| OLD | NEW |
|---|
| (Empty) | |
| 1 commit df426dcc3179a3647a695c0fde66e1b3616fa6c0 |
| 2 Author: John Alayari <jalayari@codeaurora.org> |
| 3 Date: Thu Nov 18 16:24:17 2010 -0800 |
| 4 |
| 5 commiting the PNG and ZLIB Neon Optimizations files to local git repsitory. |
| 6 |
| 7 diff --git a/Makefile.in b/Makefile.in |
| 8 index 7da5a85..276c531 100644 |
| 9 --- a/Makefile.in |
| 10 +++ b/Makefile.in |
| 11 @@ -51,7 +51,7 @@ OBJS = adler32.o compress.o crc32.o gzio.o uncompr.o deflate.o
trees.o \ |
| 12 |
| 13 PIC_OBJS = $(OBJS:%.o=%.lo) |
| 14 |
| 15 -OBJA = |
| 16 +OBJA =inflate_fast_copy_neon.o adler32_DO16_loop_neon.o |
| 17 # to use the asm code: make OBJA=match.o |
| 18 |
| 19 TEST_OBJS = example.o minigzip.o |
| 20 @@ -82,8 +82,12 @@ match.o: match.S |
| 21 %.lo: %.c |
| 22 $(CC) $(CFLAGS) -DPIC -fPIC -c $< -o $@ |
| 23 |
| 24 -$(SHAREDLIBV): $(PIC_OBJS) |
| 25 - $(LDSHARED) -o $@ $(PIC_OBJS) -lc $(LDFLAGS) |
| 26 +%.o: %.S |
| 27 + $(CC) $(CFLAGS) -DPIC -fPIC -c $< -o $@ |
| 28 + |
| 29 + |
| 30 +$(SHAREDLIBV): $(PIC_OBJS) $(OBJA) |
| 31 + $(LDSHARED) -o $@ $(PIC_OBJS) $(OBJA) -lc $(LDFLAGS) |
| 32 rm -f $(SHAREDLIB) $(SHAREDLIBM) |
| 33 ln -s $@ $(SHAREDLIB) |
| 34 ln -s $@ $(SHAREDLIBM) |
| 35 diff --git a/adler32.c b/adler32.c |
| 36 index 007ba26..a256e88 100644 |
| 37 --- a/adler32.c |
| 38 +++ b/adler32.c |
| 39 @@ -1,5 +1,6 @@ |
| 40 /* adler32.c -- compute the Adler-32 checksum of a data stream |
| 41 * Copyright (C) 1995-2004 Mark Adler |
| 42 + * Copyright (c) 2010, Code Aurora Forum. All rights reserved. |
| 43 * For conditions of distribution and use, see copyright notice in zlib.h |
| 44 */ |
| 45 |
| 46 @@ -18,6 +19,10 @@ |
| 47 #define DO8(buf,i) DO4(buf,i); DO4(buf,i+4); |
| 48 #define DO16(buf) DO8(buf,0); DO8(buf,8); |
| 49 |
| 50 +#if defined(__ARM_NEON__) |
| 51 +extern void adler32_DO16_loop_neon(unsigned char** , unsigned long *, unsigned
long * , int ); |
| 52 +#endif |
| 53 + |
| 54 /* use NO_DIVIDE if your processor does not do division in hardware */ |
| 55 #ifdef NO_DIVIDE |
| 56 # define MOD(a) \ |
| 57 @@ -96,17 +101,25 @@ uLong ZEXPORT adler32(adler, buf, len) |
| 58 /* do length NMAX blocks -- requires just one modulo operation */ |
| 59 while (len >= NMAX) { |
| 60 len -= NMAX; |
| 61 +#if defined(__ARM_NEON__) |
| 62 + adler32_DO16_loop_neon(&buf, &adler, &sum2, NMAX); |
| 63 +#else |
| 64 n = NMAX / 16; /* NMAX is divisible by 16 */ |
| 65 do { |
| 66 DO16(buf); /* 16 sums unrolled */ |
| 67 buf += 16; |
| 68 } while (--n); |
| 69 +#endif |
| 70 MOD(adler); |
| 71 MOD(sum2); |
| 72 } |
| 73 |
| 74 /* do remaining bytes (less than NMAX, still just one modulo) */ |
| 75 if (len) { /* avoid modulos if none remaining */ |
| 76 + |
| 77 +#if defined(__ARM_NEON__) |
| 78 + adler32_DO16_loop_neon(&buf, &adler, &sum2, len); |
| 79 +#else |
| 80 while (len >= 16) { |
| 81 len -= 16; |
| 82 DO16(buf); |
| 83 @@ -116,6 +129,7 @@ uLong ZEXPORT adler32(adler, buf, len) |
| 84 adler += *buf++; |
| 85 sum2 += adler; |
| 86 } |
| 87 +#endif |
| 88 MOD(adler); |
| 89 MOD(sum2); |
| 90 } |
| 91 diff --git a/adler32_DO16_loop_neon.S b/adler32_DO16_loop_neon.S |
| 92 new file mode 100755 |
| 93 index 0000000..1ba5147 |
| 94 --- /dev/null |
| 95 +++ b/adler32_DO16_loop_neon.S |
| 96 @@ -0,0 +1,195 @@ |
| 97 +# |
| 98 +# Copyright (c) 2010, Code Aurora Forum. All rights reserved. |
| 99 +# |
| 100 +# Redistribution and use in source and binary forms, with or without |
| 101 +# modification, are permitted provided that the following conditions |
| 102 +# are met: |
| 103 +# * Redistributions of source code must retain the above copyright |
| 104 +# notice, this list of conditions and the following disclaimer. |
| 105 +# * Redistributions in binary form must reproduce the above |
| 106 +# copyright notice, this list of conditions and the following |
| 107 +# disclaimer in the documentation and/or other materials provided |
| 108 +# with the distribution. |
| 109 +# * Neither the name of Code Aurora Forum, Inc. nor the names of it |
