DXR is a code search and navigation tool aimed at making sense of large projects. It supports full-text and regex searches as well as structural queries.

Mercurial (d8847129d134)

VCS Links

Line Code
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
/*
 * Copyright 2014 The Android Open Source Project
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include <SkPreConfig.h>

#if defined(__clang__) || (defined(__GNUC__) && !defined(SK_BUILD_FOR_MAC))

#define CFI_PUSH(REG) \
    .cfi_adjust_cfa_offset 4; \
    .cfi_rel_offset REG, 0

#define CFI_POP(REG) \
    .cfi_adjust_cfa_offset -4; \
    .cfi_restore REG

#define PUSH(REG) pushl REG; CFI_PUSH (REG)
#define POP(REG)  popl REG; CFI_POP (REG)
#define RETURN    POP(%edi); ret

#define EXTRACT_ALPHA(var1, var2) \
    movdqa      %var1, %var2;           /* Clone source pixels to extract alpha */\
    psrlw       $8, %var2;              /* Discard red and blue, leaving alpha and green */\
    pshufhw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (high) */\
    movdqa      %xmm6, %xmm4;           \
    pshuflw     $0xF5, %var2, %var2;    /* Repeat alpha for scaling (low) */\
    movdqa      %xmm5, %xmm3;           \
    psubw       %var2, %xmm4            /* Finalize alpha calculations */

#define SCALE_PIXELS \
    psllw       $8, %xmm5;              /* Filter out red and blue components */\
    pmulhuw     %xmm4, %xmm5;           /* Scale red and blue */\
    psrlw       $8, %xmm3;              /* Filter out alpha and green components */\
    pmullw      %xmm4, %xmm3            /* Scale alpha and green */


/*
 * void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
 *                                 const SkPMColor* SK_RESTRICT src,
 *                                 int count, U8CPU alpha)
 *
 * This function is divided into six blocks: initialization, blit 4-15 pixels,
 * blit 0-3 pixels, align destination for 16+ pixel blits,
 * blit 16+ pixels with source unaligned, blit 16+ pixels with source aligned.
 * There are some code reuse between the blocks.
 *
 * The primary optimization comes from checking the source pixels' alpha value.
 * If the alpha is zero, the pixel can be skipped entirely.
 * If the alpha is fully opaque, the pixel can be copied directly to the destination.
 * According to collected statistics, these two cases are the most common.
 * The main loop(s) uses pre-loading and unrolling in an attempt to reduce the
 * memory latency worse-case.
 */

#ifdef __clang__
    .text
#else
    .section .text.sse4.2,"ax",@progbits
    .type S32A_Opaque_BlitRow32_SSE4_asm, @function
#endif
    .p2align 4
#if defined(SK_BUILD_FOR_MAC)
    .global _S32A_Opaque_BlitRow32_SSE4_asm
    .private_extern _S32A_Opaque_BlitRow32_SSE4_asm
_S32A_Opaque_BlitRow32_SSE4_asm:
#else
    .global S32A_Opaque_BlitRow32_SSE4_asm
    .hidden S32A_Opaque_BlitRow32_SSE4_asm
S32A_Opaque_BlitRow32_SSE4_asm:
#endif
    .cfi_startproc
    movl        8(%esp), %eax           // Source pointer
    movl        12(%esp), %ecx          // Pixel count
    movl        4(%esp), %edx           // Destination pointer
    prefetcht0  (%eax)

    // Setup SSE constants
    pcmpeqd     %xmm7, %xmm7            // 0xFF000000 mask to check alpha
    pslld       $24, %xmm7
    pcmpeqw     %xmm6, %xmm6            // 16-bit 256 to calculate inv. alpha
    psrlw       $15, %xmm6
    psllw       $8, %xmm6
    pcmpeqw     %xmm0, %xmm0            // 0x00FF00FF mask (Must be in xmm0 because of pblendvb)
    psrlw       $8, %xmm0
    subl        $4, %ecx                // Check if we have only 0-3 pixels
    js          .LReallySmall
    PUSH(%edi)
    cmpl        $11, %ecx               // Do we have enough pixels to run the main loop?
    ja          .LBigBlit

    // Handle small blits (4-15 pixels)
    ////////////////////////////////////////////////////////////////////////////////
    xorl        %edi, %edi              // Reset offset to zero

