| 1 |
//go:build generate |
| 2 |
// +build generate |
| 3 |
|
| 4 |
// Copyright 2022+, Klaus Post. See LICENSE for details. |
| 5 |
|
| 6 |
package main |
| 7 |
|
| 8 |
import ( |
| 9 |
"fmt" |
| 10 |
|
| 11 |
"github.com/mmcloughlin/avo/attr" |
| 12 |
. "github.com/mmcloughlin/avo/build" |
| 13 |
. "github.com/mmcloughlin/avo/operand" |
| 14 |
"github.com/mmcloughlin/avo/reg" |
| 15 |
) |
| 16 |
|
| 17 |
type gf8ctx struct { |
| 18 |
clrMask reg.VecVirtual |
| 19 |
clrMask128 reg.VecVirtual |
| 20 |
} |
| 21 |
|
| 22 |
func genGF8() { |
| 23 |
var ctx gf8ctx |
| 24 |
// Ported from static void IFFT_DIT2 |
| 25 |
// https://github.com/catid/leopard/blob/master/LeopardFF8.cpp#L599 |
| 26 |
if pshufb { |
| 27 |
TEXT("ifftDIT28_avx2", attr.NOSPLIT, fmt.Sprintf("func(x, y []byte, table *[2*16]uint8)")) |
| 28 |
Pragma("noescape") |
| 29 |
|
| 30 |
tablePtr := Load(Param("table"), GP64()) |
| 31 |
var tables table256 |
| 32 |
tables.Lo, tables.Hi = YMM(), YMM() |
| 33 |
// Load and expand tables |
| 34 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, tables.Lo) |
| 35 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, tables.Hi) |
| 36 |
|
| 37 |
bytes := Load(Param("x").Len(), GP64()) |
| 38 |
x := Load(Param("x").Base(), GP64()) |
| 39 |
y := Load(Param("y").Base(), GP64()) |
| 40 |
// Generate mask |
| 41 |
ctx.clrMask = YMM() |
| 42 |
tmpMask := GP64() |
| 43 |
MOVQ(U32(15), tmpMask) |
| 44 |
MOVQ(tmpMask, ctx.clrMask.AsX()) |
| 45 |
VPBROADCASTB(ctx.clrMask.AsX(), ctx.clrMask) |
| 46 |
|
| 47 |
x0, x1, y0, y1 := YMM(), YMM(), YMM(), YMM() |
| 48 |
Label("loop") |
| 49 |
VMOVDQU(Mem{Base: x, Disp: 0}, x0) |
| 50 |
VMOVDQU(Mem{Base: x, Disp: 32}, x1) |
| 51 |
VMOVDQU(Mem{Base: y, Disp: 0}, y0) |
| 52 |
VMOVDQU(Mem{Base: y, Disp: 32}, y1) |
| 53 |
|
| 54 |
// Update y and store |
| 55 |
VPXOR(y0, x0, y0) |
| 56 |
VPXOR(y1, x1, y1) |
| 57 |
VMOVDQU(y0, Mem{Base: y, Disp: 0}) |
| 58 |
VMOVDQU(y1, Mem{Base: y, Disp: 32}) |
| 59 |
|
| 60 |
// Update x and store |
| 61 |
leo8MulAdd256(ctx, x0, y0, tables) |
| 62 |
leo8MulAdd256(ctx, x1, y1, tables) |
| 63 |
VMOVDQU(x0, Mem{Base: x, Disp: 0}) |
| 64 |
VMOVDQU(x1, Mem{Base: x, Disp: 32}) |
| 65 |
|
| 66 |
// Move on |
| 67 |
ADDQ(U8(64), x) |
| 68 |
ADDQ(U8(64), y) |
| 69 |
SUBQ(U8(64), bytes) |
| 70 |
JA(LabelRef("loop")) |
| 71 |
|
| 72 |
VZEROUPPER() |
| 73 |
RET() |
| 74 |
} |
| 75 |
// https://github.com/catid/leopard/blob/master/LeopardFF8.cpp#L1323 |
| 76 |
if pshufb { |
| 77 |
TEXT("fftDIT28_avx2", attr.NOSPLIT, fmt.Sprintf("func(x, y []byte, table *[2*16]uint8)")) |
| 78 |
Pragma("noescape") |
| 79 |
tablePtr := Load(Param("table"), GP64()) |
| 80 |
var tables table256 |
| 81 |
tables.Lo, tables.Hi = YMM(), YMM() |
| 82 |
// Load and expand tables |
| 83 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, tables.Lo) |