| 110 +# contributors may be used to endorse or promote products derived |
| 111 +# from this software without specific prior written permission. |
| 112 +# |
| 113 +# THIS SOFTWARE IS PROVIDED "AS IS" AND ANY EXPRESS OR IMPLIED |
| 114 +# WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF |
| 115 +# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMEN |
| 116 +# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTOR |
| 117 +# BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, O |
| 118 +# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 119 +# SUBSTITUTE GOODS OR SERVICES LOSS OF USE, DATA, OR PROFITS OR |
| 120 +# BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, |
| 121 +# WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE |
| 122 +# OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVE |
| 123 +# IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 124 +# |
| 125 +#if defined(__ARM_NEON__) |
| 126 +#====================================================================== |
| 127 + |
| 128 +#====================================================================== |
| 129 +# Code Section |
| 130 + |
| 131 + .code 32 @ Code is ARM ISA |
| 132 + |
| 133 + .global adler32_DO16_loop_neon |
| 134 + |
| 135 +#====================================================================== |
| 136 +# Function Name(s) : adler32_DO16_loop_neon |
| 137 +# |
| 138 +# Function Parameters |
| 139 +# r0 = pointer to buf |
| 140 +# r1 = pointer to adler |
| 141 +# r2 = pointer to sum |
| 142 +# r3 = len |
| 143 +# |
| 144 +# Register Usage |
| 145 +# q0, q1 = input data |
| 146 +# d18,d24 = adler |
| 147 +# d16,d28 = sum |
| 148 +# d30 = multiplication factor |
| 149 +# |
| 150 +#====================================================================== |
| 151 +# |
| 152 +# algorithm: |
| 153 +# |
| 154 +# while (length < loop_counter) |
| 155 +# do |
| 156 +# length = length - loop_counter |
| 157 +# for i = 0 to loop_counter |
| 158 +# adler = adler_begin + input[i] |
| 159 +# for i = 0 to loop_counter |
| 160 +# sum2 = sum2_begin+(loop_counter-i)*input[i] |
| 161 +# sum2 = sum2+(adler_begin*loop_counter) |
| 162 +# adler_begin = adler |
| 163 +# sum2_begin = sum2 |
| 164 +# end |
| 165 +# end |
| 166 +# end |
| 167 +# |
| 168 +# Here loop counter holds values of 16, 8 and 1 to compute |
| 169 +# adler and sum for 16 bytes, 8 bytes and 1 byte at a time |
| 170 +# adler_begin and sum2_begin are used to hold the values |
| 171 +# of adler and sum2 from previous iterations. |
| 172 +# |
| 173 +#====================================================================== |
| 174 +.balign 32 |
| 175 +.type adler32_DO16_loop_neon, %function |
| 176 + |
| 177 +adler32_DO16_loop_neon: |
| 178 + stmdb sp!,{r4-r7} |
| 179 + ldr r7,[r0] |
| 180 + vld1.32 {d18[0]},[r1] @load the input adler |
| 181 + vld1.32 {d16[0]},[r2] @load the input sum |
| 182 + ldr r5,=FACTOR16 @load the multiplication |
| 183 + @factors for data elements |
| 184 + vld1.8 {d20,d21},[r5] @load the multiplication |
| 185 + @factor for adler. |
| 186 + lsrs r4,r3,#4 @Calculate the number |
| 187 + @16 byte iterations |
| 188 + beq adler32_DO16_loop_16bytes_done |
| 189 + mov r6, #16 |
| 190 + vmov.32 d30[0],r6 |
| 191 + |
| 192 +adler32_DO16_loop_16bytes: |
| 193 + vld1.8 {d0,d1},[r7]! @load buf[0]..buf[15] |
| 194 + vpaddl.u8 d6, d0 @pair wise add to reduce |
| 195 + @8 elements to 4 and extend. |
| 196 + vpaddl.u8 d7,d1 @pair wise add to reduce 8 |
| 197 + @elements to 4 and extend. |
| 198 + vpadd.u16 d24,d6,d7 @pair wise add (i.e. no |
| 199 + @need to extend 16 bits |
| 200 + @sufficient to hold the sum). |
| 201 + vpaddl.u16 d24,d24 @pair wise add to reduce 4 |
| 202 + @elements to 2 and extend. |
| 203 + vpaddl.u32 d24,d24 @pair wise add to get the |
| 204 + @adler of 16 inputs no need |
| 205 + @to extend .. but only vpaddl |
| 206 + @adds pair wise on one |
| 207 + @doubleword. |
| 208 + vadd.u32 d24,d18,d24 @adler'=adler+adler_of_16_inputs |
