Skip to content

Instantly share code, notes, and snippets.

@stgatilov

stgatilov/gen.cpp

Created August 2, 2017 16:12
Show Gist options
  • Save stgatilov/ac832c6dd30e695e566f42b1ef8ef45a to your computer and use it in GitHub Desktop.
Save stgatilov/ac832c6dd30e695e566f42b1ef8ef45a to your computer and use it in GitHub Desktop.
Download ZIP
Vectorizing std::merge with vpermd from AVX2 and lookup table
Raw
gen.cpp
#include <stdio.h>
#include <time.h>
#include <string.h>
#include <assert.h>
#include <stdint.h>
#include <algorithm>
#include <vector>
//======================== Perfect hash generator =========================
typedef uint32_t hash_int;
//searching for a hash function of this kind:
hash_int HashFunction(hash_int param, hash_int x, int bits) {
return (param * x) >> hash_int(32 - bits);
}
static const hash_int NO_PARAM = hash_int(-1);
hash_int GeneratePerfectHash(std::vector<hash_int> values, int bits, hash_int tries = NO_PARAM) {
std::vector<int> table;
for (hash_int param = 0; param < tries; param++) {
table.assign(hash_int(1)<<bits, 0);
bool ok = true;
for (int i = 0; ok && i < values.size(); i++) {
hash_int x = values[i];
hash_int cell = HashFunction(param, x, bits);
if (table[cell]++)
ok = false;
}
if (ok)
return param;
}
return NO_PARAM;
}
//======================== Lookup table generator =========================
template<class T> void PrintTable(const T *lut, int size) {
for (int i = 0; i < size; i++) {
const unsigned char *entry = (unsigned char*) &lut[i];
printf("\t{");
for (int j = 0; j < sizeof(T); j++) {
if (j) printf(", ");
printf("0x%02X", (unsigned int)(entry[j]));
}
printf("},\n");
}
printf("\n");
}
struct m256d {
int d[8];
};
static const int HASH_BITS = 7;
m256d lookupTable[1<<HASH_BITS];
uint32_t GenerateLookupTableHash() {
std::vector<uint32_t> allMasks;
std::vector<m256d> allPerms;
for (int m = 0; m < 1<<8; m++) {
int ak = 0;
int a[8], b[8];
m256d perm;
for (int i = 0; i < 8; i++) {
if (m & (1<<i)) {
perm.d[i] = ak;
a[ak++] = i;
}
else {
perm.d[i] = 4 + (i-ak);
b[i-ak] = i;
}
}
if (ak != 4) continue;
int mask = 0;
for (int s = 0; s < 4; s++)
for (int i = 0; i < 4; i++) {
int bit = (a[i] < b[(i+s) % 4]);
mask ^= (bit << (s*4+i));
}
allMasks.push_back(mask);
allPerms.push_back(perm);
}
uint32_t param = GeneratePerfectHash(allMasks, HASH_BITS);
assert(param != NO_PARAM);
memset(lookupTable, -1, sizeof(lookupTable));
for (int i = 0; i < allMasks.size(); i++) {
uint32_t hash = HashFunction(param, allMasks[i], HASH_BITS);
lookupTable[hash] = allPerms[i];
}
return param;
}
int main() {
uint32_t param = GenerateLookupTableHash();
printf("lookupTable (%08X):\n", param);
PrintTable(lookupTable, 1 << HASH_BITS);
return 0;
}
Raw
merge.cpp
#include <stdio.h>
#include <time.h>
#include <string.h>
#include <assert.h>
#include <stdint.h>
#include <limits.h>
#include <algorithm>
#include <emmintrin.h>
#include <immintrin.h>
#include <random>
/*#include <intrin.h>
#include "iacaMarks.h"*/
#undef NDEBUG //asserts are used for correctness checks
#ifdef _MSC_VER
#define ALIGN(n) __declspec(align(n))
#else
#define ALIGN(n) __attribute__((aligned(n)))
#endif
//equivalent to _MM_SHUFFLE
//(I hate the Intel guy who introduced big-endian notation in intrinsics)
#define SHUF(i0, i1, i2, i3) (i0 + i1*4 + i2*16 + i3*64)
#if defined(__x86_64__) || defined(_WIN64)
#define _mm_cvtsi128_sz _mm_cvtsi128_si64