| 84 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, tables.Hi) |
| 85 |
|
| 86 |
bytes := Load(Param("x").Len(), GP64()) |
| 87 |
x := Load(Param("x").Base(), GP64()) |
| 88 |
y := Load(Param("y").Base(), GP64()) |
| 89 |
// Generate mask |
| 90 |
ctx.clrMask = YMM() |
| 91 |
tmpMask := GP64() |
| 92 |
MOVQ(U32(15), tmpMask) |
| 93 |
MOVQ(tmpMask, ctx.clrMask.AsX()) |
| 94 |
VPBROADCASTB(ctx.clrMask.AsX(), ctx.clrMask) |
| 95 |
|
| 96 |
x0, x1, y0, y1 := YMM(), YMM(), YMM(), YMM() |
| 97 |
Label("loop") |
| 98 |
VMOVDQU(Mem{Base: x, Disp: 0}, x0) |
| 99 |
VMOVDQU(Mem{Base: x, Disp: 32}, x1) |
| 100 |
VMOVDQU(Mem{Base: y, Disp: 0}, y0) |
| 101 |
VMOVDQU(Mem{Base: y, Disp: 32}, y1) |
| 102 |
|
| 103 |
leo8MulAdd256(ctx, x0, y0, tables) |
| 104 |
leo8MulAdd256(ctx, x1, y1, tables) |
| 105 |
VMOVDQU(x0, Mem{Base: x, Disp: 0}) |
| 106 |
VMOVDQU(x1, Mem{Base: x, Disp: 32}) |
| 107 |
|
| 108 |
VPXOR(y0, x0, y0) |
| 109 |
VPXOR(y1, x1, y1) |
| 110 |
VMOVDQU(y0, Mem{Base: y, Disp: 0}) |
| 111 |
VMOVDQU(y1, Mem{Base: y, Disp: 32}) |
| 112 |
ADDQ(U8(64), x) |
| 113 |
ADDQ(U8(64), y) |
| 114 |
SUBQ(U8(64), bytes) |
| 115 |
JA(LabelRef("loop")) |
| 116 |
|
| 117 |
VZEROUPPER() |
| 118 |
RET() |
| 119 |
} |
| 120 |
|
| 121 |
x := [8]int{} |
| 122 |
for _, withDst := range []bool{false, true} { |
| 123 |
dstString := "" |
| 124 |
dstSuf := "" |
| 125 |
if withDst { |
| 126 |
dstString = "dst, " |
| 127 |
dstSuf = "dst_" |
| 128 |
} |
| 129 |
for skipMask := range x[:] { |
| 130 |
if !pshufb { |
| 131 |
break |
| 132 |
} |
| 133 |
{ |
| 134 |
var suffix = "avx2_" + dstSuf + fmt.Sprint(skipMask) |
| 135 |
TEXT("ifftDIT48_"+suffix, attr.NOSPLIT, fmt.Sprintf("func(%swork [][]byte, dist int, t01, t23, t02 *[2*16]uint8)", dstString)) |
| 136 |
Pragma("noescape") |
| 137 |
var t01, t23, t02 table256 |
| 138 |
// Load and expand tables |
| 139 |
|
| 140 |
if (skipMask & 1) == 0 { |
| 141 |
tablePtr := Load(Param("t01"), GP64()) |
| 142 |
t01.Lo, t01.Hi = YMM(), YMM() |
| 143 |
// We need one register when loading all. |
| 144 |
if skipMask == 0 { |
| 145 |
t01.loadLo128 = &Mem{Base: tablePtr, Disp: 0} |
| 146 |
} else { |
| 147 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t01.Lo) |
| 148 |
} |
| 149 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t01.Hi) |
| 150 |
} |
| 151 |
if (skipMask & 2) == 0 { |
| 152 |
tablePtr := Load(Param("t23"), GP64()) |
| 153 |
t23.Lo, t23.Hi = YMM(), YMM() |
| 154 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t23.Lo) |
| 155 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t23.Hi) |
| 156 |
} |
| 157 |
if (skipMask & 4) == 0 { |
| 158 |
tablePtr := Load(Param("t02"), GP64()) |
| 159 |
t02.Lo, t02.Hi = YMM(), YMM() |
| 160 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t02.Lo) |