| 209 + vmull.u8 q13,d20,d0 @sum'=mul_fac_for_inputs[0...7] |
| 210 + @ * buf[0..7]. |
| 211 + vmlal.u8 q13,d21,d1 @sum'=sum'+ mul_fac_for_inputs |
| 212 + @[8...15] * buf[8..15]. |
| 213 + vpadd.u16 d28,d26,d27 @pair wise add the doublewords |
| 214 + vpaddl.u16 d28,d28 @pair wise add to reduce 4 |
| 215 + @elements to 2 and extend. |
| 216 + vpaddl.u32 d28,d28 @pair wise add |
| 217 + vadd.u32 d28,d16,d28 @sum' = sum + sum' |
| 218 + vmla.u32 d28,d18,d30 @sum' = sum' + (adler* |
| 219 + @mul_fac_for_adler). |
| 220 + vmov.u32 d18,d24 @save adler for next iteration. |
| 221 + vmov.u32 d16,d28 @save sum for next iteration. |
| 222 + sub r3,r3,#16 |
| 223 + subs r4,r4,#1 |
| 224 + bne adler32_DO16_loop_16bytes |
| 225 + |
| 226 +adler32_DO16_loop_16bytes_done: |
| 227 + lsrs r4, r3, #3 @find if there are atleast 8 bytes |
| 228 + beq adler32_DO16_loop_8bytes_done |
| 229 +adler32_DO16_loop_8bytes: |
| 230 + vld1.8 {d0},[r7]! @load buf[0] .buf[7] |
| 231 + vpaddl.u8 d24,d0 @pair wise add to |
| 232 + @reduce 8 elements to 4 |
| 233 + vpaddl.u16 d24,d24 @pair wise add to reduce |
| 234 + @4 elements to 2 |
| 235 + vpaddl.u32 d24,d24 @pair wise add to get the |
| 236 + @adler for 8 inputs |
| 237 + vadd.u32 d24,d18,d24 @adler' = adler + |
| 238 + @adler_for_8_inputs. |
| 239 + vmull.u8 q13,d21,d0 @sum' = mul_fac_for_inputs[0..7] |
| 240 + @ * buf[0..7] |
| 241 + vpadd.u16 d28,d26,d27 @pair wise add to reduce 8 |
| 242 + @elements to 4 |
| 243 + vpaddl.u16 d28,d28 @pair wise add to reduce 4 |
| 244 + @elements to 2. |
| 245 + vpaddl.u32 d28,d28 @pair wise add |
| 246 + vadd.u32 d28,d16,d28 @sum' = sum + sum' |
| 247 + vshl.u32 d30,d18,#3 @adler" = adler * 8 |
| 248 + vadd.u32 d28,d28,d30 @sum' = sum' + adler" |
| 249 + vmov.u32 d18,d24 @save adler for next iteration |
| 250 + vmov.u32 d16,d28 @save sum for next iteration |
| 251 + sub r3,r3,#8 |
| 252 + |
| 253 +adler32_DO16_loop_8bytes_done: |
| 254 + cmp r3, #0 @find if there are remaining bytes |
| 255 + beq DONE @after profiling found that a loop |
| 256 + @to compute 4 or 2 bytes at a time |
| 257 + @is less efficient than a byte by |
| 258 + @byte computation. |
| 259 + vmov.u64 d3, #0 |
| 260 + |
| 261 +adler32_DO16_loop_remaining: |
| 262 + vld1.8 {d3[0]}, [r7]! @load 1 byte of input |
| 263 + subs r3,r3,#1 |
| 264 + vadd.u32 d24,d3,d18 @adler' = adler + *buf |
| 265 + vadd.u32 d28,d24,d16 @sum' = sum + adler' |
| 266 + vmov.u32 d18,d24 |
| 267 + vmov.u32 d16,d28 |
| 268 + bne adler32_DO16_loop_remaining |
| 269 + |
| 270 +DONE: |
| 271 + vst1.32 {d24[0]},[r1] |
| 272 + vst1.32 {d28[0]},[r2] |
| 273 + str r7, [r0] |
| 274 + ldmia sp!, {r4-r7} |
| 275 + bx lr |
| 276 + |
| 277 +.size adler32_DO16_loop_neon, .-adler32_DO16_loop_neon |
| 278 + |
| 279 +.balign 16 |
| 280 +#====================================================================== |
| 281 +#FACTOR16 provides the multiplication factors for the inputs for 16 |
| 282 +#byte loops. The second half (i.e. 8 to 1) has the multiplications |
| 283 +#factors for 8 byte loops. |
| 284 +#====================================================================== |
| 285 + |
| 286 +FACTOR16: |
| 287 + .byte 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1 |
| 288 + |
| 289 +#endif |
| 290 + .END |
| 291 + |
| 292 diff --git a/inffast.c b/inffast.c |
| 293 index bbee92e..943cb92 100644 |
| 294 --- a/inffast.c |
| 295 +++ b/inffast.c |
| 296 @@ -8,6 +8,10 @@ |
| 297 #include "inflate.h" |
| 298 #include "inffast.h" |
| 299 |
| 300 +#if defined(__ARM_NEON__) |
| 301 +extern void inflate_fast_copy_neon(unsigned len, unsigned char **out, unsigned
char *from); |
| 302 +#endif |
| 303 + |
| 304 #ifndef ASMINF |
| 305 |
| 306 /* Allow machine dependent optimization for post-increment or pre-increment. |
| 307 @@ -231,6 +235,9 @@ unsigned start; /* inflate()'s starting value for st
rm->avail_out */ |
| 308 from = out - dist; /* rest from output */ |
| 309 } |
| 310 } |
| 311 +#if defined(__ARM_NEON__) |
| 312 + inflate_fast_copy_neon(len, &out, from); |
| 313 +#else |
