp;
int i;
i = gettimeofday(&tp,NULL);
if(i != 0)
{
bail_out("Time measurement impossible. gettimeofday error");
}
return ( (double) tp.tv_sec + (double) tp.tv_usec * 1.e-6 );
}
static void bail_out(char* fmt, ...)
{
char* prog_name = "aikern";
if(fmt != NULL)
{
char msgbuf[150];
va_list vl;
va_start(vl, fmt);
if(vsnprintf(msgbuf, sizeof(msgbuf), fmt, vl) < 0)
msgbuf[0] = '\0';
va_end( vl);
if(strlen(msgbuf) > 0)
(void)fprintf(stderr, "%s: %s \n", prog_name, msgbuf);
}
if(errno != 0)
(void)fprintf(stderr, "%s: %s\n", prog_name, strerror(errno));
exit(EXIT_FAILURE);
}

288
aikern.h Normal file
View file

@ -0,0 +1,288 @@
#ifndef AIKERN_H
#define AIKERN_H
typedef struct {
size_t runs; // also # of start-/endtimes
double* starts; // starttimes
double* ends; // endtimes
int flops; // flops per iteration
char* kern_name;
size_t size; // size of arrays handeld
} kern_result;
typedef enum {
SIMPLE_1_16, FMA_1_16, SIMPLE_8_1, FMA_8_1, SIMPLE_8_1_FASTMATH, FMA_8_1_MANPACK
} kernel_t;
/**
* @brief main entry point. Dispatches the kernel calls
* @param kernel the kernel to run
* @param a An array with double values of size param size
* @param b An array with double values of size param size
* @param c An array with double values of size param size
* @param size The size of the arrays
* @param runs How often the kernel should be executed
* @return kern_result containing information about the kernel execution
*
*
*/
kern_result kernel_dispatch(kernel_t kernel,
double* a, double* b, double* c,
size_t size, size_t runs);
/**
* @brief A simple 1/16 operational intensity kernel
* @param a An array with double values of size param size
* @param size Size of the three param arrays
* @param result Pointer to result storage
*
* === Warning ===
* Don't use with -O0: Stores everything on stack
*
* === Description ===
* Uses a simple floating point operation: a[i] = a[i] * a[i];
*
* Runs in a parallelized for loop.
*
* === Analysis ===
* COMM: 1 read (8 byte), 1 write = 16 bytes
* COMP: 1 FLOP
* ---------
* OI: 1/16
*
* === Optimization ===
* Nothing special
*
*/
void kernel_1_16_simple(double* a, size_t size);
/**
* @brief A 1/16 operational intensity kernel utilizing FMA
* @param a An array with double values of size param size
* @param b An array with double values of size param size
* @param c An array with double values of size param size
* @param size Size of the three param arrays
* @param result Pointer to result storage
*
* === Warning ===
* This is dangerous if FMA is not used/can't be used. Then there
* are intermediary writes (and reads) to the stack.
*
* === Description ===
* Uses a triad function: a[i] = a[i] * b[i] + c[i]; in order
* to utilize the FMA unit.
*
* Runs in a parallelized for loop.
*
* === Analysis ===
* With gcc -O2 -mavx -mfma FMA compiles to:
* vmovsd xmm0,QWORD PTR [rdi+rax*8] # 1 read (8 byte)
* vmovsd xmm1,QWORD PTR [rdx+rax*8] # 1 read
* vfmadd132sd xmm0,xmm1,QWORD PTR [rsi+rax*8] # 2 FLOPs + 1 read
* vmovsd QWORD PTR [rdi+rax*8],xmm0 # 1 write
* --------
* 1/16 OI
*
* === Optimization ===
* For packed doubles compile with -Ofast
*
*/
void kernel_1_16_fuseaware(double* a, double* b, double* c, size_t size);
/**
* @brief A simple 8/1 operational intensity kernel
* @param a An array with double values of size param size
* @param size Size of the three param arrays
* @param result Pointer to result storage
*
* === Warning ===
* Don't use with -O0: Stores everything on stack
*
* === Description ===
* Uses a simple floating point operation: a[i] = a[i] * a[i] * ...* a[i];
*
* Runs in a parallelized for loop.
*
* === Analysis ===
* With AVX and -O2 (not necessarily FMA) best results (obviously correct
* easy to read disassembly).