| 161 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t02.Hi) |
| 162 |
} |
| 163 |
dist := Load(Param("dist"), GP64()) |
| 164 |
|
| 165 |
var work [4]reg.GPVirtual |
| 166 |
workTable := Load(Param("work").Base(), GP64()) // &work[0] |
| 167 |
bytes := GP64() |
| 168 |
MOVQ(Mem{Base: workTable, Disp: 8}, bytes) |
| 169 |
|
| 170 |
var dst [4]reg.GPVirtual |
| 171 |
dstTable := GP64() |
| 172 |
if withDst { |
| 173 |
Load(Param("dst").Base(), dstTable) // &dst[0] |
| 174 |
} |
| 175 |
for i := range work { |
| 176 |
work[i] = GP64() |
| 177 |
// work[i] = &workTable[dist*i] |
| 178 |
MOVQ(Mem{Base: workTable}, work[i]) |
| 179 |
if i < len(work)-1 { |
| 180 |
ADDQ(dist, workTable) |
| 181 |
} |
| 182 |
if withDst { |
| 183 |
dst[i] = GP64() |
| 184 |
MOVQ(Mem{Base: dstTable}, dst[i]) |
| 185 |
if i < len(work)-1 { |
| 186 |
ADDQ(dist, dstTable) |
| 187 |
} |
| 188 |
} |
| 189 |
} |
| 190 |
|
| 191 |
// Generate mask |
| 192 |
ctx.clrMask = YMM() |
| 193 |
tmpMask := GP64() |
| 194 |
MOVQ(U32(15), tmpMask) |
| 195 |
MOVQ(tmpMask, ctx.clrMask.AsX()) |
| 196 |
VPBROADCASTB(ctx.clrMask.AsX(), ctx.clrMask) |
| 197 |
|
| 198 |
Label("loop") |
| 199 |
var workReg [4]reg.VecVirtual |
| 200 |
var workReg2 [4]reg.VecVirtual |
| 201 |
|
| 202 |
workReg[0] = YMM() |
| 203 |
workReg[1] = YMM() |
| 204 |
workReg2[0] = YMM() |
| 205 |
workReg2[1] = YMM() |
| 206 |
|
| 207 |
VMOVDQU(Mem{Base: work[0], Disp: 0}, workReg[0]) |
| 208 |
VMOVDQU(Mem{Base: work[1], Disp: 0}, workReg[1]) |
| 209 |
VMOVDQU(Mem{Base: work[0], Disp: 32}, workReg2[0]) |
| 210 |
VMOVDQU(Mem{Base: work[1], Disp: 32}, workReg2[1]) |
| 211 |
|
| 212 |
// work1_reg = _mm256_xor_si256(work0_reg, work1_reg); |
| 213 |
VPXOR(workReg[1], workReg[0], workReg[1]) |
| 214 |
VPXOR(workReg2[1], workReg2[0], workReg2[1]) |
| 215 |
if (skipMask & 1) == 0 { |
| 216 |
t01.prepare() |
| 217 |
leo8MulAdd256(ctx, workReg[0], workReg[1], t01) |
| 218 |
leo8MulAdd256(ctx, workReg2[0], workReg2[1], t01) |
| 219 |
} |
| 220 |
|
| 221 |
workReg[2] = YMM() |
| 222 |
workReg[3] = YMM() |
| 223 |
workReg2[2] = YMM() |
| 224 |
workReg2[3] = YMM() |
| 225 |
VMOVDQU(Mem{Base: work[2], Disp: 0}, workReg[2]) |
| 226 |
VMOVDQU(Mem{Base: work[3], Disp: 0}, workReg[3]) |
| 227 |
VMOVDQU(Mem{Base: work[2], Disp: 32}, workReg2[2]) |
| 228 |
VMOVDQU(Mem{Base: work[3], Disp: 32}, workReg2[3]) |
| 229 |
|
| 230 |
//work3_reg = _mm256_xor_si256(work2_reg, work3_reg) |
| 231 |
VPXOR(workReg[2], workReg[3], workReg[3]) |
| 232 |
VPXOR(workReg2[2], workReg2[3], workReg2[3]) |
| 233 |
if (skipMask & 2) == 0 { |
| 234 |
leo8MulAdd256(ctx, workReg[2], workReg[3], t23) |
| 235 |
leo8MulAdd256(ctx, workReg2[2], workReg2[3], t23) |
| 236 |
} |
| 237 |
|
| 238 |
// Second layer: |
| 239 |
// work2_reg = _mm256_xor_si256(work0_reg, work2_reg); |
| 240 |