| 314 while (len > 2) { |
| 315 PUP(out) = PUP(from); |
| 316 PUP(out) = PUP(from); |
| 317 @@ -242,9 +249,13 @@ unsigned start; /* inflate()'s starting value for s
trm->avail_out */ |
| 318 if (len > 1) |
| 319 PUP(out) = PUP(from); |
| 320 } |
| 321 +#endif |
| 322 } |
| 323 else { |
| 324 from = out - dist; /* copy direct from output */ |
| 325 +#if defined(__ARM_NEON__) |
| 326 + inflate_fast_copy_neon(len, &out, from); |
| 327 +#else |
| 328 do { /* minimum length is three */ |
| 329 PUP(out) = PUP(from); |
| 330 PUP(out) = PUP(from); |
| 331 @@ -256,6 +267,7 @@ unsigned start; /* inflate()'s starting value for st
rm->avail_out */ |
| 332 if (len > 1) |
| 333 PUP(out) = PUP(from); |
| 334 } |
| 335 +#endif |
| 336 } |
| 337 } |
| 338 else if ((op & 64) == 0) { /* 2nd level distance code */ |
| 339 diff --git a/inflate_fast_copy_neon.S b/inflate_fast_copy_neon.S |
| 340 new file mode 100755 |
| 341 index 0000000..ec1e4ab |
| 342 --- /dev/null |
| 343 +++ b/inflate_fast_copy_neon.S |
| 344 @@ -0,0 +1,521 @@ |
| 345 +#; Copyright (c) 2010, Code Aurora Forum. All rights reserved. |
| 346 +#; |
| 347 +#; Redistribution and use in source and binary forms, with or without |
| 348 +#; modification, are permitted provided that the following conditions are |
| 349 +#; met: |
| 350 +#; * Redistributions of source code must retain the above copyright |
| 351 +#; notice, this list of conditions and the following disclaimer. |
| 352 +#; * Redistributions in binary form must reproduce the above |
| 353 +#; copyright notice, this list of conditions and the following |
| 354 +#; disclaimer in the documentation and/or other materials provided |
| 355 +#; with the distribution. |
| 356 +#; * Neither the name of Code Aurora Forum, Inc. nor the names of its |
| 357 +#; contributors may be used to endorse or promote products derived |
| 358 +#; from this software without specific prior written permission. |
| 359 +#; |
| 360 +#; THIS SOFTWARE IS PROVIDED "AS IS" AND ANY EXPRESS OR IMPLIED |
| 361 +#; WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF |
| 362 +#; MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT |
| 363 +#; ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS |
| 364 +#; BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 365 +#; CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 366 +#; SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR |
| 367 +#; BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, |
| 368 +#; WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE |
| 369 +#; OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN |
| 370 +#; IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 371 +#; |
| 372 +#; |
| 373 +#if defined(__ARM_NEON__) |
| 374 +#;============================================================================ |
| 375 +#; Code Section |
| 376 + .code 32 @; Code is ARM ISA |
| 377 +#;============================================================================ |
| 378 + |
| 379 + .global inflate_fast_copy_neon |
| 380 + |
| 381 + |
| 382 +#;============================================================================ |
| 383 +#; INPUTS: r0 len: number of bytes to transfer |
| 384 +#; r1 **out: pointer to pointer to ``out'' buffer |
| 385 +#; r2 *from: pointer to ``from'' buffer |
| 386 +#; OUTPUTS: r1 **out: pointer to pointer to ``out'' buffer |
| 387 +#;============================================================================ |
| 388 +.balign 32 |
| 389 +.type inflate_fast_copy_neon, %function |
| 390 +inflate_fast_copy_neon: |
| 391 + push {r4-r11} @; push r4-r11 onto stack |
| 392 + |
| 393 + cmp r0,#16 @; |
| 394 + bge inflate_fast_copy_vectorized |
| 395 + |
| 396 + #;; transfer bytes one by one |
| 397 + #;; only if len < 16 bytes |
| 398 +inflate_fast_copy_default: |
| 399 + |
| 400 + cmp r0,#0 |
| 401 + beq inflate_fast_copy_exit |
| 402 + |
| 403 + ldr r3,[r1,#0] @; r3 = pointer to out |
| 404 + |
| 405 +inflate_fast_copy_default_loop: |
| 406 + |
| 407 + ldrb r12,[r2,#1]! @; r12 = *(++from) |