*
* With gcc -O2 -mavx compiles to:
* vmovsd xmm1,QWORD PTR [rdi] # 1 read
* vmulsd xmm0,xmm1,xmm1 # 1 FLOP+register shuffling
* vmulsd xmm0,xmm0,xmm1 # 127x 1 FLOP+register shuffling
* # [...]
* vmovsd QWORD PTR [rdi-0x8],xmm0 # 1 write
* --------
* 128/16 = 8/1 OI
*
* === Optimization ===
* Nothing special
*/
void kernel_8_1_simple(double* a, size_t size);
/**
* @brief A 8/1 operational intensity kernel utilizing FMA
* @param a An array with double values of size param size
* @param size Size of the three param arrays
* @param result Pointer to result storage
*
* === Warning ===
* This is dangerous if FMA is not used/can't be used. Then there
* are intermediary writes (and reads) to the stack.
*
* === Description ===
* Uses multiple triad function: a[i] = a[i] * a[i] + a[i]; in order
* to utilize the FMA unit.
*
* Runs in a parallelized for loop.
*
* === Analysis ===
* With gcc -O2 -mavx -mfma FMA compiles to:
* vmovsd xmm0,QWORD PTR [rdi] # 1 read
* vfmadd132sd xmm0,xmm0,xmm0 # 64 x 2 FLOPs+register shuffling
* vmovsd QWORD PTR [rdi-0x8],xmm0 # 1 write
* --------
* 128/16 = 8/1 OI
*
* === Optimization ===
* For packed doubles compile with -Ofast
*
*/
void kernel_8_1_fuseaware(double* a, size_t size);
/**
* @brief A simple 8/1 operational intensity kernel which
* undermines evil fastmath optimization
* @param a An array with double values of size param size
* @param size Size of the three param arrays
* @param result Pointer to result storage
*
* === Warning ===
* Don't use with anything other than -Ofast / -ffast-math
*
* === Description ===
* Uses a simple floating point operation that more closely resembles
* that of 8_1_fuseaware:
* a[i] = a[i]*a[i]; # 128x
*
* Runs in a parallelized for loop.
*
* === Analysis ===
* -Ofast/-ffast-math does not preserve strict IEEE compliance. It
* therefore is allowed to ignore non-associativity of floating
* point operations.
*
* x = x*x*x*x*x*x*x*x; is optimized to x *= x;x *= x;x *= x;
*
* This cleary breaks the whole OI calculation of 8_1_simple.
*
* This kernel does not introduce more byte write-outs than
* 8_1_simple at a high optimization level since a[i] is held
* in a register and only written out once at the end of an
* iteration.
*
*
* === Optimization ===
* Nothing special
*/
void kernel_8_1_simple_fastmath(double* a, size_t size);
/********************************************
* Kernels which potentially compile to *
* different operational intensities than *
* specified *
********************************************/
/**
* @brief A 1/16 operational intensity which might compile to a flawed oi kernel
* @param a An array with double values of size param size
* @param b An array with double values of size param size
* @param size Size of the three param arrays
*
* === Problem ===
* As soon as volatile is used gcc uses the stack for tmp.
* Even if "register" is in place. Resulting in one additional write per loop.
* Omitting volatile results in optimizing away the whole loop
* (checked at -O2, which is necessary for FMA to eventually step in).
* Maybe the value stays in cache, maybe not. It does not live a register.
*
* Even with -O3:
* movsd xmm0,QWORD PTR [rdi+rax*8] # 1 read
* mulsd xmm0,QWORD PTR [rsi+rax*8] # 1 read (+ write to xmm0, not counted)
* # [...] # instructions for loop
* movsd QWORD PTR [rsp-0x8],xmm0 # malicious write
*
* Without volatile (-O3):
* repz ret # that's it
*/
void kernel_1_16_simple_dangerous(double* a, double* b, size_t size);
/**
* @brief A 8/1 operational intensity which might compile to a flawed oi kernel
* @param a An array with double values of size param size
* @param size Size of the three param arrays
*
* === Problem ==
* Same as for kernel_1_16_simple_dangerous
*/
void kernel_8_1_simple_dangerous(double* a, size_t size);
/**
* @brief A 1/8 operational intensity which might compile to a flawed oi kernel
* @param a An array with double values of size param size
* @param size Size of the three param arrays
*
* === Problem ==
* Same as for kernel_1_16_simple_dangerous
*
* Without volatile the loop is optimized away completely.
* With volatile tmp is written to the stack in every loop
* (-O3). tmp could be cached or not. This might depend on
* how large the array is and how the cpu work internally
* -> unpredictable.
