Compare commits
2 Commits
46ee06406f
...
ba31dd9f8c
Author | SHA1 | Date | |
---|---|---|---|
ba31dd9f8c | |||
f355ab2d25 |
1
haversine_02/.gitignore
vendored
1
haversine_02/.gitignore
vendored
@ -4,6 +4,7 @@
|
||||
compile_commands.json
|
||||
count_and_distances
|
||||
pairs.json
|
||||
cache_test
|
||||
main
|
||||
genhavr
|
||||
prochavr
|
||||
|
@ -9,6 +9,9 @@
|
||||
|
||||
#define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR)
|
||||
|
||||
u64 *g_cache_output = NULL;
|
||||
u64 g_size = 1024 * 1024 * 1024 / 128 * sizeof(u64);
|
||||
|
||||
extern "C" void mov_all_bytes_asm(char *buffer, u64 size);
|
||||
extern "C" void nop_all_bytes_asm(u64 size);
|
||||
extern "C" void nop_1x3_all_bytes_asm(u64 size);
|
||||
@ -47,6 +50,7 @@ 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_16x4_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);
|
||||
|
||||
void test_fread(reptester *tester, alloc_type type);
|
||||
void test_read(reptester *tester, alloc_type type);
|
||||
@ -89,6 +93,20 @@ 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_16x4_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_32k(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_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);
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
@ -108,6 +126,8 @@ int main(int argc, char *argv[]) {
|
||||
break;
|
||||
}
|
||||
|
||||
g_cache_output = (u64 *)calloc(1, g_size);
|
||||
|
||||
// clang-format off
|
||||
reptester tester = {
|
||||
{filename, NULL, 0, 0}, // params
|
||||
@ -182,15 +202,36 @@ int main(int argc, char *argv[]) {
|
||||
// {{"READ 2x2", "READ 2x2 WITH MALLOC"}, test_read_2x2},
|
||||
// {{"READ 4x2", "READ 4x2 WITH MALLOC"}, test_read_4x2},
|
||||
// {{"READ 8x2", "READ 8x2 WITH MALLOC"}, test_read_8x2},
|
||||
{{"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 16x2_simd", "READ 16x2_simd WITH MALLOC"}, test_read_16x2_simd},
|
||||
{{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"},
|
||||
test_read_32x2_simd_offset},
|
||||
{{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"},
|
||||
test_read_32x2_simd_no_offset},
|
||||
{{"READ 16x4_simd", "READ 16x4_simd WITH MALLOC"}, test_read_16x4_simd},
|
||||
{{"READ 32x4_simd", "READ 32x4_simd WITH MALLOC"}, test_read_32x4_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 16x2_simd", "READ 16x2_simd WITH MALLOC"},
|
||||
// test_read_16x2_simd},
|
||||
// {{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"},
|
||||
// test_read_32x2_simd_offset},
|
||||
// {{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"},
|
||||
// test_read_32x2_simd_no_offset},
|
||||
// {{"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 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 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);
|
||||
@ -586,7 +627,6 @@ void test_align63_loop(reptester *tester, alloc_type type) {
|
||||
}
|
||||
|
||||
void test_align75_loop(reptester *tester, alloc_type type) {
|
||||
|
||||
u64 start = read_cpu_timer();
|
||||
u64 fault_count_start = page_fault_count();
|
||||
|
||||
@ -612,7 +652,6 @@ void test_align75_loop(reptester *tester, alloc_type type) {
|
||||
}
|
||||
|
||||
void test_align90_loop(reptester *tester, alloc_type type) {
|
||||
|
||||
u64 start = read_cpu_timer();
|
||||
u64 fault_count_start = page_fault_count();
|
||||
|
||||
@ -1112,7 +1151,6 @@ void test_read_4x2_simd(reptester *tester, alloc_type type) {
|
||||
}
|
||||
|
||||
void test_read_8x2_simd(reptester *tester, alloc_type type) {
|
||||
|
||||
u64 start = read_cpu_timer();
|
||||
u64 fault_count_start = page_fault_count();
|
||||
|
||||
@ -1138,7 +1176,6 @@ void test_read_8x2_simd(reptester *tester, alloc_type type) {
|
||||
}
|
||||
|
||||
void test_read_16x2_simd(reptester *tester, alloc_type type) {
|
||||
|
||||
u64 start = read_cpu_timer();
|
||||
u64 fault_count_start = page_fault_count();
|
||||
|
||||
@ -1264,6 +1301,356 @@ void test_read_32x4_simd(reptester *tester, alloc_type 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_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_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_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) {
|
||||
if (!fp) {
|
||||
return 0;
|
||||
|
@ -36,6 +36,7 @@ global read_32x2_simd_offset
|
||||
global read_32x2_simd_no_offset
|
||||
global read_16x4_simd
|
||||
global read_32x4_simd
|
||||
global cache_test ; Expects 3 inputs (pointer, read_count, mask)
|
||||
|
||||
mov_all_bytes_asm:
|
||||
xor rax, rax
|
||||
@ -492,3 +493,19 @@ read_32x4_simd:
|
||||
jb .loop
|
||||
|
||||
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
|
||||
|
Loading…
Reference in New Issue
Block a user