Compare commits

...

7 Commits

Author SHA1 Message Date
94167e05fd Remove unused variables 2024-06-23 21:52:58 +01:00
7177add4ce Switch to using mmap and test unaligned loads 2024-06-23 21:10:14 +01:00
37d3340df9 Rename C++ files 2024-06-22 13:54:45 +01:00
fcdaf41495 Write to test buffer before working with it 2024-05-19 00:32:39 +01:00
cec0662e68 Add 256k cache test function 2024-05-19 00:10:18 +01:00
ba31dd9f8c Add cache testing functions 2024-05-06 22:55:44 +01:00
f355ab2d25 Update .gitignore 2024-05-06 22:55:32 +01:00
26 changed files with 890 additions and 372 deletions

View File

@@ -1,2 +1,2 @@
all: all:
clang++ -g dasm.cpp -o dasm clang++ -g dasm.cc -o dasm

View File

@@ -1,2 +1,2 @@
all: all:
clang++ -g dasm.cpp -o dasm clang++ -g dasm.cc -o dasm

View File

@@ -1,2 +1,2 @@
all: all:
clang++ -g dasm.cpp -o dasm clang++ -g dasm.cc -o dasm

View File

@@ -1,7 +1,7 @@
CC=clang++ CC=clang++
CFLAGS=-g -O0 -Wall -Wextra CFLAGS=-g -O0 -Wall -Wextra
LIBS=-Wl,-rpath,./lib -L./lib -lsim86 LIBS=-Wl,-rpath,./lib -L./lib -lsim86
SRC=*.cpp SRC=*.cc
OUT=sim86 OUT=sim86
all: all:

View File

@@ -1,17 +1,17 @@
mk_haversine_fscanf: mk_haversine_fscanf:
clang++ -g cpp/fscanf.cpp cpp/haversine.cpp -o cpp/haverscan clang++ -g cpp/fscanf.cc cpp/haversine.cc -o cpp/haverscan
run_haversine_fscanf: run_haversine_fscanf:
cd ./cpp && ./haverscan cd ./cpp && ./haverscan
mk_haversine_strtok: mk_haversine_strtok:
clang++ -g cpp/strtok.cpp cpp/haversine.cpp -o cpp/haverstrtok clang++ -g cpp/strtok.cc cpp/haversine.cc -o cpp/haverstrtok
run_haversine_strtok: run_haversine_strtok:
cd ./cpp && ./haverstrtok cd ./cpp && ./haverstrtok
mk_test: mk_test:
clang++ -g -lpthread cpp/test.cpp cpp/haversine.cpp -o cpp/test clang++ -g -lpthread cpp/test.cc cpp/haversine.cc -o cpp/test
run_test: run_test:
cd ./cpp && ./test cd ./cpp && ./test

View File

@@ -4,6 +4,7 @@
compile_commands.json compile_commands.json
count_and_distances count_and_distances
pairs.json pairs.json
cache_test
main main
genhavr genhavr
prochavr prochavr