#else
#define _mm_cvtsi128_sz _mm_cvtsi128_si32
#endif
//for GCC: https://stackoverflow.com/a/32630658/556899
#ifndef _mm256_setr_m128i
#define _mm256_setr_m128i(v0, v1) _mm256_insertf128_si256(_mm256_castsi128_si256(v0), (v1), 1)
#endif
//=========================== merge: scalar =============================
int Merge_Std(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
std::merge(aArr, aArr + aCnt, bArr, bArr + bCnt, dst);
return aCnt + bCnt;
}
int Merge_ScalarTrivial(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
int i = 0, j = 0, k = 0;
while (i < aCnt && j < bCnt) {
if (aArr[i] < bArr[j])
dst[k++] = aArr[i++];
else
dst[k++] = bArr[j++];
}
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
int Merge_ScalarOptimized_2(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
int i = 0, j = 0, k = 0;
while (i < aCnt - 32 && j < bCnt - 32) {
for (int t = 0; t < 32; t++) {
if (aArr[i] < bArr[j])
dst[k++] = aArr[i++];
else
dst[k++] = bArr[j++];
}
}
while (i < aCnt && j < bCnt) {
if (aArr[i] < bArr[j])
dst[k++] = aArr[i++];
else
dst[k++] = bArr[j++];
}
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
int Merge_ScalarOptimized_3(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
int lastA = (aCnt > 0 ? aArr[aCnt-1] : INT_MIN);
int lastB = (bCnt > 0 ? bArr[bCnt-1] : INT_MIN);
if (lastA < lastB)
return Merge_ScalarOptimized_3(bArr, bCnt, aArr, aCnt, dst);
int aCap = aCnt;
while (aCap > 0 && aArr[aCap-1] >= lastB)
aCap--;
int i = 0, j = 0, k = 0;
int aX = aArr[0], bX = bArr[0];
for (; k < aCap + bCnt; k++) {
if (aX < bX) {
dst[k] = aX;
aX = aArr[++i];
}
else {
dst[k] = bX;
bX = bArr[++j];
}
}
assert(i == aCnt || j == bCnt);
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
int Merge_ScalarBranchless_3(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
int lastA = (aCnt > 0 ? aArr[aCnt-1] : INT_MIN);
int lastB = (bCnt > 0 ? bArr[bCnt-1] : INT_MIN);
if (lastA < lastB)
return Merge_ScalarBranchless_3(bArr, bCnt, aArr, aCnt, dst);
int aCap = aCnt;
while (aCap > 0 && aArr[aCap-1] >= lastB)
aCap--;
size_t i = 0, j = 0, k = 0;
for (; k < aCap + bCnt; k++) {
int aX = aArr[i], bX = bArr[j];
dst[k] = (aX < bX ? aX : bX);
i += (aX < bX);
j += 1 - (aX < bX);
}
assert(i == aCnt || j == bCnt);
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
int Merge_ScalarBranchless_2(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
size_t i = 0, j = 0, k = 0;
while (i < aCnt - 32 && j < bCnt - 32) {
for (int t = 0; t < 32; t++) {
int aX = aArr[i], bX = bArr[j];
dst[k++] = (aX < bX ? aX : bX);
i += (aX < bX);
j += 1 - (aX < bX);
}
}
while (i < aCnt && j < bCnt) {
if (aArr[i] < bArr[j])
dst[k++] = aArr[i++];
else
dst[k++] = bArr[j++];
}
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
//=========================== merge: vector =============================
ALIGN(4096) uint8_t lookupTable[128][32] = {
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF},
{0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00},
{0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00}
};
int Merge_Simd_KeysOnly(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
intptr_t i = aCnt - 1, j = bCnt - 1, k = aCnt + bCnt - 1;
while (i >= 0 && j >= 0 && (i >= aCnt - 8 || j >= bCnt - 8)) {
if (aArr[i] < bArr[j])
dst[k--] = bArr[j--];
else
dst[k--] = aArr[i--];
//TODO: any vectorized code if loop does not finish soon?