// work3_reg = _mm256_xor_si256(work1_reg, work3_reg); |
| 241 |
VPXOR(workReg[0], workReg[2], workReg[2]) |
| 242 |
VPXOR(workReg[1], workReg[3], workReg[3]) |
| 243 |
VPXOR(workReg2[0], workReg2[2], workReg2[2]) |
| 244 |
VPXOR(workReg2[1], workReg2[3], workReg2[3]) |
| 245 |
|
| 246 |
if (skipMask & 4) == 0 { |
| 247 |
leo8MulAdd256(ctx, workReg[0], workReg[2], t02) |
| 248 |
leo8MulAdd256(ctx, workReg[1], workReg[3], t02) |
| 249 |
leo8MulAdd256(ctx, workReg2[0], workReg2[2], t02) |
| 250 |
leo8MulAdd256(ctx, workReg2[1], workReg2[3], t02) |
| 251 |
} |
| 252 |
|
| 253 |
// Store + Next loop: |
| 254 |
for i := range work { |
| 255 |
if withDst { |
| 256 |
VMOVDQU(workReg[i], Mem{Base: dst[i], Disp: 0}) |
| 257 |
VMOVDQU(workReg2[i], Mem{Base: dst[i], Disp: 32}) |
| 258 |
ADDQ(U8(64), dst[i]) |
| 259 |
} else { |
| 260 |
VMOVDQU(workReg[i], Mem{Base: work[i], Disp: 0}) |
| 261 |
VMOVDQU(workReg2[i], Mem{Base: work[i], Disp: 32}) |
| 262 |
} |
| 263 |
ADDQ(U8(64), work[i]) |
| 264 |
} |
| 265 |
|
| 266 |
SUBQ(U8(64), bytes) |
| 267 |
JA(LabelRef("loop")) |
| 268 |
|
| 269 |
VZEROUPPER() |
| 270 |
RET() |
| 271 |
} |
| 272 |
// No destination version needed at this point. |
| 273 |
if !withDst { |
| 274 |
var suffix = "avx2_" + dstSuf + fmt.Sprint(skipMask) |
| 275 |
TEXT("fftDIT48_"+suffix, attr.NOSPLIT, fmt.Sprintf("func(%swork [][]byte, dist int, t01, t23, t02 *[2*16]uint8)", dstString)) |
| 276 |
Pragma("noescape") |
| 277 |
var t01, t23, t02 table256 |
| 278 |
// Load and expand tables |
| 279 |
|
| 280 |
if (skipMask & 2) == 0 { |
| 281 |
tablePtr := Load(Param("t01"), GP64()) |
| 282 |
t01.Lo, t01.Hi = YMM(), YMM() |
| 283 |
if skipMask == 0 { |
| 284 |
t01.loadLo128 = &Mem{Base: tablePtr, Disp: 0} |
| 285 |
} else { |
| 286 |
// We need additional registers |
| 287 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t01.Lo) |
| 288 |
} |
| 289 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t01.Hi) |
| 290 |
} |
| 291 |
if (skipMask & 4) == 0 { |
| 292 |
tablePtr := Load(Param("t23"), GP64()) |
| 293 |
t23.Lo, t23.Hi = YMM(), YMM() |
| 294 |
if skipMask == 0 { |
| 295 |
t23.loadLo128 = &Mem{Base: tablePtr, Disp: 0} |
| 296 |
} else { |
| 297 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t23.Lo) |
| 298 |
} |
| 299 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t23.Hi) |
| 300 |
} |
| 301 |
if (skipMask & 1) == 0 { |
| 302 |
tablePtr := Load(Param("t02"), GP64()) |
| 303 |
|
| 304 |
t02.Lo, t02.Hi = YMM(), YMM() |
| 305 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 0}, t02.Lo) |
| 306 |
VBROADCASTI128(Mem{Base: tablePtr, Disp: 16}, t02.Hi) |
| 307 |
} |
| 308 |
dist := Load(Param("dist"), GP64()) |
| 309 |
|
| 310 |
var work [4]reg.GPVirtual |
| 311 |
workTable := Load(Param("work").Base(), GP64()) // &work[0] |
| 312 |
bytes := GP64() |