*/
void kernel_1_8_vo_dangerous(double* a, size_t size);
#ifdef INTRINS
void kernel_8_1_fuseaware_manpack(double* a, size_t size);
#endif
/****************************************
* Helper macros for repeating things *
****************************************/
#define REP0(X)
#define REP1(X) X
#define REP2(X) REP1(X) REP1(X)
#define REP3(X) REP2(X) REP1(X)
#define REP4(X) REP3(X) REP1(X)
#define REP5(X) REP4(X) REP1(X)
#define REP6(X) REP5(X) REP1(X)
#define REP7(X) REP6(X) REP1(X)
#define REP8(X) REP7(X) REP1(X)
#define REP9(X) REP8(X) REP1(X)
#define REP10(X) REP9(X) REP1(X)
#define REP20(X) REP10(X) REP10(X)
#define REP30(X) REP20(X) REP10(X)
#define REP40(X) REP30(X) REP10(X)
#define REP50(X) REP40(X) REP10(X)
#define REP60(X) REP50(X) REP10(X)
#define REP100(X) REP50(X) REP50(X)
#ifdef ENDEBUG
#define DEBUG(...) do { fprintf(stderr, __VA_ARGS__); fprintf(stderr, "\n"); } while(0)
#else
#define DEBUG(...)
#endif
#endif /* AIKERN_H */

199
prof Normal file
View file

@ -0,0 +1,199 @@
Flat profile:
Each sample counts as 0.01 seconds.
% cumulative self self total
time seconds seconds calls Ts/call Ts/call name
100.01 0.74 0.74 bail_out
0.00 0.74 0.00 50 0.00 0.00 pin_time
0.00 0.74 0.00 5 0.00 0.00 kernel_1_16_fuseaware
0.00 0.74 0.00 5 0.00 0.00 kernel_1_16_simple
0.00 0.74 0.00 5 0.00 0.00 kernel_8_1_fuseaware
0.00 0.74 0.00 5 0.00 0.00 kernel_8_1_simple
0.00 0.74 0.00 5 0.00 0.00 kernel_8_1_simple_fastmath
0.00 0.74 0.00 5 0.00 0.00 kernel_dispatch
0.00 0.74 0.00 5 0.00 0.00 print_kernresult
0.00 0.74 0.00 2 0.00 0.00 pin_time
0.00 0.74 0.00 1 0.00 0.00 get_int
0.00 0.74 0.00 1 0.00 0.00 get_size
0.00 0.74 0.00 1 0.00 0.00 testkern
% the percentage of the total running time of the
time program used by this function.
cumulative a running sum of the number of seconds accounted
seconds for by this function and those listed above it.
self the number of seconds accounted for by this
seconds function alone. This is the major sort for this
listing.
calls the number of times this function was invoked, if
this function is profiled, else blank.
self the average number of milliseconds spent in this
ms/call function per call, if this function is profiled,
else blank.
total the average number of milliseconds spent in this
ms/call function and its descendents per call, if this
function is profiled, else blank.
name the name of the function. This is the minor sort
for this listing. The index shows the location of
the function in the gprof listing. If the index is
in parenthesis it shows where it would appear in
the gprof listing if it were to be printed.
Copyright (C) 2012-2014 Free Software Foundation, Inc.
Copying and distribution of this file, with or without modification,
are permitted in any medium without royalty provided the copyright
notice and this notice are preserved.