.LSmallLoop:
    lddqu       (%eax, %edi), %xmm1     // Load four source pixels
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LSmallAlphaNotOpaqueOrZero
    jz          .LSmallAlphaZero        // If all alphas are zero, skip the pixels completely
    movdqu      %xmm1, (%edx, %edi)     // Store four destination pixels
.LSmallAlphaZero:
    addl        $16, %edi
    subl        $4, %ecx                // Check if there are four additional pixels, at least
    jns         .LSmallLoop
    jmp         .LSmallRemaining

    // Handle mixed alphas (calculate and scale)
    .p2align 4
.LSmallAlphaNotOpaqueOrZero:
    lddqu       (%edx, %edi), %xmm5     // Load four destination pixels
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    addl        $16, %edi
    subl        $4, %ecx                // Check if there are four additional pixels, at least
    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
    paddb       %xmm3, %xmm1            // Add source and destination pixels together
    movdqu      %xmm1, -16(%edx, %edi)  // Store four destination pixels
    jns         .LSmallLoop

    // Handle the last 0-3 pixels (also used by the main loops)
.LSmallRemaining:
    cmpl        $-4, %ecx               // Check if we are done
    je          .LSmallExit
    sall        $2, %ecx                // Calculate offset for last pixels
    addl        %ecx, %edi

    lddqu       (%eax, %edi), %xmm1     // Load last four source pixels (overlapping)
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    jc          .LSmallRemainingStoreAll// If all alphas are opaque, just store (overlapping)
    jz          .LSmallExit             // If all alphas are zero, skip the pixels completely

    // Handle mixed alphas (calculate and scale)
    lddqu       (%edx, %edi), %xmm5     // Load last four destination pixels (overlapping)
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value

    psllw       $8, %xmm3               // Filter out red and blue components
    pmulhuw     %xmm4, %xmm3            // Scale red and blue
    movdqa      %xmm5, %xmm2
    psrlw       $8, %xmm2               // Filter out alpha and green components
    pmullw      %xmm4, %xmm2            // Scale alpha and green

    cmpl        $-8, %ecx               // Check how many pixels should be written
    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm2, %xmm1            // Add source and destination pixels together
    jb          .LSmallPixelsLeft1
    ja          .LSmallPixelsLeft3      // To avoid double-blending the overlapping pixels...
    pblendw     $0xF0, %xmm1, %xmm5     // Merge only the final two pixels to the destination
    movdqu      %xmm5, (%edx, %edi)     // Store last two destination pixels
.LSmallExit:
    RETURN

.LSmallPixelsLeft1:
    pblendw     $0xC0, %xmm1, %xmm5     // Merge only the final pixel to the destination
    movdqu      %xmm5, (%edx, %edi)     // Store last destination pixel
    RETURN

.LSmallPixelsLeft3:
    pblendw     $0xFC, %xmm1, %xmm5     // Merge only the final three pixels to the destination
    movdqu      %xmm5, (%edx, %edi)     // Store last three destination pixels
    RETURN

.LSmallRemainingStoreAll:
    movdqu      %xmm1, (%edx, %edi)     // Store last destination pixels (overwrite)
    RETURN

    // Handle really small blits (0-3 pixels)
    ////////////////////////////////////////////////////////////////////////////////
.LReallySmall:
    addl        $4, %ecx
    jle         .LReallySmallExit
    pcmpeqd     %xmm1, %xmm1
    cmp         $2, %ecx                // Check how many pixels should be read
    pinsrd      $0x0, (%eax), %xmm1     // Load one source pixel
    pinsrd      $0x0, (%edx), %xmm5     // Load one destination pixel
    jb          .LReallySmallCalc
    pinsrd      $0x1, 4(%eax), %xmm1    // Load second source pixel
    pinsrd      $0x1, 4(%edx), %xmm5    // Load second destination pixel
    je          .LReallySmallCalc
    pinsrd      $0x2, 8(%eax), %xmm1    // Load third source pixel
    pinsrd      $0x2, 8(%edx), %xmm5    // Load third destination pixel

.LReallySmallCalc:
    ptest       %xmm7, %xmm1            // Check if all alphas are opaque
    jc          .LReallySmallStore      // If all alphas are opaque, just store

    // Handle mixed alphas (calculate and scale)
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value

    pand        %xmm0, %xmm5            // Filter out red and blue components
    pmullw      %xmm4, %xmm5            // Scale red and blue
    psrlw       $8, %xmm3               // Filter out alpha and green components
    pmullw      %xmm4, %xmm3            // Scale alpha and green

    psrlw       $8, %xmm5               // Combine results
    pblendvb    %xmm5, %xmm3            // Mask in %xmm0, implicitly
    paddb       %xmm3, %xmm1            // Add source and destination pixels together