| 408 + subs r0,r0,#1 @; len-- |
| 409 + strb r12,[r3,#1]! @; *(++out) = r12 |
| 410 + |
| 411 + bne inflate_fast_copy_default_loop |
| 412 + |
| 413 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 414 + @; to out |
| 415 + b inflate_fast_copy_exit |
| 416 + |
| 417 + #;; vectorized copy routines |
| 418 + #;; only if len > 16 bytes |
| 419 +inflate_fast_copy_vectorized: |
| 420 + |
| 421 + ldr r3,[r1,#0] @; r3 = pointer to out |
| 422 + @; DON'T TOUCH r1 UNTIL FINAL |
| 423 + @; UPDATE OF r1 WITH ADDRESS OF r3 |
| 424 + cmp r3,r2 @ |
| 425 + sublt r4,r2,r3 @ |
| 426 + subge r4,r3,r2 @;r4 = gap = |out-from| |
| 427 + |
| 428 + cmp r4,#0 |
| 429 + beq inflate_fast_copy_exit |
| 430 + |
| 431 + cmp r4,#1 |
| 432 + beq inflate_fast_copy_gap1b_proc |
| 433 + |
| 434 + cmp r4,#2 |
| 435 + beq inflate_fast_copy_gap2b_proc |
| 436 + |
| 437 + cmp r4,#3 |
| 438 + beq inflate_fast_copy_gap3b_proc |
| 439 + |
| 440 + cmp r4,#4 |
| 441 + beq inflate_fast_copy_gap4b_proc |
| 442 + |
| 443 + cmp r4,#8 |
| 444 + blt inflate_fast_copy_gap5to7b_proc |
| 445 + beq inflate_fast_copy_gap8b_proc |
| 446 + |
| 447 + cmp r4,#16 |
| 448 + blt inflate_fast_copy_gap9to15b_proc |
| 449 + bge inflate_fast_copy_gap16b_proc |
| 450 + |
| 451 + |
| 452 + #;; ------------------------------------------------------------------ |
| 453 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 454 + #;; buffers is 1 byte |
| 455 + #;; INPUTS: |
| 456 + #;; r0 = len |
| 457 + #;; r2 = pointer to from |
| 458 + #;; r3 = pointer to out |
| 459 + #;; OUTPUTS: |
| 460 + #;; r1 = pointer to pointer to out |
| 461 + #;; ------------------------------------------------------------------ |
| 462 +inflate_fast_copy_gap1b_proc: |
| 463 + |
| 464 + add r3,r3,#1 @; out++ |
| 465 + @ |
| 466 + ldrb r12,[r2,#1]! @; r12 = *(++from) |
| 467 + vdup.8 q0, r12 @; duplicate r12 16 times in q0 |
| 468 + @ |
| 469 + lsrs r4,r0,#4 @; r4 = floor(len/16) |
| 470 + @; = iteration count for loop16 |
| 471 + beq inflate_fast_copy_gap1b_proc_16bytes_loop_done |
| 472 + |
| 473 +inflate_fast_copy_gap1b_proc_16bytes_loop: |
| 474 + |
| 475 + vst1.8 {q0},[r3]! @; store 16 bytes in out and |
| 476 + @; increment out pointer |
| 477 + sub r0,r0,#16 @; subtract 16 from len |
| 478 + subs r4,r4,#1 @; decrement iteration count |
| 479 + bne inflate_fast_copy_gap1b_proc_16bytes_loop |
| 480 + |
| 481 +inflate_fast_copy_gap1b_proc_16bytes_loop_done: |
| 482 + |
| 483 + cmp r0,#0 |
| 484 + subeq r3,r3,#1 @; out-- |
| 485 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 486 + @; to out |
| 487 + beq inflate_fast_copy_exit |
| 488 + |
| 489 +inflate_fast_copy_gap1b_proc_lastfewbytes_loop: |
| 490 + |
| 491 + strb r12,[r3],#1 @; *out = r12, out++ |
| 492 + subs r0,r0,#1 @; len-- |
| 493 + bne inflate_fast_copy_gap1b_proc_lastfewbytes_loop |
| 494 + |
| 495 + sub r3,r3,#1 @; out-- |
| 496 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 497 + @; to out |
| 498 + b inflate_fast_copy_exit |
| 499 + |
| 500 + #;; ------------------------------------------------------------------ |
| 501 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 502 + #;; buffers is 2 bytes |
| 503 + #;; INPUTS: |
| 504 + #;; r0 = len |
| 505 + #;; r2 = pointer to from |
| 506 + #;; r3 = pointer to out |
| 507 + #;; OUTPUTS: |
| 508 + #;; r1 = pointer to pointer to out |
| 509 + #;; ------------------------------------------------------------------ |
| 510 +inflate_fast_copy_gap2b_proc: |
| 511 + |
| 512 + add r2,r2,#1 @; from++ |
| 513 + add r3,r3,#1 @; out++ |
| 514 + @ |
| 515 + vld1.16 {d0[0]},[r2] @; load 2 bytes into d0[0] |
| 516 + vdup.16 q0,d0[0] @; duplicate those 2 bytes 8 times |
| 517 + @; to fill up q0 |
| 518 + @ |
| 519 + lsrs r4,r0,#4 @; r4 = floor(len/16) |
| 520 + @; = iteration count for loop16 |
| 521 + beq inflate_fast_copy_gap2b_proc_16bytes_loop_done |
| 522 + |
| 523 +inflate_fast_copy_gap2b_proc_16bytes_loop: |
| 524 + |
| 525 + vst1.8 {q0},[r3]! @; store 16 bytes in out and |
| 526 + @; increment out pointer |
