1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
|
package shared;
import jdk.incubator.vector.ByteVector;
import jdk.incubator.vector.DoubleVector;
import jdk.incubator.vector.FloatVector;
import jdk.incubator.vector.IntVector;
import jdk.incubator.vector.LongVector;
import jdk.incubator.vector.ShortVector;
import jdk.incubator.vector.VectorMask;
import jdk.incubator.vector.VectorOperators;
import jdk.incubator.vector.VectorSpecies;
import ml.Cell;
/**
* Holds SIMD methods.
* @author Brian Bushnell
* @date Sep 12, 2013
*
*/
final class SIMD {
//Example from https://medium.com/@Styp/java-18-vector-api-do-we-get-free-speed-up-c4510eda50d2
@SuppressWarnings("restriction")
private static final VectorSpecies<Float> FSPECIES=FloatVector.SPECIES_256;//FloatVector.SPECIES_PREFERRED; //This needs to be final or performance drops.
private static final int FWIDTH=FSPECIES.length();
// private static final int boundMask=~(FWIDTH-1);
@SuppressWarnings("restriction")
private static final VectorSpecies<Byte> BSPECIES=ByteVector.SPECIES_256;
private static final int BWIDTH=BSPECIES.length();
@SuppressWarnings("restriction")
private static final VectorSpecies<Integer> ISPECIES=IntVector.SPECIES_256;
private static final int IWIDTH=ISPECIES.length();
@SuppressWarnings("restriction")
private static final VectorSpecies<Short> SSPECIES=ShortVector.SPECIES_256;
private static final int SWIDTH=SSPECIES.length();
@SuppressWarnings("restriction")
private static final VectorSpecies<Double> DSPECIES=DoubleVector.SPECIES_256;
private static final int DWIDTH=DSPECIES.length();
@SuppressWarnings("restriction")
private static final VectorSpecies<Long> LSPECIES=LongVector.SPECIES_256;
private static final int LWIDTH=LSPECIES.length();
@SuppressWarnings("restriction")
/**
* Vectorized version of "c+=a[i]*b[i]" where a and b are equal-length arrays.
* @param a A vector to multiply.
* @param b A vector to multiply.
* @return Sum of products of vector elements.
*/
static final float fma(final float[] a, final float[] b){
assert(a.length==b.length);
//Note: FSPECIES=FloatVector.SPECIES_256 and FWIDTH=8
final int limit=FSPECIES.loopBound(a.length);
FloatVector sum=FloatVector.zero(FSPECIES);
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
sum=va.fma(vb, sum);
}
float c=sum.reduceLanes(VectorOperators.ADD);
for (; i<a.length; i++) {//Residual scalar loop
c+=a[i]*b[i];
}
return c;
}
@SuppressWarnings("restriction")
/**
* Vectorized version of "c+=a[i]*b[bSet[i]]" where a and bSet are equal-length arrays,
* and bSet stores indices of b, in ascending contiguous blocks of 8.
* @param a A vector to multiply.
* @param b A vector to multiply.
* @return Sum of products of vector elements.
*/
static final float fmaSparse(final float[] a, final float[] b, int[] bSet){
assert(a.length==bSet.length);
assert(a.length<b.length);//Otherwise should do normal fma
//Note: FSPECIES=FloatVector.SPECIES_256 and FWIDTH=8
final int limit=FSPECIES.loopBound(bSet.length);
// assert(FWIDTH==8);
// int elements=0;
FloatVector sum=FloatVector.zero(FSPECIES);
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
int idx=bSet[i];
// elements+=FWIDTH;
// assert(idx%8==0) : idx+", "+i+", "+Arrays.toString(bSet);
// assert(bSet[i+1]==idx+1);
// assert(bSet[i+7]==idx+7);
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, idx);
sum=va.fma(vb, sum);
}
float c=sum.reduceLanes(VectorOperators.ADD);
for (; i<bSet.length; i++) {//Residual scalar loop
// elements++;
c+=a[i]*b[bSet[i]];
}
// float c2=0;
// for (int j=0; j<bSet.length; j++) {//Verification loop
// c2+=a[j]*b[bSet[j]];
// }
// assert(Tools.absdif(c, c2)<0.0001f) : c+", "+c2;
// assert(elements==bSet.length);
return c;
}
/**
* This is here to keep all the vector operations in a single loop,
* to prevent going in and out of SIMD mode too often... hopefully.