View File

@@ -43,11 +43,11 @@ fi
# GENERATOR # GENERATOR
GENSRC="./src/generator/gen_argparser.cpp \ GENSRC="./src/generator/gen_argparser.cc \
./src/generator/generator.cpp \ ./src/generator/generator.cc \
./src/haversine.cpp \ ./src/haversine.cc \
./src/point_types.cpp \ ./src/point_types.cc \
./src/generator/main.cpp" ./src/generator/main.cc"
GENOUT=genhavr GENOUT=genhavr
(set -x ; $CXX $CFLAGS $GENSRC -o $GENOUT) (set -x ; $CXX $CFLAGS $GENSRC -o $GENOUT)
@@ -64,10 +64,10 @@ JSONFLAGS="-c "
JSON_BUILD_DIR=json_build JSON_BUILD_DIR=json_build
PROCSRC="./$JSON_BUILD_DIR/*.o \ PROCSRC="./$JSON_BUILD_DIR/*.o \
./src/haversine.cpp \ ./src/haversine.cc \
./src/point_types.cpp \ ./src/point_types.cc \
./src/processor/proc_argparser.cpp \ ./src/processor/proc_argparser.cc \
./src/processor/main.cpp " ./src/processor/main.cc "
PROCOUT=prochavr PROCOUT=prochavr
# MEMTESTER # MEMTESTER
@@ -113,7 +113,7 @@ if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
cd ../ cd ../
# REPETITION TESTING # REPETITION TESTING
REPTESTSRC="./src/repetition_testing/*.cpp ./$PROF_BUILD_DIR/*.o $ASM_LIB" REPTESTSRC="./src/repetition_testing/*.cc ./$PROF_BUILD_DIR/*.o $ASM_LIB"
REPTESTOUT=reptest REPTESTOUT=reptest
(set -x ; $CXX $CFLAGS $REPTESTFLAGS $REPTESTSRC -o $REPTESTOUT) (set -x ; $CXX $CFLAGS $REPTESTFLAGS $REPTESTSRC -o $REPTESTOUT)