}
i++, j++, k++;
if (i == 0) {
memcpy(dst, bArr, j * sizeof(dst[0]));
return aCnt + bCnt;
}
if (j == 0) {
memcpy(dst, aArr, i * sizeof(dst[0]));
return aCnt + bCnt;
}
size_t maxK = k, tail = aCnt + bCnt - k;
const int *aPtr = aArr;
const int *bPtr = bArr;
intptr_t abDiff = intptr_t(bArr) - intptr_t(aArr);
__m128i srcA = _mm_load_si128((__m128i*)aPtr);
__m128i srcB = _mm_load_si128((__m128i*)bPtr);
for (k = 0; k < maxK; ) {
__m128i mask0 = _mm_cmplt_epi32(srcA, srcB);
__m128i mask1 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(1, 2, 3, 0)));
__m128i mask2 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(2, 3, 0, 1)));
__m128i mask3 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(3, 0, 1, 2)));
__m128i mask = _mm_packs_epi16(_mm_packs_epi16(mask0, mask1), _mm_packs_epi16(mask2, mask3));
uint32_t bits = _mm_movemask_epi8(mask);
uint32_t hash = ((bits * 0x256150A9U) >> 25);
__m256i perm = ((__m256i*)lookupTable)[hash];
__m256i allSrc = _mm256_setr_m128i(srcA, srcB);
__m256i sorted = _mm256_permutevar8x32_epi32(allSrc, perm);
__m128i sortedLow = _mm256_castsi256_si128(sorted);
__m128i sortedHigh = _mm256_extractf128_si256(sorted, 1);
size_t srcAless = _mm_cvtsi128_sz(_mm_shuffle_epi32(mask0, SHUF(3, 3, 3, 3)));
//conditionally swap ptrA and ptrB (abDiff)
aPtr = (const int *)(size_t(aPtr) + (abDiff & srcAless));
abDiff ^= srcAless; abDiff -= srcAless;
abDiff += 4 * sizeof(int);
bPtr = (const int *)(size_t(aPtr) + abDiff);
srcA = sortedHigh;
srcB = _mm_load_si128((__m128i*)bPtr);
_mm_store_si128((__m128i*)(dst + k), sortedLow);
k += 4;
}
return aCnt + bCnt;
}
//early version: do not use it =)
int Merge_Simd_Trash(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
size_t i = 0, j = 0, k = 0;
const int *aBlock = aArr + aCnt - 128;
const int *bBlock = bArr + bCnt - 128;
const int *aPtr = aArr;
const int *bPtr = bArr;
__m128i srcA = _mm_load_si128((__m128i*)aPtr);
__m128i srcB = _mm_load_si128((__m128i*)bPtr);
size_t revPtr = 0;
while (aPtr < aBlock && bPtr < bBlock) {
for (int t = 0; t < 32; t++) {
__m128i mask0 = _mm_cmplt_epi32(srcA, srcB);
__m128i mask1 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(1, 2, 3, 0)));
__m128i mask2 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(2, 3, 0, 1)));
__m128i mask3 = _mm_cmplt_epi32(srcA, _mm_shuffle_epi32(srcB, SHUF(3, 0, 1, 2)));
__m128i mask = _mm_packs_epi16(_mm_packs_epi16(mask0, mask1), _mm_packs_epi16(mask2, mask3));
uint32_t bits = _mm_movemask_epi8(mask);
uint32_t hash = ((bits * 0x256150A9U) >> 25);
__m256i perm = ((__m256i*)lookupTable)[hash];
__m256i allSrc = _mm256_setr_m128i(srcA, srcB);
__m256i sorted = _mm256_permutevar8x32_epi32(allSrc, perm);
__m128i sortedLow = _mm256_castsi256_si128(sorted);
__m128i sortedHigh = _mm256_extractf128_si256(sorted, 1);
size_t srcAless = _mm_cvtsi128_sz(_mm_shuffle_epi32(mask0, SHUF(3, 3, 3, 3)));
size_t ptrAless = revPtr ^ srcAless; //0 or -1
bPtr += ptrAless*4 + 4;
aPtr -= ptrAless*4;
srcA = sortedHigh;
srcB = _mm_load_si128((__m128i*)(ptrAless ? aPtr : bPtr));
revPtr = ptrAless;