Call graph (explanation follows)
granularity: each sample hit covers 2 byte(s) for 1.35% of 0.74 seconds
index % time self children called name
<spontaneous>
[1] 100.0 0.74 0.00 bail_out [1]
-----------------------------------------------
0.00 0.00 50/50 kernel_dispatch [8]
[2] 0.0 0.00 0.00 50 pin_time [2]
-----------------------------------------------
0.00 0.00 5/5 kernel_dispatch [8]
[3] 0.0 0.00 0.00 5 kernel_1_16_fuseaware [3]
-----------------------------------------------
0.00 0.00 5/5 kernel_dispatch [8]
[4] 0.0 0.00 0.00 5 kernel_1_16_simple [4]
-----------------------------------------------
0.00 0.00 5/5 kernel_dispatch [8]
[5] 0.0 0.00 0.00 5 kernel_8_1_fuseaware [5]
-----------------------------------------------
0.00 0.00 5/5 kernel_dispatch [8]
[6] 0.0 0.00 0.00 5 kernel_8_1_simple [6]
-----------------------------------------------
0.00 0.00 5/5 kernel_dispatch [8]
[7] 0.0 0.00 0.00 5 kernel_8_1_simple_fastmath [7]
-----------------------------------------------
0.00 0.00 5/5 main [23]
[8] 0.0 0.00 0.00 5 kernel_dispatch [8]
0.00 0.00 50/50 pin_time [2]
0.00 0.00 5/5 kernel_1_16_simple [4]
0.00 0.00 5/5 kernel_1_16_fuseaware [3]
0.00 0.00 5/5 kernel_8_1_simple [6]
0.00 0.00 5/5 kernel_8_1_fuseaware [5]
0.00 0.00 5/5 kernel_8_1_simple_fastmath [7]
-----------------------------------------------
0.00 0.00 5/5 main [23]
[9] 0.0 0.00 0.00 5 print_kernresult [9]
-----------------------------------------------
0.00 0.00 2/2 main [23]
[10] 0.0 0.00 0.00 2 pin_time [10]
-----------------------------------------------
0.00 0.00 1/1 main [23]
[11] 0.0 0.00 0.00 1 get_int [11]
-----------------------------------------------
0.00 0.00 1/1 main [23]
[12] 0.0 0.00 0.00 1 get_size [12]
-----------------------------------------------
0.00 0.00 1/1 main [23]
[13] 0.0 0.00 0.00 1 testkern [13]
-----------------------------------------------
This table describes the call tree of the program, and was sorted by
the total amount of time spent in each function and its children.
Each entry in this table consists of several lines. The line with the
index number at the left hand margin lists the current function.
The lines above it list the functions that called this function,
and the lines below it list the functions this one called.
This line lists:
index A unique number given to each element of the table.
Index numbers are sorted numerically.
The index number is printed next to every function name so
it is easier to look up where the function is in the table.
% time This is the percentage of the `total' time that was spent
in this function and its children. Note that due to
different viewpoints, functions excluded by options, etc,
these numbers will NOT add up to 100%.
self This is the total amount of time spent in this function.
children This is the total amount of time propagated into this
function by its children.
called This is the number of times the function was called.
If the function called itself recursively, the number
only includes non-recursive calls, and is followed by
a `+' and the number of recursive calls.
name The name of the current function. The index number is
printed after it. If the function is a member of a
cycle, the cycle number is printed between the
function's name and the index number.
For the function's parents, the fields have the following meanings:
self This is the amount of time that was propagated directly
from the function into this parent.
children This is the amount of time that was propagated from
the function's children into this parent.
called This is the number of times this parent called the
function `/' the total number of times the function
was called. Recursive calls to the function are not
included in the number after the `/'.
name This is the name of the parent. The parent's index
number is printed after it. If the parent is a
member of a cycle, the cycle number is printed between
the name and the index number.
If the parents of the function cannot be determined, the word
`<spontaneous>' is printed in the `name' field, and all the other
fields are blank.
For the function's children, the fields have the following meanings:
self This is the amount of time that was propagated directly
from the child into the function.
children This is the amount of time that was propagated from the
child's children to the function.
called This is the number of times the function called
this child `/' the total number of times the child
was called. Recursive calls by the child are not
listed in the number after the `/'.
name This is the name of the child. The child's index
number is printed after it. If the child is a
member of a cycle, the cycle number is printed
between the name and the index number.
If there are any cycles (circles) in the call graph, there is an
entry for the cycle-as-a-whole. This entry shows who called the
cycle (as parents) and the members of the cycle (as children.)
The `+' recursive calls entry shows the number of function calls that
were internal to the cycle, and the calls entry for each member shows,
for that member, how many times it was called from other members of
the cycle.
Copyright (C) 2012-2014 Free Software Foundation, Inc.
Copying and distribution of this file, with or without modification,
are permitted in any medium without royalty provided the copyright
notice and this notice are preserved.