| 313 |
MOVQ(Mem{Base: workTable, Disp: 8}, bytes) |
| 314 |
|
| 315 |
var dst [4]reg.GPVirtual |
| 316 |
dstTable := GP64() |
| 317 |
if withDst { |
| 318 |
Load(Param("dst").Base(), dstTable) // &dst[0] |
| 319 |
} |
| 320 |
for i := range work { |
| 321 |
work[i] = GP64() |
| 322 |
// work[i] = &workTable[dist*i] |
| 323 |
MOVQ(Mem{Base: workTable}, work[i]) |
| 324 |
if i < len(work)-1 { |
| 325 |
ADDQ(dist, workTable) |
| 326 |
} |
| 327 |
if withDst { |
| 328 |
dst[i] = GP64() |
| 329 |
MOVQ(Mem{Base: dstTable}, dst[i]) |
| 330 |
if i < len(work)-1 { |
| 331 |
ADDQ(dist, dstTable) |
| 332 |
} |
| 333 |
} |
| 334 |
} |
| 335 |
|
| 336 |
// Generate mask |
| 337 |
ctx.clrMask = YMM() |
| 338 |
tmpMask := GP64() |
| 339 |
MOVQ(U32(15), tmpMask) |
| 340 |
MOVQ(tmpMask, ctx.clrMask.AsX()) |
| 341 |
VPBROADCASTB(ctx.clrMask.AsX(), ctx.clrMask) |
| 342 |
|
| 343 |
Label("loop") |
| 344 |
var workReg [4]reg.VecVirtual |
| 345 |
var workReg2 [4]reg.VecVirtual |
| 346 |
|
| 347 |
for i := range workReg { |
| 348 |
workReg[i] = YMM() |
| 349 |
workReg2[i] = YMM() |
| 350 |
} |
| 351 |
|
| 352 |
VMOVDQU(Mem{Base: work[0], Disp: 0}, workReg[0]) |
| 353 |
VMOVDQU(Mem{Base: work[0], Disp: 32}, workReg2[0]) |
| 354 |
VMOVDQU(Mem{Base: work[2], Disp: 0}, workReg[2]) |
| 355 |
VMOVDQU(Mem{Base: work[2], Disp: 32}, workReg2[2]) |
| 356 |
VMOVDQU(Mem{Base: work[1], Disp: 0}, workReg[1]) |
| 357 |
VMOVDQU(Mem{Base: work[1], Disp: 32}, workReg2[1]) |
| 358 |
VMOVDQU(Mem{Base: work[3], Disp: 0}, workReg[3]) |
| 359 |
VMOVDQU(Mem{Base: work[3], Disp: 32}, workReg2[3]) |
| 360 |
|
| 361 |
// work1_reg = _mm256_xor_si256(work0_reg, work1_reg); |
| 362 |
if (skipMask & 1) == 0 { |
| 363 |
leo8MulAdd256(ctx, workReg[0], workReg[2], t02) |
| 364 |
leo8MulAdd256(ctx, workReg2[0], workReg2[2], t02) |
| 365 |
|
| 366 |
leo8MulAdd256(ctx, workReg[1], workReg[3], t02) |
| 367 |
leo8MulAdd256(ctx, workReg2[1], workReg2[3], t02) |
| 368 |
} |
| 369 |
// work2_reg = _mm256_xor_si256(work0_reg, work2_reg); |
| 370 |
// work3_reg = _mm256_xor_si256(work1_reg, work3_reg); |
| 371 |
VPXOR(workReg[0], workReg[2], workReg[2]) |
| 372 |
VPXOR(workReg[1], workReg[3], workReg[3]) |
| 373 |
VPXOR(workReg2[0], workReg2[2], workReg2[2]) |
| 374 |
VPXOR(workReg2[1], workReg2[3], workReg2[3]) |
| 375 |
|
| 376 |
// Second layer: |
| 377 |
if (skipMask & 2) == 0 { |
| 378 |
t01.prepare() |
| 379 |
leo8MulAdd256(ctx, workReg[0], workReg[1], t01) |
| 380 |
leo8MulAdd256(ctx, workReg2[0], workReg2[1], t01) |
| 381 |
} |
| 382 |
//work1_reg = _mm256_xor_si256(work0_reg, work1_reg); |
| 383 |
VPXOR(workReg[1], workReg[0], workReg[1]) |
| 384 |
VPXOR(workReg2[1], workReg2[0], workReg2[1]) |
| 385 |
|
| 386 |
if (skipMask & 4) == 0 { |
| 387 |
t23.prepare() |
| 388 |
leo8MulAdd256(ctx, workReg[2], workReg[3], t23) |