.LReallySmallStore:
    cmp         $2, %ecx                // Check how many pixels should be written
    pextrd      $0x0, %xmm1, (%edx)     // Store one destination pixel
    jb          .LReallySmallExit
    pextrd      $0x1, %xmm1, 4(%edx)    // Store second destination pixel
    je          .LReallySmallExit
    pextrd      $0x2, %xmm1, 8(%edx)    // Store third destination pixel
.LReallySmallExit:
    ret

    // Handle bigger blit operations (16+ pixels)
    ////////////////////////////////////////////////////////////////////////////////
    .p2align 4
.LBigBlit:
    // Align destination?
    testl       $0xF, %edx
    lddqu       (%eax), %xmm1           // Pre-load four source pixels
    jz          .LAligned

    movl        %edx, %edi              // Calculate alignment of destination pointer
    negl        %edi
    andl        $0xF, %edi

    // Handle 1-3 pixels to align destination
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    jz          .LAlignDone             // If all alphas are zero, just skip
    lddqu       (%edx), %xmm5           // Load four destination pixels
    jc          .LAlignStore            // If all alphas are opaque, just store

    // Handle mixed alphas (calculate and scale)
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value

    psllw       $8, %xmm3               // Filter out red and blue components
    pmulhuw     %xmm4, %xmm3            // Scale red and blue
    movdqa      %xmm5, %xmm2
    psrlw       $8, %xmm2               // Filter out alpha and green components
    pmullw      %xmm4, %xmm2            // Scale alpha and green

    pblendvb    %xmm3, %xmm2            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm2, %xmm1            // Add source and destination pixels together

.LAlignStore:
    cmp         $8, %edi                // Check how many pixels should be written
    jb          .LAlignPixelsLeft1
    ja          .LAlignPixelsLeft3
    pblendw     $0x0F, %xmm1, %xmm5     // Blend two pixels
    jmp .LAlignStorePixels

.LAlignPixelsLeft1:
    pblendw     $0x03, %xmm1, %xmm5     // Blend one pixel
    jmp .LAlignStorePixels

.LAlignPixelsLeft3:
    pblendw     $0x3F, %xmm1, %xmm5     // Blend three pixels

.LAlignStorePixels:
    movdqu      %xmm5, (%edx)           // Store destination pixels

.LAlignDone:
    addl        %edi, %eax              // Adjust pointers and pixel count
    addl        %edi, %edx
    shrl        $2, %edi
    lddqu       (%eax), %xmm1           // Pre-load new source pixels (after alignment)
    subl        %edi, %ecx

.LAligned:                              // Destination is guaranteed to be 16 byte aligned
    xorl        %edi, %edi              // Reset offset to zero
    subl        $8, %ecx                // Decrease counter (Reserve four pixels for the cleanup)
    testl       $0xF, %eax              // Check alignment of source pointer
    jz          .LAlignedLoop

    // Source not aligned to destination
    ////////////////////////////////////////////////////////////////////////////////
    .p2align 4
.LUnalignedLoop:                        // Main loop for unaligned, handles eight pixels per iteration
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero00
    lddqu       16(%eax, %edi), %xmm2   // Pre-load four source pixels
    jz          .LAlphaZero00
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels

.LAlphaZero00:
    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero01
    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
    jz          .LAlphaZero01
    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels

.LAlphaZero01:
    addl        $32, %edi               // Adjust offset and pixel count
    subl        $8, %ecx
    jae         .LUnalignedLoop
    addl        $8, %ecx                // Adjust pixel count
    jmp         .LLoopCleanup0

    .p2align 4
.LAlphaNotOpaqueOrZero00:
    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    lddqu       16(%eax, %edi), %xmm2   // Pre-load four source pixels
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm1            // Add source and destination pixels together
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels

    // Handle next four pixels
    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero01
    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
    jz          .LAlphaZero02
    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
.LAlphaZero02:
    addl        $32, %edi               // Adjust offset and pixel count
    subl        $8, %ecx
    jae         .LUnalignedLoop
    addl        $8, %ecx                // Adjust pixel count
    jmp         .LLoopCleanup0

    .p2align 4
.LAlphaNotOpaqueOrZero01:
    movdqa      16(%edx, %edi), %xmm5   // Load four destination pixels
    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    lddqu       32(%eax, %edi), %xmm1   // Pre-load four source pixels
    addl        $32, %edi
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm2            // Add source and destination pixels together
    subl        $8, %ecx
    movdqa      %xmm2, -16(%edx, %edi)  // Store four destination pixels
    jae         .LUnalignedLoop
    addl        $8, %ecx                // Adjust pixel count