* ~20% measured speed increase compared to calling fma() for ScoreSequence.
*/
@SuppressWarnings("restriction")
public static void feedForward(final Cell[] layer, final float[] b) {
assert(false) : "This was giving incorrect results for nets made made with simd=f and vice versa. Needs validation.";
final int limit=FSPECIES.loopBound(b.length);
for(int cnum=0; cnum<layer.length; cnum++) {
final Cell cell=layer[cnum];
final float[] a=cell.weights;
FloatVector sum=FloatVector.zero(FSPECIES);
for(int i=0; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
sum=va.fma(vb, sum);
}
cell.sum=sum.reduceLanes(VectorOperators.ADD);
}
for(int cnum=0; cnum<layer.length; cnum++) {
final Cell cell=layer[cnum];
final float[] a=cell.weights;
float residual=cell.bias;
for (int i=limit+FWIDTH; i<a.length; i++) {//Residual scalar loop
residual+=a[i]*b[i];
}
cell.sum+=residual;
final float v=(float)cell.activation(cell.sum);
cell.setValue(v);
}
}
/**
* This is here to keep all the vector operations in a single loop,
* to prevent going in and out of SIMD mode too often... hopefully.
* ~20% measured speed increase compared to calling fma() for Train.
*/
public static void backPropFma(Cell[] layer, float[] a, float[][] bb) {
final int limit=FSPECIES.loopBound(a.length);
for(int cnum=0; cnum<layer.length; cnum++) {
Cell cell=layer[cnum];
float[] b=bb[cnum];
FloatVector sum=FloatVector.zero(FSPECIES);
for(int i=0; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
sum=va.fma(vb, sum);
}
cell.eTotalOverOut=sum.reduceLanes(VectorOperators.ADD);
}
if(limit+FWIDTH>=a.length) {return;}//Shortcut when length is divisible by 8.
for(int cnum=0; cnum<layer.length; cnum++) {
Cell cell=layer[cnum];
float[] b=bb[cnum];
float residual=0;
for (int i=limit+FWIDTH; i<a.length; i++) {//Residual scalar loop
residual+=a[i]*b[i];
}
cell.eTotalOverOut+=residual;
}
}
/**
* Performs "a+=incr" where a and incr are equal-length arrays.
* @param a A vector to increment.
* @param b Increment amount.
*/
@SuppressWarnings("restriction")
static final void add(final float[] a, final float[] b){
//final int width=SPECIES.length();
final int limit=FSPECIES.loopBound(a.length);
// final int limit=a.length&boundMask;
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
FloatVector sum=va.add(vb);
sum.intoArray(a, i);
}
for (; i<a.length; i++) {//Residual scalar loop; TODO: replace with vector mask
a[i]+=b[i];
}
}
/**
* Performs "a+=b*mult" where a and b are equal-length arrays.
* @param a A vector to increment.
* @param b Increment amount.
* @param mult Increment multiplier.