View File

@@ -5,6 +5,7 @@
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
#include <string.h> #include <string.h>
#include <sys/mman.h>
#include <unistd.h> #include <unistd.h>
#define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR) #define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR)
@@ -47,6 +48,8 @@ extern "C" void read_32x2_simd_offset(char *buffer, u64 size);
extern "C" void read_32x2_simd_no_offset(char *buffer, u64 size); extern "C" void read_32x2_simd_no_offset(char *buffer, u64 size);
extern "C" void read_16x4_simd(char *buffer, u64 size); extern "C" void read_16x4_simd(char *buffer, u64 size);
extern "C" void read_32x4_simd(char *buffer, u64 size); extern "C" void read_32x4_simd(char *buffer, u64 size);
extern "C" void cache_test(char *buffer, u64 size, u64 mask);
extern "C" void cache_test_unaligned(char *buffer, u64 size, u64 mask);
void test_fread(reptester *tester, alloc_type type); void test_fread(reptester *tester, alloc_type type);
void test_read(reptester *tester, alloc_type type); void test_read(reptester *tester, alloc_type type);
@@ -89,6 +92,23 @@ void test_read_32x2_simd_offset(reptester *tester, alloc_type type);
void test_read_32x2_simd_no_offset(reptester *tester, alloc_type type); void test_read_32x2_simd_no_offset(reptester *tester, alloc_type type);
void test_read_16x4_simd(reptester *tester, alloc_type type); void test_read_16x4_simd(reptester *tester, alloc_type type);
void test_read_32x4_simd(reptester *tester, alloc_type type); void test_read_32x4_simd(reptester *tester, alloc_type type);
void test_cache_test_16k(reptester *tester, alloc_type type);
void test_cache_test_16k_unaligned(reptester *tester, alloc_type type);
void test_cache_test_32k(reptester *tester, alloc_type type);
void test_cache_test_32k_unaligned(reptester *tester, alloc_type type);
void test_cache_test_64k(reptester *tester, alloc_type type);
void test_cache_test_128k(reptester *tester, alloc_type type);
void test_cache_test_256k(reptester *tester, alloc_type type);
void test_cache_test_512k(reptester *tester, alloc_type type);
void test_cache_test_1m(reptester *tester, alloc_type type);
void test_cache_test_2m(reptester *tester, alloc_type type);
void test_cache_test_4m(reptester *tester, alloc_type type);
void test_cache_test_8m(reptester *tester, alloc_type type);
void test_cache_test_16m(reptester *tester, alloc_type type);
void test_cache_test_32m(reptester *tester, alloc_type type);
void test_cache_test_64m(reptester *tester, alloc_type type);
void test_cache_test_512m(reptester *tester, alloc_type type);
void test_cache_test_full(reptester *tester, alloc_type type);
u64 get_file_length(FILE *fp); u64 get_file_length(FILE *fp);
int main(int argc, char *argv[]) { int main(int argc, char *argv[]) {
@@ -182,22 +202,58 @@ int main(int argc, char *argv[]) {
// {{"READ 2x2", "READ 2x2 WITH MALLOC"}, test_read_2x2}, // {{"READ 2x2", "READ 2x2 WITH MALLOC"}, test_read_2x2},
// {{"READ 4x2", "READ 4x2 WITH MALLOC"}, test_read_4x2}, // {{"READ 4x2", "READ 4x2 WITH MALLOC"}, test_read_4x2},
// {{"READ 8x2", "READ 8x2 WITH MALLOC"}, test_read_8x2}, // {{"READ 8x2", "READ 8x2 WITH MALLOC"}, test_read_8x2},
{{"READ 4x2_simd", "READ 4x2_simd WITH MALLOC"}, test_read_4x2_simd}, // {{"READ 4x2_simd", "READ 4x2_simd WITH MALLOC"}, test_read_4x2_simd},
{{"READ 8x2_simd", "READ 8x2_simd WITH MALLOC"}, test_read_8x2_simd}, // {{"READ 8x2_simd", "READ 8x2_simd WITH MALLOC"}, test_read_8x2_simd},
{{"READ 16x2_simd", "READ 16x2_simd WITH MALLOC"}, test_read_16x2_simd}, // {{"READ 16x2_simd", "READ 16x2_simd WITH MALLOC"},
{{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"}, // test_read_16x2_simd},
test_read_32x2_simd_offset}, // {{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"},
{{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"}, // test_read_32x2_simd_offset},
test_read_32x2_simd_no_offset}, // {{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"},
{{"READ 16x4_simd", "READ 16x4_simd WITH MALLOC"}, test_read_16x4_simd}, // test_read_32x2_simd_no_offset},
{{"READ 32x4_simd", "READ 32x4_simd WITH MALLOC"}, test_read_32x4_simd}, // {{"READ 16x4_simd", "READ 16x4_simd WITH MALLOC"},
// test_read_16x4_simd},
// {{"READ 32x4_simd", "READ 32x4_simd WITH MALLOC"},
// test_read_32x4_simd},
{{"CACHE TEST 16K", "CACHE TEST 16K WITH MALLOC"}, test_cache_test_16k},
{{"CACHE TEST 16K UNALIGNED", "CACHE TEST 16K UNALIGNED WITH MALLOC"},
test_cache_test_16k_unaligned},
// {{"CACHE TEST 32K", "CACHE TEST 32K WITH MALLOC"},
// test_cache_test_32k},
// {{"CACHE TEST 64K", "CACHE TEST 64K WITH MALLOC"},
// test_cache_test_64k},
// {{"CACHE TEST 128K", "CACHE TEST 128K WITH MALLOC"},
// test_cache_test_128k},
// {{"CACHE TEST 256K", "CACHE TEST 256K WITH MALLOC"},
// test_cache_test_256k},
// {{"CACHE TEST 512K", "CACHE TEST 512K WITH MALLOC"},
// test_cache_test_512k},
// {{"CACHE TEST 1M", "CACHE TEST 1M WITH MALLOC"}, test_cache_test_1m},
// {{"CACHE TEST 2M", "CACHE TEST 2M WITH MALLOC"}, test_cache_test_2m},
// {{"CACHE TEST 4M", "CACHE TEST 4M WITH MALLOC"}, test_cache_test_4m},
// {{"CACHE TEST 8M", "CACHE TEST 8M WITH MALLOC"}, test_cache_test_8m},
// {{"CACHE TEST 16M", "CACHE TEST 16M WITH MALLOC"},
// test_cache_test_16m},
// {{"CACHE TEST 32M", "CACHE TEST 32M WITH MALLOC"},
// test_cache_test_32m},
// {{"CACHE TEST 64M", "CACHE TEST 64M WITH MALLOC"},