Index by function name
[1] bail_out (aikern.c) [5] kernel_8_1_fuseaware [2] pin_time (aikern.c)
[11] get_int (roofline.c) [6] kernel_8_1_simple [9] print_kernresult (roofline.c)
[12] get_size (roofline.c) [7] kernel_8_1_simple_fastmath [13] testkern (roofline.c)
[3] kernel_1_16_fuseaware [8] kernel_dispatch
[4] kernel_1_16_simple [10] pin_time (roofline.c)

342
roofline.c Normal file
View file

@ -0,0 +1,342 @@
# include <stdlib.h>
# include <stdio.h>
# include <unistd.h>
# include <ctype.h>
# include <sys/time.h>
# include <sys/stat.h>
# include <errno.h>
# include <string.h>
# include <stdint.h>
# include <getopt.h>
# include <stdarg.h>
# include <limits.h>
# include "aikern.h"
/* === Macros === */
#ifdef ENDEBUG
#define DEBUG(...) do { fprintf(stderr, __VA_ARGS__); fprintf(stderr, "\n"); } while(0)
#else
#define DEBUG(...)
#endif
/* === Constants === */
/* === Global Variables === */
char* prog_name;
/* === Prototypes === */
/**
* @brief print usage message
*/
static void usage(void);
/**
* @brief terminate program on program error
* @param msg additional message to print
* @param ret exit value
*/
static void bail_out(char* fmt, ...);
/**
* @brief converts the argument to size_t if possible.
* bails out on error.
* @param oparg the argument to convert
*/
static size_t get_size(char* oparg);
/**
* @brief converts the argument to int if possible.
* bails out on error.
* @param oparg the argument to convert
*/
static int get_int(char* oparg);
/**
* @brief microseconds since epoch
*/
static double pin_time(void);
/**
* @brief a simple test kernel with ai of 1/16
*/
static void testkern(double* a, double* b, double* c, size_t size);
/**
* @brief pretty prints a kern_result
*/
static void print_kernresult(kern_result* result, const char* logname);
int main(int argc, char* argv[]) {
prog_name = argv[0];
int opt;
char *size_arg = NULL;
char *runs_arg = NULL;
while((opt = getopt(argc, argv, "s:r:")) != -1)
{
switch(opt)
{
case 's':
size_arg = optarg;
break;
case 'r':
runs_arg = optarg;
break;
case '?':
usage();
default:
usage();
}
}
if(optind < argc)
{
for (int index = optind; index < argc; index++)
bail_out ("Non-option argument %s\n", argv[index]);
usage();
}
if(size_arg == NULL || runs_arg == NULL)
usage();
size_t size = get_size(size_arg);
int runs = get_int(runs_arg);
// Allocating arrays
printf("Will run with array sizes of %zu elements\n", size);
printf("Will calculate min, max, avg for %d runs\n", runs);
double* a = malloc(sizeof(double)*(size));
double* b = malloc(sizeof(double)*(size));
double* c = malloc(sizeof(double)*(size));
if(a==NULL || b==NULL || c == NULL)
bail_out("One of the mallocs failed\n. a = %p, b=%p, c=%p", a, b, c);
printf("Allocated 3 arrays (3*%.2f MB = %.2f GB)\n", (sizeof(double)*(size)/1024.0/1024.0), (sizeof(double)*(size)*3/1024.0/1024.0/1024));
printf("Filling arrays with dummy values. This will also warm the cache\n");
// Filling arrays with arbitrary numbers
#pragma omp parallel for
for (size_t j=0; j<size; j++)
{
a[j] = 1.0;
b[j] = 2.0;
c[j] = 3.0;
}
double t;
printf("Heating up machine\n");
t = pin_time();
testkern(a,b,c, size);
t = pin_time() - t;
printf("Machine heating took %.4f microseconds = %.4f seconds (with test OI kernel)\n", (t*1.0E6), t);
printf("Starting tests...\n\n\n");
// Executing kernels
kern_result simple16 = kernel_dispatch(SIMPLE_1_16, a, b, c, size, runs);
kern_result fma16 = kernel_dispatch(FMA_1_16, a, b, c, size, runs);
kern_result simple8 = kernel_dispatch(SIMPLE_8_1, a, b, c, size, runs);
kern_result fma8 = kernel_dispatch(FMA_8_1, a, b, c, size, runs);
kern_result simple8fm = kernel_dispatch(SIMPLE_8_1_FASTMATH, a, b, c, size, runs);
#ifdef INTRINS
DEBUG("Running manpack now");