*/
@SuppressWarnings("restriction")
static final void addProduct(final float[] a, final float[] b, final float mult){
//final int width=SPECIES.length();
final int limit=FSPECIES.loopBound(a.length);
// final int limit=a.length&boundMask;
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
FloatVector sum=va.add(vb.mul(mult));
sum.intoArray(a, i);
}
for (; i<a.length; i++) {//Residual scalar loop; TODO: replace with vector mask
a[i]+=b[i]*mult;
}
}
@SuppressWarnings("restriction")
static final void addProductSparse(final float[] a, final float[] b, final int[] bSet, final float mult){
//final int width=SPECIES.length();
final int limit=FSPECIES.loopBound(bSet.length);
// final int limit=a.length&boundMask;
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
int idx=bSet[i];
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
FloatVector vb=FloatVector.fromArray(FSPECIES, b, idx);
FloatVector sum=va.add(vb.mul(mult));
sum.intoArray(a, i);
}
for (; i<bSet.length; i++) {//Residual scalar loop; TODO: replace with vector mask
a[i]+=b[bSet[i]]*mult;
}
}
//a is dest
@SuppressWarnings("restriction")
static final void copy(final float[] a, final float[] b){
final int length=Tools.min(a.length, b.length);
//final int width=SPECIES.length();
final int limit=FSPECIES.loopBound(length);
// final int limit=a.length&boundMask;
int i=0;
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector vb=FloatVector.fromArray(FSPECIES, b, i);
vb.intoArray(a, i);
}
for (; i<length; i++) {//Residual scalar loop; TODO: replace with vector mask
a[i]=b[i];
}
}
/** Returns number of matches */
@SuppressWarnings("restriction")
static final int countMatches(final byte[] s1, final byte[] s2, int a1, int b1, int a2, int b2){
final int length=b2-a2+1;
final int limit0=BSPECIES.loopBound(length);
final int limit=a2+limit0;
int i=a1, j=a2;
int matches=0;
for(; j<limit; i+=BWIDTH, j+=BWIDTH) {//SIMD loop
ByteVector v1=ByteVector.fromArray(BSPECIES, s1, i);
ByteVector v2=ByteVector.fromArray(BSPECIES, s2, j);
VectorMask<Byte> x=v1.eq(v2);
matches+=x.trueCount();//This might be slow, or might not
}
for(; j<=b2; i++, j++) {
final byte x=s1[i], y=s2[j];
final int m=((x==y) ? 1 : 0);
matches+=m;
}
return matches;
}
/** Returns index of symbol */
@SuppressWarnings("restriction")
static final int find(final byte[] a, final byte symbol, final int from, final int to){//15% Slower than scalar code, at least for ByteFile1
final int length=to-from;//Intentionally exclusive
final int limit0=BSPECIES.loopBound(length);
final int limit=from+limit0;
int pos=from;
for(; pos<limit; pos+=BWIDTH) {//SIMD loop
ByteVector v=ByteVector.fromArray(BSPECIES, a, pos);
VectorMask<Byte> x=v.eq(symbol);
int t=x.firstTrue();
if(t<BWIDTH) {return pos+t;}
// if(x.anyTrue()) {break;}
}
while(pos<to && a[pos]!=symbol){pos++;}
return pos;
}
@SuppressWarnings("restriction")
/**
* Sums the array.
* @param a A vector.
* @return The sum.
*/
static final float sum(final float[] a, final int from, final int to){
final int length=to-from+1;//Intentionally inclusive
final int limit0=FSPECIES.loopBound(length);
final int limit=from+limit0;
FloatVector sum=FloatVector.zero(FSPECIES);
int i=from;
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
sum=sum.add(va);
}
float c=sum.reduceLanes(VectorOperators.ADD);
for (; i<=to; i++) {c+=a[i];}//Residual scalar loop
return c;
}
@SuppressWarnings("restriction")
/**
* Sums the array.
* @param a A vector.
* @return The sum.
*/
static final long sum(final long[] a, final int from, final int to){
final int length=to-from+1;
final int limit0=LSPECIES.loopBound(length);
final int limit=from+limit0;
LongVector sum=LongVector.zero(LSPECIES);
int i=from;
for(; i<limit; i+=LWIDTH) {//SIMD loop
LongVector va=LongVector.fromArray(LSPECIES, a, i);
sum=sum.add(va);
}
long c=sum.reduceLanes(VectorOperators.ADD);
for (; i<=to; i++) {c+=a[i];}//Residual scalar loop
return c;
}
@SuppressWarnings("restriction")
/**
* Sums the array.
* @param a A vector.
* @return The sum.