| 389 |
leo8MulAdd256(ctx, workReg2[2], workReg2[3], t23) |
| 390 |
} |
| 391 |
// work3_reg = _mm256_xor_si256(work2_reg, work3_reg); |
| 392 |
VPXOR(workReg[2], workReg[3], workReg[3]) |
| 393 |
VPXOR(workReg2[2], workReg2[3], workReg2[3]) |
| 394 |
|
| 395 |
// Store + Next loop: |
| 396 |
for i := range work { |
| 397 |
if withDst { |
| 398 |
VMOVDQU(workReg[i], Mem{Base: dst[i], Disp: 0}) |
| 399 |
VMOVDQU(workReg2[i], Mem{Base: dst[i], Disp: 32}) |
| 400 |
ADDQ(U8(64), dst[i]) |
| 401 |
} else { |
| 402 |
VMOVDQU(workReg[i], Mem{Base: work[i], Disp: 0}) |
| 403 |
VMOVDQU(workReg2[i], Mem{Base: work[i], Disp: 32}) |
| 404 |
} |
| 405 |
ADDQ(U8(64), work[i]) |
| 406 |
} |
| 407 |
|
| 408 |
SUBQ(U8(64), bytes) |
| 409 |
JA(LabelRef("loop")) |
| 410 |
|
| 411 |
VZEROUPPER() |
| 412 |
RET() |
| 413 |
} |
| 414 |
} |
| 415 |
|
| 416 |
// GFNI |
| 417 |
for skipMask := range x[:] { |
| 418 |
{ |
| 419 |
var suffix = "gfni_" + dstSuf + fmt.Sprint(skipMask) |
| 420 |
TEXT("ifftDIT48_"+suffix, attr.NOSPLIT, fmt.Sprintf("func(%swork [][]byte, dist int, t01, t23, t02 uint64)", dstString)) |
| 421 |
Pragma("noescape") |
| 422 |
var t01, t23, t02 table512 = ZMM(), ZMM(), ZMM() |
| 423 |
// Load and expand tables |
| 424 |
|
| 425 |
if (skipMask & 1) == 0 { |
| 426 |
tablePtr, _ := Param("t01").Resolve() |
| 427 |
VBROADCASTF32X2(tablePtr.Addr, t01) |
| 428 |
} |
| 429 |
if (skipMask & 2) == 0 { |
| 430 |
tablePtr, _ := Param("t23").Resolve() |
| 431 |
VBROADCASTF32X2(tablePtr.Addr, t23) |
| 432 |
} |
| 433 |
if (skipMask & 4) == 0 { |
| 434 |
tablePtr, _ := Param("t02").Resolve() |
| 435 |
VBROADCASTF32X2(tablePtr.Addr, t02) |
| 436 |
} |
| 437 |
dist := Load(Param("dist"), GP64()) |
| 438 |
|
| 439 |
var work [4]reg.GPVirtual |
| 440 |
workTable := Load(Param("work").Base(), GP64()) // &work[0] |
| 441 |
bytes := GP64() |
| 442 |
MOVQ(Mem{Base: workTable, Disp: 8}, bytes) |
| 443 |
|
| 444 |
var dst [4]reg.GPVirtual |
| 445 |
dstTable := GP64() |
| 446 |
if withDst { |
| 447 |
Load(Param("dst").Base(), dstTable) // &dst[0] |
| 448 |
} |
| 449 |
for i := range work { |
| 450 |
work[i] = GP64() |
| 451 |
// work[i] = &workTable[dist*i] |
| 452 |
MOVQ(Mem{Base: workTable}, work[i]) |
| 453 |
if i < len(work)-1 { |
| 454 |
ADDQ(dist, workTable) |
| 455 |
} |
| 456 |
if withDst { |
| 457 |
dst[i] = GP64() |
| 458 |
MOVQ(Mem{Base: dstTable}, dst[i]) |
| 459 |
if i < len(work)-1 { |
| 460 |
ADDQ(dist, dstTable) |
| 461 |
} |
| 462 |
} |
| 463 |
} |
| 464 |
|
| 465 |
Label("loop") |
| 466 |
var workReg [4]reg.VecVirtual |
| 467 |
for i := range workReg[:] { |
| 468 |
workReg[i] = ZMM() |
| 469 |
VMOVDQU64(Mem{Base: work[i], Disp: 0}, workReg[i]) |
| 470 |
} |
| 471 |
|
| 472 |
// First layer: |
| 473 |
// work1_reg = _mm256_xor_si256(work0_reg, work1_reg); |
| 474 |
