AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

compress_r11eac_blocks() required more alignment than dst has.

Browse Source 
This shouldn't change any behavior except that the stores to dst
will no longer require 8-byte alignment.

Empirically it seems like we can use 4-byte alignment here,
but u8 (i.e. 1-byte alignment) is always safe.

BUG=skia:5637
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2264103002
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review-Url: https://codereview.chromium.org/2264103002
This commit is contained in:
mtklein 2016-08-22 09:30:53 -07:00 committed by Commit bot
parent 2ec417257b
commit a61b6d4f9e
3 changed files with 12 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
infra/bots/recipes/swarm_test.expected/Test-Android-GCC-Nexus6-GPU-Adreno420-Arm7-Debug.json
View File

@ -898,9 +898,7 @@
"_", "_",
"image", "image",
"_", "_",
".SRW", ".SRW"
"--match",
"~CompressCheckerboard"
], ],
"env": { "env": {
"ANDROID_HOME": "[SLAVE_BUILD]/android_sdk/android-sdk", "ANDROID_HOME": "[SLAVE_BUILD]/android_sdk/android-sdk",

3
infra/bots/recipes/swarm_test.py
View File

@ -298,9 +298,6 @@ def dm_flags(bot):
if 'Valgrind' in bot: # skia:3021 if 'Valgrind' in bot: # skia:3021
match.append('~Threaded') match.append('~Threaded')
if 'Nexus6' in bot: # skia:5637
match.append('~CompressCheckerboard')
if 'GalaxyS3' in bot: # skia:1699 if 'GalaxyS3' in bot: # skia:1699
match.append('~WritePixels') match.append('~WritePixels')

23
src/opts/SkTextureCompressor_opts.h
View File

@ -150,14 +150,14 @@ namespace SK_OPTS_NS {
} }
#endif #endif
static void compress_r11eac_blocks(uint64_t* dst, const uint8_t* src, size_t rowBytes) { static void compress_r11eac_blocks(uint8_t* dst, const uint8_t* src, size_t rowBytes) {
// Try to avoid switching between vector and non-vector ops... // Try to avoid switching between vector and non-vector ops...
const uint8_t *const src1 = src; const uint8_t *const src1 = src;
const uint8_t *const src2 = src + rowBytes; const uint8_t *const src2 = src + rowBytes;
const uint8_t *const src3 = src + 2*rowBytes; const uint8_t *const src3 = src + 2*rowBytes;
const uint8_t *const src4 = src + 3*rowBytes; const uint8_t *const src4 = src + 3*rowBytes;
uint64_t *const dst1 = dst; uint8_t *const dst1 = dst;
uint64_t *const dst2 = dst + 2; uint8_t *const dst2 = dst + 16;
const uint8x16_t alphaRow1 = vld1q_u8(src1); const uint8x16_t alphaRow1 = vld1q_u8(src1);
const uint8x16_t alphaRow2 = vld1q_u8(src2); const uint8x16_t alphaRow2 = vld1q_u8(src2);
@ -175,12 +175,12 @@ namespace SK_OPTS_NS {
if (is_zero(alphaRow1)) { if (is_zero(alphaRow1)) {
static const uint64x2_t kTransparent = { 0x0020000000002000ULL, static const uint64x2_t kTransparent = { 0x0020000000002000ULL,
0x0020000000002000ULL }; 0x0020000000002000ULL };
vst1q_u64(dst1, kTransparent); vst1q_u8(dst1, vreinterpretq_u8_u64(kTransparent));
vst1q_u64(dst2, kTransparent); vst1q_u8(dst2, vreinterpretq_u8_u64(kTransparent));
return; return;
} else if (is_zero(nAlphaRow1)) { } else if (is_zero(nAlphaRow1)) {
vst1q_u64(dst1, vreinterpretq_u64_u8(cmp)); vst1q_u8(dst1, cmp);
vst1q_u64(dst2, vreinterpretq_u64_u8(cmp)); vst1q_u8(dst2, cmp);
return; return;
} }
} }
@ -205,8 +205,8 @@ namespace SK_OPTS_NS {
const uint64x2_t d1 = vcombine_u64(vget_low_u64(indicesLeft), vget_low_u64(indicesRight)); const uint64x2_t d1 = vcombine_u64(vget_low_u64(indicesLeft), vget_low_u64(indicesRight));
const uint64x2_t d2 = vcombine_u64(vget_high_u64(indicesLeft), vget_high_u64(indicesRight)); const uint64x2_t d2 = vcombine_u64(vget_high_u64(indicesLeft), vget_high_u64(indicesRight));
vst1q_u64(dst1, d1); vst1q_u8(dst1, vreinterpretq_u8_u64(d1));
vst1q_u64(dst2, d2); vst1q_u8(dst2, vreinterpretq_u8_u64(d2));
} }
static bool compress_a8_r11eac(uint8_t* dst, const uint8_t* src, static bool compress_a8_r11eac(uint8_t* dst, const uint8_t* src,
@ -224,12 +224,11 @@ namespace SK_OPTS_NS {
SkASSERT((blocksX % 4) == 0); SkASSERT((blocksX % 4) == 0);
uint64_t* encPtr = reinterpret_cast<uint64_t*>(dst);
for (int y = 0; y < blocksY; ++y) { for (int y = 0; y < blocksY; ++y) {
for (int x = 0; x < blocksX; x+=4) { for (int x = 0; x < blocksX; x+=4) {
// Compress it // Compress it
compress_r11eac_blocks(encPtr, src + 4*x, rowBytes); compress_r11eac_blocks(dst, src + 4*x, rowBytes);
encPtr += 4; dst += 32;
} }
src += 4 * rowBytes; src += 4 * rowBytes;
} }