kern_result fma8manpack = kernel_dispatch(FMA_8_1_MANPACK, a, b, c, size, runs);
DEBUG("manpack run successful");
#endif
// Freeing arrays
free(a);
free(b);
free(c);
// Printing results
print_kernresult(&simple16, "simple16");
print_kernresult(&fma16, "fma16");
print_kernresult(&simple8, "simple8");
print_kernresult(&fma8, "fma8");
print_kernresult(&simple8fm, "simple8fastmath");
#ifdef INTRINS
print_kernresult(&fma8manpack, "fma8manpack");
#endif
printf("\n\n\n");
printf("Please refer to the log files in the log/ folder for details about the GFLOP/s of every kernel.");
printf("\n");
printf("Exiting...");
exit(EXIT_SUCCESS);
}
static void testkern(double* a, double* b, double* c, size_t size)
{
#pragma omp parallel for
for (size_t j = 0; j < size; j++)
{
/* 3*8 Bytes read + 3*8 Bytes write, 3 FLOPs -> AI = 3/(2*3*8) = 1/16 */
a[j] = 2.0E0 * a[j];
b[j] = 2.0E0 * b[j];
c[j] = 2.0E0 * c[j];
}
}
/* === Helper Functions === */
static double pin_time(void)
{
struct timeval tp;
int i;
i = gettimeofday(&tp,NULL);
if(i != 0)
{
bail_out("Time measurement impossible. gettimeofday error");
}
return ( (double) tp.tv_sec + (double) tp.tv_usec * 1.e-6 );
}
static size_t get_size(char *oparg)
{
long long int llsize = strtoll(oparg, NULL, 10);
if(llsize <= 0)
usage();
unsigned long long int u_llsize = (unsigned long long int) llsize;
if(u_llsize > SIZE_MAX)
{
bail_out("Only size between 1 to %zu allowed.", SIZE_MAX);
}
return (size_t) llsize;
}
static int get_int(char *oparg)
{
long long int llsize = strtoll(oparg, NULL, 10);
if(llsize <= 0)
usage();
unsigned long long int u_llsize = (unsigned long long int) llsize;
if(u_llsize > INT_MAX)
{
bail_out("Only size between 1 to %d allowed.", INT_MAX);
}
return (int) llsize;
}
static void usage()
{
fprintf(stderr, "USAGE: ./roofline -s <size> -r <runs> \n");
fprintf(stderr, "e.g.: ./roofline -s 100000 -r 5 \n");
bail_out("Invalid paramers");
}
static void bail_out(char* fmt, ...)
{
if(fmt != NULL)
{
char msgbuf[150];
va_list vl;
va_start(vl, fmt);
if(vsnprintf(msgbuf, sizeof(msgbuf), fmt, vl) < 0)
msgbuf[0] = '\0';
va_end( vl);
if(strlen(msgbuf) > 0)
(void)fprintf(stderr, "%s: %s \n", prog_name, msgbuf);
}
if(errno != 0)
(void)fprintf(stderr, "%s: %s\n", prog_name, strerror(errno));
exit(EXIT_FAILURE);
}
static void print_kernresult(kern_result* result, const char* logname)
{
struct stat st = {};
if (stat("log", &st) == -1)
{
if(mkdir("log", 0700))
{
bail_out("Couldn't create log directory for %s", result->kern_name);
}
}
char logpath[20];
snprintf(logpath, sizeof(logpath), "%s/%s", "log", logname);
FILE* log = fopen(logpath, "w");
if(log == NULL)
bail_out("Couldn't open log file for %s", result->kern_name);
if(fputs("run,start,end,delta,GFLOP/s\n", log) == EOF)
{
fclose(log);
bail_out("Couldn't write header to log file");
}
printf("=== %s ===\n", result->kern_name);
double min;
double max;
double sum = 0.0;
double deltas[result->runs];
deltas[0] = result->ends[0] - result->starts[0];
min=deltas[0];
max=deltas[0];
sum+=deltas[0];
for(size_t i=1; i<result->runs; i++)
{
deltas[i] = result->ends[i] - result->starts[i];
sum+=deltas[i];
if(deltas[i] < min) min=deltas[i];
if(deltas[i] > max) max=deltas[i];
double gflops = ((result->flops * result->size) / deltas[i]) / 1.0E9;
if(fprintf(log, "%zu,%.4f,%.4f,%.4f,%.4f\n",
i, result->starts[i],
result->ends[i], deltas[i],
gflops) == EOF)
{
fclose(log);
bail_out("Couldn't write to log file");
}
}
printf("%d flop(s) per run\t %zu run(s)\n\n", result->flops, result->runs);
printf("Min: %.4f \t Max: %.4f \t Avg: %.4f\n", min, max, (sum/result->runs));
printf("\n\n\n");
if(fclose(log))
{
bail_out("Couldn't close log file for %s", result->kern_name);
}
}