VXORPD(workReg[1], workReg[0], workReg[1]) |
| 475 |
if (skipMask & 1) == 0 { |
| 476 |
leo8MulAdd512(ctx, workReg[0], workReg[1], t01, nil) |
| 477 |
} |
| 478 |
|
| 479 |
//work3_reg = _mm256_xor_si256(work2_reg, work3_reg) |
| 480 |
VXORPD(workReg[2], workReg[3], workReg[3]) |
| 481 |
if (skipMask & 2) == 0 { |
| 482 |
leo8MulAdd512(ctx, workReg[2], workReg[3], t23, workReg[0]) |
| 483 |
} else { |
| 484 |
// Merged above when run... |
| 485 |
VXORPD(workReg[0], workReg[2], workReg[2]) |
| 486 |
} |
| 487 |
|
| 488 |
// Second layer: |
| 489 |
// work2_reg = _mm256_xor_si256(work0_reg, work2_reg); |
| 490 |
// work3_reg = _mm256_xor_si256(work1_reg, work3_reg); |
| 491 |
VXORPD(workReg[1], workReg[3], workReg[3]) |
| 492 |
|
| 493 |
if (skipMask & 4) == 0 { |
| 494 |
leo8MulAdd512(ctx, workReg[0], workReg[2], t02, nil) |
| 495 |
leo8MulAdd512(ctx, workReg[1], workReg[3], t02, nil) |
| 496 |
} |
| 497 |
|
| 498 |
// Store + Next loop: |
| 499 |
for i := range work { |
| 500 |
if withDst { |
| 501 |
VMOVDQU64(workReg[i], Mem{Base: dst[i], Disp: 0}) |
| 502 |
ADDQ(U8(64), dst[i]) |
| 503 |
} else { |
| 504 |
VMOVDQU64(workReg[i], Mem{Base: work[i], Disp: 0}) |
| 505 |
} |
| 506 |
ADDQ(U8(64), work[i]) |
| 507 |
} |
| 508 |
|
| 509 |
SUBQ(U8(64), bytes) |
| 510 |
JA(LabelRef("loop")) |
| 511 |
|
| 512 |
VZEROUPPER() |
| 513 |
RET() |
| 514 |
} |
| 515 |
|
| 516 |
// No destination version needed at this point. |
| 517 |
if !withDst { |
| 518 |
var suffix = "gfni_" + dstSuf + fmt.Sprint(skipMask) |
| 519 |
TEXT("fftDIT48_"+suffix, attr.NOSPLIT, fmt.Sprintf("func(%swork [][]byte, dist int, t01, t23, t02 uint64)", dstString)) |
| 520 |
Pragma("noescape") |
| 521 |
var t01, t23, t02 table512 = ZMM(), ZMM(), ZMM() |
| 522 |
// Load and expand tables |
| 523 |
|
| 524 |
if (skipMask & 2) == 0 { |
| 525 |
tablePtr, _ := Param("t01").Resolve() |
| 526 |
VBROADCASTF32X2(tablePtr.Addr, t01) |
| 527 |
} |
| 528 |
if (skipMask & 4) == 0 { |
| 529 |
tablePtr, _ := Param("t23").Resolve() |
| 530 |
VBROADCASTF32X2(tablePtr.Addr, t23) |
| 531 |
} |
| 532 |
if (skipMask & 1) == 0 { |
| 533 |
tablePtr, _ := Param("t02").Resolve() |
| 534 |
VBROADCASTF32X2(tablePtr.Addr, t02) |
| 535 |
} |
| 536 |
dist := Load(Param("dist"), GP64()) |
| 537 |
|
| 538 |
var work [4]reg.GPVirtual |
| 539 |
workTable := Load(Param("work").Base(), GP64()) // &work[0] |
| 540 |
bytes := GP64() |
| 541 |
MOVQ(Mem{Base: workTable, Disp: 8}, bytes) |
| 542 |
|
| 543 |
var dst [4]reg.GPVirtual |
| 544 |
dstTable := GP64() |
| 545 |
if withDst { |
| 546 |
Load(Param("dst").Base(), dstTable) // &dst[0] |
| 547 |
} |
| 548 |
for i := range work { |
| 549 |
work[i] = GP64() |
| 550 |
// work[i] = &workTable[dist*i] |
| 551 |
MOVQ(Mem{Base: workTable}, work[i]) |
| 552 |
if i < len(work)-1 { |
| 553 |
ADDQ(dist, workTable) |