// test_cache_test_64m},
// {{"CACHE TEST 512M", "CACHE TEST 512M WITH MALLOC"},
// test_cache_test_512m},
// {{"CACHE TEST FULL", "CACHE TEST FULL WITH MALLOC"},
// test_cache_test_full},
}; };
tester.params.read_size = get_file_length(fp); tester.params.read_size = get_file_length(fp);
tester.params.read_count = 1; tester.params.read_count = 1;
tester.params.buffer = (char *)malloc(tester.params.read_size + 1); tester.params.buffer =
(char *)mmap(NULL, tester.params.read_size + 1, PROT_READ | PROT_WRITE,
MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
memset(tester.params.buffer, 0, tester.params.read_size + 1); memset(tester.params.buffer, 0, tester.params.read_size + 1);
for (u64 i = 0; i < tester.params.read_size; ++i) {
tester.params.buffer[i] = (char)i;
}
for (u64 i = 0; i < waves; ++i) { for (u64 i = 0; i < waves; ++i) {
for (u64 j = 0; j < ARR_LEN(funcs); ++j) { for (u64 j = 0; j < ARR_LEN(funcs); ++j) {
for (u64 k = 0; k < ALLOC_TYPE_WITH_MALLOC; ++k) { for (u64 k = 0; k < ALLOC_TYPE_WITH_MALLOC; ++k) {
@@ -208,7 +264,7 @@ int main(int argc, char *argv[]) {
fclose(fp); fclose(fp);
free(tester.params.buffer); munmap(tester.params.buffer, tester.params.read_size + 1);
return 0; return 0;
} }
@@ -586,7 +642,6 @@ void test_align63_loop(reptester *tester, alloc_type type) {
} }
void test_align75_loop(reptester *tester, alloc_type type) { void test_align75_loop(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer(); u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count(); u64 fault_count_start = page_fault_count();
@@ -612,7 +667,6 @@ void test_align75_loop(reptester *tester, alloc_type type) {
} }
void test_align90_loop(reptester *tester, alloc_type type) { void test_align90_loop(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer(); u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count(); u64 fault_count_start = page_fault_count();
@@ -1112,7 +1166,6 @@ void test_read_4x2_simd(reptester *tester, alloc_type type) {
} }
void test_read_8x2_simd(reptester *tester, alloc_type type) { void test_read_8x2_simd(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer(); u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count(); u64 fault_count_start = page_fault_count();
@@ -1138,7 +1191,6 @@ void test_read_8x2_simd(reptester *tester, alloc_type type) {
} }
void test_read_16x2_simd(reptester *tester, alloc_type type) { void test_read_16x2_simd(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer(); u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count(); u64 fault_count_start = page_fault_count();
@@ -1264,6 +1316,431 @@ void test_read_32x4_simd(reptester *tester, alloc_type type) {
handle_free(tester, type); handle_free(tester, type);
} }
void test_cache_test_16k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x3fff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_16k_unaligned(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test_unaligned(tester->params.buffer, total_size, 0x3fff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_32k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x7fff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_32k_unaligned(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test_unaligned(tester->params.buffer, total_size, 0x7fff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_64k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0xffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_128k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x1ffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_256k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x3ffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_512k(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x7ffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_1m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0xfffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_2m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x1fffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_4m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x3fffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_8m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x7fffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_16m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0xffffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_32m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x1ffffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_64m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x3ffffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_512m(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0x1fffffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
void test_cache_test_full(reptester *tester, alloc_type type) {
u64 start = read_cpu_timer();
u64 fault_count_start = page_fault_count();
handle_alloc(tester, type);
u64 total_size = tester->params.read_size * tester->params.read_count;
cache_test(tester->params.buffer, total_size, 0xffffffffffffffff);
u64 fault_count_end = page_fault_count();
u64 end = read_cpu_timer();
u64 read_time = end - start;
u64 page_faults = fault_count_end - fault_count_start;
tester->results = {
total_size,
read_time,
page_faults,
};
handle_free(tester, type);
}
u64 get_file_length(FILE *fp) { u64 get_file_length(FILE *fp) {
if (!fp) { if (!fp) {
return 0; return 0;