| 527 + sub r0,r0,#16 @; subtract 16 from len |
| 528 + subs r4,r4,#1 @; decrement iteration count |
| 529 + bne inflate_fast_copy_gap2b_proc_16bytes_loop |
| 530 + |
| 531 +inflate_fast_copy_gap2b_proc_16bytes_loop_done: |
| 532 + |
| 533 + cmp r0,#0 |
| 534 + subeq r3,r3,#1 @; out-- |
| 535 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 536 + @; to out |
| 537 + beq inflate_fast_copy_exit |
| 538 + |
| 539 +inflate_fast_copy_gap2b_proc_lastfewbytes_loop: |
| 540 + |
| 541 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 542 + subs r0,r0,#1 @; len-- |
| 543 + strb r12,[r3],#1 @; *out = r12, out++ |
| 544 + @ |
| 545 + bne inflate_fast_copy_gap2b_proc_lastfewbytes_loop |
| 546 + |
| 547 + sub r3,r3,#1 @; out-- |
| 548 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 549 + @; to out |
| 550 + b inflate_fast_copy_exit |
| 551 + |
| 552 + #;; ------------------------------------------------------------------ |
| 553 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 554 + #;; buffers is 3 bytes |
| 555 + #;; INPUTS: |
| 556 + #;; r0 = len |
| 557 + #;; r2 = pointer to from |
| 558 + #;; r3 = pointer to out |
| 559 + #;; r4 = 3 |
| 560 + #;; OUTPUTS: |
| 561 + #;; r1 = pointer to pointer to out |
| 562 + #;; ------------------------------------------------------------------ |
| 563 +inflate_fast_copy_gap3b_proc: |
| 564 + |
| 565 + add r2,r2,#1 @; from++ |
| 566 + add r3,r3,#1 @; out++ |
| 567 + @ |
| 568 + vld1.32 {d0[0]},[r2] @; load 4 bytes into d0[0] |
| 569 + |
| 570 +inflate_fast_copy_gap3b_proc_3bytes_loop: |
| 571 + |
| 572 + cmp r0,#3 @; exit loop if len < 3 |
| 573 + blt inflate_fast_copy_gap3b_proc_3bytes_loop_done |
| 574 + |
| 575 + vst1.32 {d0[0]},[r3],r4 @; store 4 bytes in out |
| 576 + @; out+=3 |
| 577 + |
| 578 + sub r0,r0,#3 @; len-=3 |
| 579 + b inflate_fast_copy_gap3b_proc_3bytes_loop |
| 580 + |
| 581 +inflate_fast_copy_gap3b_proc_3bytes_loop_done: |
| 582 + |
| 583 + cmp r0,#0 |
| 584 + subeq r3,r3,#1 @; out-- |
| 585 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 586 + @; to out |
| 587 + beq inflate_fast_copy_exit |
| 588 + |
| 589 +inflate_fast_copy_gap3b_proc_lastfewbytes_loop: |
| 590 + |
| 591 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 592 + subs r0,r0,#1 @; len-- |
| 593 + strb r12,[r3],#1 @; *out = r12, out++ |
| 594 + |
| 595 + bne inflate_fast_copy_gap3b_proc_lastfewbytes_loop |
| 596 + |
| 597 + sub r3,r3,#1 @; out-- |
| 598 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 599 + @; to out |
| 600 + b inflate_fast_copy_exit |
| 601 + |
| 602 + #;; ------------------------------------------------------------------ |
| 603 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 604 + #;; buffers is 4 bytes |
| 605 + #;; INPUTS: |
| 606 + #;; r0 = len |
| 607 + #;; r2 = pointer to from |
| 608 + #;; r3 = pointer to out |
| 609 + #;; OUTPUTS: |
| 610 + #;; r1 = pointer to pointer to out |
| 611 + #;; ------------------------------------------------------------------ |
| 612 +inflate_fast_copy_gap4b_proc: |
| 613 + |
| 614 + add r2,r2,#1 @; from++ |
| 615 + add r3,r3,#1 @; out++ |
| 616 + @ |
| 617 + vld1.32 {d0[0]},[r2] @; load 4 bytes into d0[0] |
| 618 + vdup.32 q0,d0[0] @; duplicate those 4 bytes 4 times |
| 619 + @; to fill up q0 |
| 620 + @ |
| 621 + lsrs r4,r0,#4 @; r4 = floor(len/16) |
| 622 + @; = iteration count for loop16 |
| 623 + beq inflate_fast_copy_gap4b_proc_16bytes_loop_done |
| 624 + |
| 625 +inflate_fast_copy_gap4b_proc_16bytes_loop: |
| 626 + |
| 627 + vst1.32 {q0},[r3]! @; store 16 bytes in out and |
| 628 + @; increment out pointer |
| 629 + sub r0,r0,#16 @; subtract 16 from len |
| 630 + subs r4,r4,#1 @; decrement iteration count |
| 631 + bne inflate_fast_copy_gap4b_proc_16bytes_loop |
| 632 + |
| 633 +inflate_fast_copy_gap4b_proc_16bytes_loop_done: |
| 634 + |
| 635 + cmp r0,#0 |
| 636 + subeq r3,r3,#1 @; out-- |
| 637 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 638 + @; to out |
| 639 + beq inflate_fast_copy_exit |
| 640 + |