    // Cleanup - handle pending pixels from loop
.LLoopCleanup0:
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero02
    jz          .LAlphaZero03
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
.LAlphaZero03:
    addl        $16, %edi
    subl        $4, %ecx
    js          .LSmallRemaining        // Reuse code from small loop

.LRemain0:
    lddqu       (%eax, %edi), %xmm1     // Load four source pixels
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero02
    jz          .LAlphaZero04
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
.LAlphaZero04:
    addl        $16, %edi
    subl        $4, %ecx
    jmp         .LSmallRemaining        // Reuse code from small loop

.LAlphaNotOpaqueOrZero02:
    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    addl        $16, %edi
    subl        $4, %ecx
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm1            // Add source and destination pixels together
    movdqa      %xmm1, -16(%edx, %edi)  // Store four destination pixels
    js          .LSmallRemaining        // Reuse code from small loop
    jmp         .LRemain0

    // Source aligned to destination
    ////////////////////////////////////////////////////////////////////////////////
    .p2align 4
.LAlignedLoop:                          // Main loop for aligned, handles eight pixels per iteration
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero10
    movdqa      16(%eax, %edi), %xmm2   // Pre-load four source pixels
    jz          .LAlphaZero10
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels

.LAlphaZero10:
    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero11
    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels
    jz          .LAlphaZero11
    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels

.LAlphaZero11:
    addl        $32, %edi               // Adjust offset and pixel count
    subl        $8, %ecx
    jae         .LAlignedLoop
    addl        $8, %ecx                // Adjust pixel count
    jmp         .LLoopCleanup1

    .p2align 4
.LAlphaNotOpaqueOrZero10:
    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    movdqa      16(%eax, %edi), %xmm2   // Pre-load four source pixels
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm1            // Add source and destination pixels together
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels

    // Handle next four pixels
    ptest       %xmm7, %xmm2            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero11
    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels
    jz          .LAlphaZero12
    movdqa      %xmm2, 16(%edx, %edi)   // Store four destination pixels
.LAlphaZero12:
    addl        $32, %edi               // Adjust offset and pixel count
    subl        $8, %ecx
    jae         .LAlignedLoop
    addl        $8, %ecx                // Adjust pixel count
    jmp         .LLoopCleanup1

    .p2align 4
.LAlphaNotOpaqueOrZero11:
    movdqa      16(%edx, %edi), %xmm5   // Load four destination pixels
    EXTRACT_ALPHA(xmm2, xmm1)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha
    movdqa      32(%eax, %edi), %xmm1   // Pre-load four source pixels

    addl        $32, %edi
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm2            // Add source and destination pixels together
    subl        $8, %ecx
    movdqa      %xmm2, -16(%edx, %edi)  // Store four destination pixels
    jae         .LAlignedLoop
    addl        $8, %ecx                // Adjust pixel count

    // Cleanup - handle pending pixels from loop
.LLoopCleanup1:
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero12
    jz          .LAlphaZero13
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
.LAlphaZero13:
    addl        $16, %edi
    subl        $4, %ecx
    js          .LSmallRemaining        // Reuse code from small loop

.LRemain1:
    movdqa      (%eax, %edi), %xmm1     // Load four source pixels
    ptest       %xmm7, %xmm1            // Check if all alphas are zero or opaque
    ja          .LAlphaNotOpaqueOrZero12
    jz          .LAlphaZero14
    movdqa      %xmm1, (%edx, %edi)     // Store four destination pixels
.LAlphaZero14:
    addl        $16, %edi
    subl        $4, %ecx
    jmp         .LSmallRemaining        // Reuse code from small loop

.LAlphaNotOpaqueOrZero12:
    movdqa      (%edx, %edi), %xmm5     // Load four destination pixels
    EXTRACT_ALPHA(xmm1, xmm2)           // Extract and clone alpha value
    SCALE_PIXELS                        // Scale pixels using alpha

    addl        $16, %edi
    subl        $4, %ecx
    pblendvb    %xmm5, %xmm3            // Combine results (mask in %xmm0, implicitly)
    paddb       %xmm3, %xmm1            // Add source and destination pixels together
    movdqa      %xmm1, -16(%edx, %edi)  // Store four destination pixels
    js          .LSmallRemaining        // Reuse code from small loop
    jmp         .LRemain1

    .cfi_endproc
#ifndef __clang__
    .size S32A_Opaque_BlitRow32_SSE4_asm, .-S32A_Opaque_BlitRow32_SSE4_asm
#endif
#endif