View File

@@ -36,6 +36,8 @@ global read_32x2_simd_offset
global read_32x2_simd_no_offset global read_32x2_simd_no_offset
global read_16x4_simd global read_16x4_simd
global read_32x4_simd global read_32x4_simd
global cache_test ; Expects 3 inputs (pointer, read_count, mask)
global cache_test_unaligned ; Expects 3 inputs (pointer, read_count, mask)
mov_all_bytes_asm: mov_all_bytes_asm:
xor rax, rax xor rax, rax
@@ -492,3 +494,36 @@ read_32x4_simd:
jb .loop jb .loop
ret ret
cache_test:
xor r10, r10 ; Zero loop counter
mov rbx, rdi ; Save original pointer
.loop:
add rdi, r10 ; Advance the pointer
add r10, 128 ; Increment loop counter
and r10, rdx ; Mask offset
vmovdqu ymm0, [rdi + 0]
vmovdqu ymm1, [rdi + 32]
vmovdqu ymm2, [rdi + 64]
vmovdqu ymm3, [rdi + 96]
mov rdi, rbx ; Restore original pointer
sub rsi, 128 ; Decrement count
ja .loop
ret
cache_test_unaligned:
xor r10, r10 ; Zero loop counter
add rdi, 5 ; Unalign pointer
mov rbx, rdi ; Save original pointer
.loop:
add rdi, r10 ; Advance the pointer
add r10, 128 ; Increment loop counter
and r10, rdx ; Mask offset
vmovdqu ymm0, [rdi + 0]
vmovdqu ymm1, [rdi + 32]
vmovdqu ymm2, [rdi + 64]
vmovdqu ymm3, [rdi + 96]
mov rdi, rbx ; Restore original pointer
sub rsi, 128 ; Decrement count
ja .loop
ret

View File

@@ -3,6 +3,7 @@
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
#include <string.h> #include <string.h>
#include <sys/mman.h>
#include <sys/resource.h> #include <sys/resource.h>
#include <sys/time.h> #include <sys/time.h>
@@ -10,7 +11,9 @@ void handle_alloc(reptester *tester, alloc_type type) {
switch (type) { switch (type) {
case ALLOC_TYPE_WITH_MALLOC: case ALLOC_TYPE_WITH_MALLOC:
if (!(tester->params.buffer)) { if (!(tester->params.buffer)) {
tester->params.buffer = (char *)malloc(tester->params.read_size + 1); tester->params.buffer = (char *)mmap(
NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
memset(tester->params.buffer, 0, tester->params.read_size + 1); memset(tester->params.buffer, 0, tester->params.read_size + 1);
} }
@@ -24,7 +27,7 @@ void handle_free(reptester *tester, alloc_type type) {
switch (type) { switch (type) {
case ALLOC_TYPE_WITH_MALLOC: case ALLOC_TYPE_WITH_MALLOC:
if (tester->params.buffer) { if (tester->params.buffer) {
free(tester->params.buffer); munmap(tester->params.buffer, tester->params.read_size + 1);
tester->params.buffer = NULL; tester->params.buffer = NULL;
} }
@@ -58,7 +61,9 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
if (type == ALLOC_TYPE_WITH_MALLOC) { if (type == ALLOC_TYPE_WITH_MALLOC) {
buffer = tester->params.buffer; buffer = tester->params.buffer;
tester->params.buffer = (char *)malloc(tester->params.read_size + 1); tester->params.buffer =
(char *)mmap(NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
memset(tester->params.buffer, 0, tester->params.read_size + 1); memset(tester->params.buffer, 0, tester->params.read_size + 1);
} }
@@ -100,7 +105,7 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
} }
if (type == ALLOC_TYPE_WITH_MALLOC) { if (type == ALLOC_TYPE_WITH_MALLOC) {
free(tester->params.buffer); munmap(tester->params.buffer, tester->params.read_size + 1);
tester->params.buffer = buffer; tester->params.buffer = buffer;
} }