| 641 +inflate_fast_copy_gap4b_proc_lastfewbytes_loop: |
| 642 + |
| 643 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 644 + subs r0,r0,#1 @; len-- |
| 645 + strb r12,[r3],#1 @; *out = r12, out++ |
| 646 + |
| 647 + bne inflate_fast_copy_gap4b_proc_lastfewbytes_loop |
| 648 + |
| 649 + sub r3,r3,#1 @; out-- |
| 650 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 651 + @; to out |
| 652 + b inflate_fast_copy_exit |
| 653 + |
| 654 + #;; ------------------------------------------------------------------ |
| 655 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 656 + #;; buffers is {5-7} bytes |
| 657 + #;; INPUTS: |
| 658 + #;; r0 = len |
| 659 + #;; r2 = pointer to from |
| 660 + #;; r3 = pointer to out |
| 661 + #;; r4 = {5-7} |
| 662 + #;; OUTPUTS: |
| 663 + #;; r1 = pointer to pointer to out |
| 664 + #;; ------------------------------------------------------------------ |
| 665 +inflate_fast_copy_gap5to7b_proc: |
| 666 + |
| 667 + add r2,r2,#1 @; from++ |
| 668 + add r3,r3,#1 @; out++ |
| 669 + @ |
| 670 + vld1.8 {d0},[r2] @; load 8 bytes into d0 |
| 671 + |
| 672 +inflate_fast_copy_gap5to7b_proc_5to7bytes_loop: |
| 673 + |
| 674 + cmp r0,r4 @; exit loop if len < {5-7} |
| 675 + blt inflate_fast_copy_gap5to7b_proc_5to7bytes_loop_done |
| 676 + |
| 677 + vst1.8 {d0},[r3],r4 @; store 8 bytes in out |
| 678 + @; out+={5-7} |
| 679 + |
| 680 + sub r0,r0,r4 @; len-={5-7} |
| 681 + b inflate_fast_copy_gap5to7b_proc_5to7bytes_loop |
| 682 + |
| 683 +inflate_fast_copy_gap5to7b_proc_5to7bytes_loop_done: |
| 684 + |
| 685 + cmp r0,#0 |
| 686 + subeq r3,r3,#1 @; out-- |
| 687 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 688 + @; to out |
| 689 + beq inflate_fast_copy_exit |
| 690 + |
| 691 +inflate_fast_copy_gap5to7b_proc_lastfewbytes_loop: |
| 692 + |
| 693 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 694 + subs r0,r0,#1 @; len-- |
| 695 + strb r12,[r3],#1 @; *out = r12, out++ |
| 696 + |
| 697 + bne inflate_fast_copy_gap5to7b_proc_lastfewbytes_loop |
| 698 + |
| 699 + sub r3,r3,#1 @; out-- |
| 700 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 701 + @; to out |
| 702 + b inflate_fast_copy_exit |
| 703 + |
| 704 + #;; ------------------------------------------------------------------ |
| 705 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 706 + #;; buffers is 8 bytes |
| 707 + #;; INPUTS: |
| 708 + #;; r0 = len |
| 709 + #;; r2 = pointer to from |
| 710 + #;; r3 = pointer to out |
| 711 + #;; OUTPUTS: |
| 712 + #;; r1 = pointer to pointer to out |
| 713 + #;; ------------------------------------------------------------------ |
| 714 +inflate_fast_copy_gap8b_proc: |
| 715 + |
| 716 + add r2,r2,#1 @; from++ |
| 717 + add r3,r3,#1 @; out++ |
| 718 + @ |
| 719 + vld1.8 {d0},[r2] @; load 8 bytes into d0 |
| 720 + vmov d1,d0 @; duplicate the 8 bytes to fill up |
| 721 + @; q0 |
| 722 + @ |
| 723 + lsrs r4,r0,#4 @; r4 = floor(len/16) |
| 724 + @; = iteration count for loop16 |
| 725 + beq inflate_fast_copy_gap8b_proc_16bytes_loop_done |
| 726 + |
| 727 +inflate_fast_copy_gap8b_proc_16bytes_loop: |
| 728 + |
| 729 + vst1.8 {q0},[r3]! @; store 16 bytes in out and |
| 730 + @; increment out pointer |
| 731 + sub r0,r0,#16 @; subtract 16 from len |
| 732 + subs r4,r4,#1 @; decrement iteration count |
| 733 + bne inflate_fast_copy_gap8b_proc_16bytes_loop |
| 734 + |
| 735 +inflate_fast_copy_gap8b_proc_16bytes_loop_done: |
| 736 + |
| 737 + cmp r0,#0 |
| 738 + subeq r3,r3,#1 @; out-- |
| 739 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 740 + @; to out |
| 741 + beq inflate_fast_copy_exit |
| 742 + |
| 743 +inflate_fast_copy_gap8b_proc_lastfewbytes_loop: |
| 744 + |
| 745 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 746 + subs r0,r0,#1 @; len-- |
| 747 + strb r12,[r3],#1 @; *out = r12, out++ |
| 748 + |
| 749 + bne inflate_fast_copy_gap8b_proc_lastfewbytes_loop |
| 750 + |
| 751 + sub r3,r3,#1 @; out-- |
| 752 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 753 + @; to out |
| 754 + b inflate_fast_copy_exit |