*/
static final long sum(final int[] a, final int from, final int to){//Tested as 1.5x scalar speed
final int length=to-from+1;
final int limit0=ISPECIES.loopBound(length);
final int limit=from+limit0;
int i=from;
long c=0;
for(; i<limit; i+=IWIDTH) {//SIMD loop
IntVector va=IntVector.fromArray(ISPECIES, a, i);
c+=va.reduceLanesToLong(VectorOperators.ADD);//This is probably slow
}
for (; i<=to; i++) {c+=a[i];}//Residual scalar loop
return c;
}
@SuppressWarnings("restriction")
/**
* Sums the array.
* @param a A vector.
* @return The sum.
*/
static final long sum(final byte[] a, final int from, final int to){//Tested as 4x scalar speed
//TODO: Test speed.
final int length=to-from+1;
final int limit0=BSPECIES.loopBound(length);
final int limit=from+limit0;
int i=from;
long c=0;
for(; i<limit; i+=BWIDTH) {//SIMD loop
ByteVector va=ByteVector.fromArray(BSPECIES, a, i);
c+=va.reduceLanesToLong(VectorOperators.ADD);
}
for (; i<=to; i++) {c+=a[i];}//Residual scalar loop
return c;
}
@SuppressWarnings("restriction")
/**
* Sums the array.
* @param a A vector.
* @return The sum.
*/
static final double sum(final double[] a, final int from, final int to){
final int length=to-from+1;
final int limit0=DSPECIES.loopBound(length);
final int limit=from+limit0;
DoubleVector sum=DoubleVector.zero(DSPECIES);
int i=from;
for(; i<limit; i+=DWIDTH) {//SIMD loop
DoubleVector va=DoubleVector.fromArray(DSPECIES, a, i);
sum=sum.add(va);
}
double c=sum.reduceLanes(VectorOperators.ADD);
for (; i<=to; i++) {c+=a[i];}//Residual scalar loop
return c;
}
@SuppressWarnings("restriction")
/**
* Finds the maximum value.
* @param a A vector.
* @return The max.
*/
static final int max(final int[] a, final int from, final int to){//Tested as 5x scalar speed
final int length=to-from+1;
final int limit0=ISPECIES.loopBound(length);
final int limit=from+limit0;
int i=from;
IntVector max=IntVector.broadcast(ISPECIES, a[from]);
for(; i<limit; i+=IWIDTH) {//SIMD loop
IntVector va=IntVector.fromArray(ISPECIES, a, i);
max=max.max(va);
}
int c=max.reduceLanes(VectorOperators.MAX);
for (; i<=to; i++) {//Residual scalar loop
final int x=a[i];
c=(x>c ? x : c);
}
return c;
}
@SuppressWarnings("restriction")
/**
* Finds the maximum value.
* @param a A vector.
* @return The max.
*/
static final long max(final long[] a, final int from, final int to){
final int length=to-from+1;
final int limit0=LSPECIES.loopBound(length);
final int limit=from+limit0;
int i=from;
LongVector max=LongVector.broadcast(LSPECIES, a[from]);
for(; i<limit; i+=LWIDTH) {//SIMD loop
LongVector va=LongVector.fromArray(LSPECIES, a, i);
max=max.max(va);
}
long c=max.reduceLanes(VectorOperators.MAX);
for (; i<=to; i++) {//Residual scalar loop
final long x=a[i];
c=(x>c ? x : c);
}
return c;
}
@SuppressWarnings("restriction")
/**
* Finds the maximum value.
* @param a A vector.
* @return The max.
*/
static final float max(final float[] a, final int from, final int to){
final int length=to-from+1;
final int limit0=FSPECIES.loopBound(length);
final int limit=from+limit0;
int i=from;
FloatVector max=FloatVector.broadcast(FSPECIES, a[from]);
for(; i<limit; i+=FWIDTH) {//SIMD loop
FloatVector va=FloatVector.fromArray(FSPECIES, a, i);
max=max.max(va);
}
float c=max.reduceLanes(VectorOperators.MAX);
for (; i<=to; i++) {//Residual scalar loop
final float x=a[i];
c=(x>c ? x : c);
}
return c;
}
}
|