Arm: Avoid over-read in 4-wide intrabc convolutions av1_convolve_x_sr_intrabc_neon() and the horizontal pass of av1_convolve_2d_sr_intrabc_neon() load 8 bytes from src and src + 1 for blocks 4 or fewer pixels wide, though only w + 1 pixels per row are used. The extra pixels lie to the right of the reference block. With IntraBC in row based multithreading, as in DatarateTestSVC.BasicRateTargetingSVC2TL1SLScreenDropFrame1920x10804Thread, those pixels can be in a superblock that another thread is reconstructing at the same time. The read values are discarded, but ThreadSanitizer reports a data race between the read in av1_convolve_2d_sr_intrabc_neon() and the intra prediction writes of the other thread. Load only 4 bytes per row from src and src + 1, as av1_convolve_y_sr_intrabc_neon() already does. Change-Id: I3aeb96861d384ed2fb3c0c231d86478961e9944b
diff --git a/av1/common/arm/convolve_neon.c b/av1/common/arm/convolve_neon.c index fc42f9e..68521f2 100644 --- a/av1/common/arm/convolve_neon.c +++ b/av1/common/arm/convolve_neon.c
@@ -1648,13 +1648,15 @@ if (w <= 4) { do { - uint8x8_t s0_0 = vld1_u8(src); - uint8x8_t s0_1 = vld1_u8(src + 1); - uint8x8_t s1_0 = vld1_u8(src + src_stride); - uint8x8_t s1_1 = vld1_u8(src + src_stride + 1); + // Load at most 5 pixels per row. A full 8-byte load would read + // pixels to the right of the reference block, which may be in a region + // that is concurrently being reconstructed by another encoder thread. + uint8x8_t s0 = load_unaligned_u8(src, src_stride); + uint8x8_t s1 = load_unaligned_u8(src + 1, src_stride); - uint8x8_t d0 = vrhadd_u8(s0_0, s0_1); - uint8x8_t d1 = vrhadd_u8(s1_0, s1_1); + // Row 0 is in the low half of d0, row 1 in the high half. + uint8x8_t d0 = vrhadd_u8(s0, s1); + uint8x8_t d1 = vext_u8(d0, d0, 4); if (w == 2) { store_u8_2x1(dst + 0 * dst_stride, d0); @@ -1806,8 +1808,11 @@ // Horizontal filter. if (w <= 4) { do { - uint8x8_t s0 = vld1_u8(src); - uint8x8_t s1 = vld1_u8(src + 1); + // Load at most 5 pixels per row. A full 8-byte load would read + // pixels to the right of the reference block, which may be in a region + // that is concurrently being reconstructed by another encoder thread. + uint8x8_t s0 = load_unaligned_u8_4x1(src); + uint8x8_t s1 = load_unaligned_u8_4x1(src + 1); uint16x4_t sum = vget_low_u16(vaddl_u8(s0, s1));