| 755 + |
| 756 + #;; ------------------------------------------------------------------ |
| 757 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 758 + #;; buffers is {9-15} bytes |
| 759 + #;; INPUTS: |
| 760 + #;; r0 = len |
| 761 + #;; r2 = pointer to from |
| 762 + #;; r3 = pointer to out |
| 763 + #;; r4 = {9-15} |
| 764 + #;; OUTPUTS: |
| 765 + #;; r1 = pointer to pointer to out |
| 766 + #;; ------------------------------------------------------------------ |
| 767 +inflate_fast_copy_gap9to15b_proc: |
| 768 + |
| 769 + add r2,r2,#1 @; from++ |
| 770 + add r3,r3,#1 @; out++ |
| 771 + @ |
| 772 + vld1.8 {q0},[r2] @; load 16 bytes into q0 |
| 773 + |
| 774 +inflate_fast_copy_gap9to15b_proc_9to15bytes_loop: |
| 775 + |
| 776 + cmp r0, r4 @; exit loop if len < {9-15} |
| 777 + blt inflate_fast_copy_gap9to15b_proc_9to15bytes_loop_done |
| 778 + |
| 779 + vst1.8 {q0},[r3],r4 @; store 16 bytes in out |
| 780 + @; out+={9-15} |
| 781 + |
| 782 + sub r0,r0,r4 @; len-={9-15} |
| 783 + b inflate_fast_copy_gap9to15b_proc_9to15bytes_loop |
| 784 + |
| 785 +inflate_fast_copy_gap9to15b_proc_9to15bytes_loop_done: |
| 786 + |
| 787 + cmp r0,#0 |
| 788 + subeq r3,r3,#1 @; out-- |
| 789 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 790 + @; to out |
| 791 + beq inflate_fast_copy_exit |
| 792 + |
| 793 +inflate_fast_copy_gap9to15b_proc_lastfewbytes_loop: |
| 794 + |
| 795 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 796 + subs r0,r0,#1 @; len-- |
| 797 + strb r12,[r3],#1 @; *out = r12, out++ |
| 798 + |
| 799 + bne inflate_fast_copy_gap9to15b_proc_lastfewbytes_loop |
| 800 + |
| 801 + sub r3,r3,#1 @; out-- |
| 802 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 803 + @; to out |
| 804 + b inflate_fast_copy_exit |
| 805 + |
| 806 + #;; ------------------------------------------------------------------ |
| 807 + #;; vectorized copy routine when gap between ``from'' and ``out'' |
| 808 + #;; buffers is 16 bytes or more |
| 809 + #;; INPUTS: |
| 810 + #;; r0 = len |
| 811 + #;; r2 = pointer to from |
| 812 + #;; r3 = pointer to out |
| 813 + #;; OUTPUTS: |
| 814 + #;; r1 = pointer to pointer to out |
| 815 + #;; ------------------------------------------------------------------ |
| 816 +inflate_fast_copy_gap16b_proc: |
| 817 + |
| 818 + add r2,r2,#1 @; from++ |
| 819 + add r3,r3,#1 @; out++ |
| 820 + @ |
| 821 + lsrs r4,r0,#4 @; r4 = floor(len/16) |
| 822 + @; = iteration count for loop16 |
| 823 + beq inflate_fast_copy_gap16b_proc_16bytes_loop_done |
| 824 + |
| 825 +inflate_fast_copy_gap16b_proc_16bytes_loop: |
| 826 + |
| 827 + vld1.8 {q0},[r2]! @; load 16 bytes into q0 and |
| 828 + @; increment from pointer |
| 829 + vst1.8 {q0},[r3]! @; store 16 bytes in out and |
| 830 + @; increment out pointer |
| 831 + sub r0,r0,#16 @; subtract 16 from len |
| 832 + subs r4,r4,#1 @; decrement iteration count |
| 833 + bne inflate_fast_copy_gap16b_proc_16bytes_loop |
| 834 + |
| 835 +inflate_fast_copy_gap16b_proc_16bytes_loop_done: |
| 836 + |
| 837 + cmp r0,#0 |
| 838 + subeq r3,r3,#1 @; out-- |
| 839 + streq r3,[r1,#0] @; r1 = updated pointer to pointer |
| 840 + @; to out |
| 841 + beq inflate_fast_copy_exit |
| 842 + |
| 843 +inflate_fast_copy_gap16b_proc_lastfewbytes_loop: |
| 844 + |
| 845 + ldrb r12,[r2],#1 @; r12 = *from, from++ |
| 846 + subs r0,r0,#1 @; len-- |
| 847 + strb r12,[r3],#1 @; *out = r12, out++ |
| 848 + |
| 849 + bne inflate_fast_copy_gap16b_proc_lastfewbytes_loop |
| 850 + |
| 851 + sub r3,r3,#1 @; out-- |
| 852 + str r3,[r1,#0] @; r1 = updated pointer to pointer |
| 853 + @; to out |
| 854 + |
| 855 +inflate_fast_copy_exit: |
| 856 + |
| 857 + pop {r4-r11} @; pop r4-r11 from stack |
| 858 + bx lr @; return |
| 859 + |
| 860 +.size inflate_fast_copy_neon, .-inflate_fast_copy_neon |
| 861 + |
| 862 +#endif |
| 863 + .END |
| 864 + |
| 865 + |
| OLD | NEW |
|---|
Chromium Code Reviews
Issue 5176006:
Applying Neon optimization patch to the ZLIB library.
Base URL: http://git.chromium.org/git/portage.git@master