| 554 |
} |
| 555 |
if withDst { |
| 556 |
dst[i] = GP64() |
| 557 |
MOVQ(Mem{Base: dstTable}, dst[i]) |
| 558 |
if i < len(work)-1 { |
| 559 |
ADDQ(dist, dstTable) |
| 560 |
} |
| 561 |
} |
| 562 |
} |
| 563 |
|
| 564 |
Label("loop") |
| 565 |
var workReg [4]reg.VecVirtual |
| 566 |
|
| 567 |
for i := range workReg { |
| 568 |
workReg[i] = ZMM() |
| 569 |
VMOVDQU64(Mem{Base: work[i], Disp: 0}, workReg[i]) |
| 570 |
} |
| 571 |
|
| 572 |
// First layer: |
| 573 |
if (skipMask & 1) == 0 { |
| 574 |
leo8MulAdd512(ctx, workReg[0], workReg[2], t02, nil) |
| 575 |
leo8MulAdd512(ctx, workReg[1], workReg[3], t02, nil) |
| 576 |
} |
| 577 |
// work2_reg = _mm256_xor_si256(work0_reg, work2_reg); |
| 578 |
// work3_reg = _mm256_xor_si256(work1_reg, work3_reg); |
| 579 |
VXORPD(workReg[0], workReg[2], workReg[2]) |
| 580 |
VXORPD(workReg[1], workReg[3], workReg[3]) |
| 581 |
|
| 582 |
// Second layer: |
| 583 |
if (skipMask & 2) == 0 { |
| 584 |
leo8MulAdd512(ctx, workReg[0], workReg[1], t01, nil) |
| 585 |
} |
| 586 |
//work1_reg = _mm256_xor_si256(work0_reg, work1_reg); |
| 587 |
VXORPD(workReg[1], workReg[0], workReg[1]) |
| 588 |
|
| 589 |
if (skipMask & 4) == 0 { |
| 590 |
leo8MulAdd512(ctx, workReg[2], workReg[3], t23, nil) |
| 591 |
} |
| 592 |
// work3_reg = _mm256_xor_si256(work2_reg, work3_reg); |
| 593 |
VXORPD(workReg[2], workReg[3], workReg[3]) |
| 594 |
|
| 595 |
// Store + Next loop: |
| 596 |
for i := range work { |
| 597 |
if withDst { |
| 598 |
VMOVDQU64(workReg[i], Mem{Base: dst[i], Disp: 0}) |
| 599 |
ADDQ(U8(64), dst[i]) |
| 600 |
} else { |
| 601 |
VMOVDQU64(workReg[i], Mem{Base: work[i], Disp: 0}) |
| 602 |
} |
| 603 |
ADDQ(U8(64), work[i]) |
| 604 |
} |
| 605 |
|
| 606 |
SUBQ(U8(64), bytes) |
| 607 |
JA(LabelRef("loop")) |
| 608 |
|
| 609 |
VZEROUPPER() |
| 610 |
RET() |
| 611 |
} |
| 612 |
} //end with_dst |
| 613 |
} |
| 614 |
|
| 615 |
} |
| 616 |
|
| 617 |
// x updated, y preserved... |
| 618 |
func leo8MulAdd256(ctx gf8ctx, x, y reg.VecVirtual, table table256) { |
| 619 |
Comment("LEO_MULADD_256") |
| 620 |
lo, hi := YMM(), YMM() |
| 621 |
|
| 622 |
VPAND(y, ctx.clrMask, lo) |
| 623 |
VPSRLQ(U8(4), y, hi) |
| 624 |
VPSHUFB(lo, table.Lo, lo) |
| 625 |
|
| 626 |
// Do high |
| 627 |
VPAND(hi, ctx.clrMask, hi) |
| 628 |
VPSHUFB(hi, table.Hi, hi) |
| 629 |
VPXOR3way(lo, hi, x) |
| 630 |
} |
| 631 |
|
| 632 |
// multiply y with table and xor result into x. |
| 633 |
// If z is set it is xored into x as. |
| 634 |
func leo8MulAdd512(ctx gf8ctx, x, y reg.VecVirtual, table table512, z reg.VecVirtual) { |
| 635 |
Comment("LEO_MULADD_512") |
| 636 |
tmp := ZMM() |
| 637 |
VGF2P8AFFINEQB(U8(0), table, y, tmp) |
| 638 |
if z == nil { |
| 639 |
VXORPD(x, tmp, x) |
| 640 |
} else { |
| 641 |
VPTERNLOGD(U8(0x96), tmp, z, x) |
| 642 |
} |
| 643 |
} |