_mm_store_si128((__m128i*)(dst + k), sortedLow);
k += 4;
}
}
i = aPtr - aArr;
j = bPtr - bArr;
if (revPtr) std::swap(srcA, srcB);
int tempA[4], tempB[4];
_mm_storeu_si128((__m128i*)tempA, srcA);
_mm_storeu_si128((__m128i*)tempB, srcB);
int t;
if (memcmp(tempA, aPtr, sizeof(tempA)) == 0) {
for (t = 3; t >= 0; t--)
if (std::find(tempB, tempB+4, bPtr[t]) == tempB+4)
break;
t++;
i -= t;
j += t;
}
else if (memcmp(tempB, bPtr, sizeof(tempB)) == 0) {
for (t = 3; t >= 0; t--)
if (std::find(tempA, tempA+4, aPtr[t]) == tempA+4)
break;
t++;
i += t;
j -= t;
}
else assert(0);
while (i < aCnt && j < bCnt) {
if (aArr[i] < bArr[j])
dst[k++] = aArr[i++];
else
dst[k++] = bArr[j++];
}
memcpy(dst + k, aArr + i, (aCnt - i) * sizeof(dst[0]));
k += (aCnt - i);
memcpy(dst + k, bArr + j, (bCnt - j) * sizeof(dst[0]));
k += (bCnt - j);
return k;
}
//============================= testing ===============================
std::mt19937 rnd;
uint32_t trandom(uint32_t maxValue) {
return std::uniform_int_distribution<uint32_t>(0, maxValue)(rnd);
}
//number of elements in each array
const int MAXN = 16 << 10;
int an, bn, resn;
ALIGN(16) int a[MAXN], b[MAXN], res[2 * MAXN];
int check_res[2 * MAXN];
//input arrays generation (random)
void GenMergeInputs() {
an = bn = MAXN;
for (int i = 0; i < an; i++)
a[i] = trandom(3 * MAXN);
for (int i = 0; i < bn; i++)
b[i] = trandom(3 * MAXN);
std::sort(a, a + an);
std::sort(b, b + bn);
// an = std::unique(a, a + an) - a;
// bn = std::unique(b, b + bn) - b;
}
void TestMergeFunction(int (*mergeFunction)(const int*, int, const int*, int, int*), const char *name, bool checkCorrectness = false) {
static const int ELEMENTS = 100000000; //how long to test
int tries = ELEMENTS / MAXN;
printf("Testing %s:\n", name);
int start = clock();
for (int q = 0; q < tries; q++) {
if (checkCorrectness) {
GenMergeInputs();
Merge_Std(a, an, b, bn, check_res);
}
resn = mergeFunction(a, an, b, bn, res);
if (checkCorrectness) {
assert(resn == an + bn);
for (int i = 0; i < resn; i++)
assert(res[i] == check_res[i]);
}
}
int end = clock();
double elapsed = double(end - start) / CLOCKS_PER_SEC;
if (elapsed <= 0.1)
printf(" WARNING: testing took too little time (%0.3lf ms), timings may be imprecise!\n", elapsed * 1e+3);
if (checkCorrectness)
printf(" WARNING: correctness checking mode, timings are incorrect!\n");
printf(" Time per element: %0.2lf ns\n", 1e+9 * elapsed / (tries * (an + bn)));
printf("\n");
assert(std::is_sorted(res, res+resn));
}
#define FUNC(func) &func,#func
int main() {
GenMergeInputs();
//for stress-testing correctness of merge
//TestMergeFunction(FUNC(Merge_Simd_KeysOnly), true);
TestMergeFunction(FUNC(Merge_Simd_KeysOnly));
TestMergeFunction(FUNC(Merge_Simd_Trash));
TestMergeFunction(FUNC(Merge_ScalarBranchless_3));
TestMergeFunction(FUNC(Merge_ScalarBranchless_2));
TestMergeFunction(FUNC(Merge_ScalarOptimized_3));
TestMergeFunction(FUNC(Merge_ScalarOptimized_2));
TestMergeFunction(FUNC(Merge_ScalarTrivial));
TestMergeFunction(FUNC(Merge_Std));
return 0;
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment