From c885ec409f9b6ffa25e03851729b1bc2ad2005b3 Mon Sep 17 00:00:00 2001 From: Dale Curtis Date: Fri, 16 Dec 2022 22:37:46 +0000 Subject: [Backport] Security bug 1401571 Manual update of libdav1d to match the version introduced by patch https://chromium-review.googlesource.com/c/chromium/src/+/4114163: Roll src/third_party/dav1d/libdav1d/ 87f9a81cd..ed63a7459 (104 commits) This roll required a few changes to get working: - "properties" => "built in options" crossfile configuration change due to Meson deprecation. - generic config creation never worked, so fixed. - PPC64 configs were never checked in, so switched to generic. - copyright header changes for generate_sources. - Updated readme.chromium with potential issues that can arise. https://chromium.googlesource.com/external/github.com/videolan/dav1d.git/+log/87f9a81cd770..ed63a7459376 $ git log 87f9a81cd..ed63a7459 --date=short --no-merges --format='%ad %ae %s' 2022-12-09 jamrial dav1d: add an option to skip decoding some frame types 2022-12-08 jamrial picture: support creating and freeing refs without tile data 2022-12-07 gramner x86: Add 10bpc 8x32/32x8 itx AVX-512 (Ice Lake) asm 2022-12-07 gramner x86: Add minor DC-only IDCT optimizations 2022-12-13 gramner getbits: Fix assertion failure 2022-12-13 gramner checkasm: Fix integer overflow in refmvs test 2022-01-26 gramner dav1dplay: Update to new libplacebo API 2022-12-09 gramner Add minor getbits improvements 2022-12-09 gramner Add a separate getbits function for getting a single bit 2022-12-09 gramner Remove redundant zeroing in sequence header parsing 2022-12-09 gramner Set the correct default value of initial_display_delay 2022-12-09 jamrial tools: remove the null last entry in inloop_filters_tbl 2022-12-04 lu_zero Do not assume the picture allocation starts as the left edge 2022-11-21 lu_zero ppc: Allocate the correct temp buffer size 2022-11-21 lu_zero ppc: Do not use static const with vec_splats 2022-11-02 charlie.c.hayden Add info to dav1d_send_data docs 2022-10-30 jbeich build: drop -D_DARWIN_C_SOURCE on macOS/iOS after 6b611d36acab 2022-10-30 jbeich build: drop -D_POSIX_C_SOURCE on non-Linux after 6b611d36acab 2022-06-28 victorien threading: Add a pending list for async task insertion 2022-10-26 martin Implement atomic_compare_exchange_strong in the atomic compat headers 2022-10-06 victorien threading: Fix a race around frame completion (frame-mt) 2022-10-07 sebastian Handle host_machine.system() 'ios' and 'tvos' the same way as 'darwin' 2022-09-23 gramner x86: Add 10-bit 8x8/8x16/16x8/16x16 itx AVX-512 (Ice Lake) asm 2022-09-30 gramner Specify hidden visibility for global data symbol declarations 2022-09-28 gramner build: strip() the result of cc.get_define() 2022-09-26 gramner checkasm: Move printf format string to .rodata on x86 2022-09-26 gramner checkasm: Improve 32-bit parameter clobbering on x86-64 2022-09-26 gramner x86: Fix incorrect 32-bit parameter usage in high bit-depth AVX-512 mc 2022-09-09 martin arm: itx: Add clipping to row_clip_min/max in the 10 bpc codepaths 2022-09-15 gramner x86: Fix overflows in 12bpc AVX2 IDCT/IADST 2022-09-15 gramner x86: Fix overflows in 12bpc AVX2 DC-only IDCT 2022-09-15 gramner x86: Fix clipping in high bit-depth AVX2 4x16 IDCT 2022-03-21 martin Don't use gas-preprocessor with clang-cl for arm targets 2022-06-07 david_conrad Fix checking the reference dimesions for the projection process 2022-06-07 david_conrad Fix calculation of OBMC lap dimensions 2022-06-07 david_conrad Support film grain application whose only effect is clipping to video range 2022-06-07 david_conrad Ignore T.35 metadata if the OBU contains no payload 2022-06-07 david_conrad Fix chroma deblock filter size calculation for lossless 2022-06-07 david_conrad Fix rounding in the calculation of initialSubpelX 2022-06-07 david_conrad Fix overflow when saturating dequantized coefficients clipped to 0 2022-06-08 david_conrad Fix overflow in 8-bit NEON ADST 2022-09-14 martin tools: Allocate the priv structs with proper alignment 2022-09-08 gramner x86: Fix clipping in 10bpc SSE4.1 IDCT asm 2022-09-08 gramner build: Improve Windows linking options 2022-09-08 gramner tools: Improve demuxer probing 2022-08-30 code CI: Disable trimming on some tests 2022-08-30 code CI: Remove git 'safe.directory' config 2022-08-30 code gcovr: Ignore parsing errors 2022-08-30 code crossfiles: Update Android toolchains 2022-08-30 code CI: Update images (...) 2022-09-01 victorien checkasm: Add short options 2022-09-01 victorien checkasm: Add pattern matching to --test 2022-09-01 victorien checkasm: Remove pattern matching from --bench 2022-08-29 victorien checkasm: Add a --function option 2022-08-30 victorien threading: Fix copy_lpf_progress initialization 2022-08-19 jamrial data: don't overwrite the Dav1dDataProps size value 2022-07-18 gramner Adjust inlining attributes on some functions 2022-07-19 gramner x86: Remove leftover instruction in loopfilter AVX2 asm 2022-06-07 david_conrad Enable pointer authentication in assembly when building arm64e 2022-06-07 david_conrad Don't trash the return stack buffer in the NEON loop filter 2022-07-03 thresh CI: Removed snap package generation 2022-07-06 gramner Eliminate unused C DSP functions at compile time 2022-07-06 gramner cpu: Inline dav1d_get_cpu_flags() 2022-06-22 gramner x86: Add minor loopfilter asm improvements 2022-06-15 gramner checkasm: Speed up signal handling 2022-06-15 gramner checkasm: Improve seed generation on Windows 2022-06-20 gramner ci: Don't specify a specific MacOS version 2022-06-14 gramner x86: Add high bit-depth loopfilter AVX-512 (Ice Lake) asm 2022-06-13 victorien checkasm/lpf: Use operating dimensions 2022-06-03 gramner checkasm: Print the cpu model and cpuid signature on x86 2022-06-03 gramner checkasm: Add a vzeroupper check on x86 2022-06-02 gramner x86: Add a workaround for quirky AVX-512 hardware behavior 2022-05-31 victorien checkasm: Fix uninitialized variable 2022-05-14 code CI: Update coverage collecting 2022-05-05 code CI: Add a build with the minimum requirements 2022-05-05 code CI: Deactivate git 'safe.directory' 2022-03-24 code CI: Update images 2022-05-25 victorien Fix typo 2022-05-19 gramner x86: Add high bit-depth cdef_filter AVX-512 (Ice Lake) asm 2022-05-20 gramner checkasm: Print --help message to stderr instead of stdout 2022-05-20 gramner checkasm: Split cdef test into separate pri/sec/pri+sec parts 2022-05-20 gramner checkasm: Improve benchmarking of functions that modify their input 2022-05-18 b x86/itx_avx2: fix typo 2022-04-22 code CI: Add gcc12 and clang14 builds with mold linker 2022-04-26 code CI: Trigger documentation rebuild if configuration changes 2022-04-24 code meson/doc: Fix doxygen config 2022-04-28 gramner Use a relaxed memory ordering in dav1d_ref_inc() 2022-04-28 gramner Remove redundant code in dav1d_cdf_thread_unref() 2022-04-28 gramner Inline dav1d_ref_inc() 2022-04-24 code x86/itx: Add 32x8 12bpc AVX2 transforms 2022-04-24 code x86/itx: Add 8x32 12bpc AVX2 transforms 2022-04-24 code x86/itx: Deduplicate dconly code 2022-04-23 code lib: Fix typo in documentation 2022-04-07 jamrial obu: don't output invisible but showable key frames more than once 2022-04-07 jamrial obu: check that the frame referenced by existing_frame_idx is showable 2022-04-07 jamrial obu: check refresh_frame_flags is not equal to allFrames on Intra Only frames 2022-03-29 robux4 remove multipass wait from dav1d_decode_frame 2022-04-07 jamrial picture: ensure the new seq header and op param info flags are attached to the next visible picture in display order 2022-03-31 jamrial lib: add a function to query the decoder frame delay 2022-03-31 jamrial lib: split calculating thread count to its own function Created with: roll-dep src/third_party/dav1d/libdav1d Fixed: 1401571 Change-Id: Ic3cef540a87a2cf411abe6071fd4c9963ea61f75 Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/4114163 Reviewed-by: Wan-Teh Chang Commit-Queue: Dale Curtis Cr-Commit-Position: refs/heads/main@{#1084574} Reviewed-on: https://codereview.qt-project.org/c/qt/qtwebengine-chromium/+/468619 Reviewed-by: Michal Klocek --- .../third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c | 487 +++++++++++++++++++++ 1 file changed, 487 insertions(+) create mode 100644 chromium/third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c (limited to 'chromium/third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c') diff --git a/chromium/third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c b/chromium/third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c new file mode 100644 index 00000000000..e2e759810f7 --- /dev/null +++ b/chromium/third_party/dav1d/libdav1d/src/ppc/cdef_tmpl.c @@ -0,0 +1,487 @@ +/* + * Copyright © 2019, Luca Barbato + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are met: + * + * 1. Redistributions of source code must retain the above copyright notice, this + * list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright notice, + * this list of conditions and the following disclaimer in the documentation + * and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND + * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include "src/ppc/dav1d_types.h" +#include "src/ppc/cdef.h" + +#if BITDEPTH == 8 +static inline i16x8 vconstrain(const i16x8 diff, const int16_t threshold, + const int damping) +{ + const i16x8 zero = vec_splat_s16(0); + if (!threshold) return zero; + const uint16_t shift = imax(0, damping - ulog2(threshold)); + const i16x8 abs_diff = vec_abs(diff); + const b16x8 mask = vec_cmplt(diff, zero); + const i16x8 thr = vec_splats(threshold); + const i16x8 sub = vec_sub(thr, vec_sra(abs_diff, vec_splats(shift))); + const i16x8 max = vec_max(zero, sub); + const i16x8 min = vec_min(abs_diff, max); + const i16x8 neg = vec_sub(zero, min); + return vec_sel(min, neg, mask); +} + +static inline void copy4xN(uint16_t *tmp, const ptrdiff_t tmp_stride, + const uint8_t *src, const ptrdiff_t src_stride, + const uint8_t (*left)[2], const uint8_t *const top, + const uint8_t *const bottom, const int w, const int h, + const enum CdefEdgeFlags edges) +{ + const u16x8 fill = vec_splats((uint16_t)INT16_MAX); + + u16x8 l0; + u16x8 l1; + + int y_start = -2, y_end = h + 2; + + // Copy top and bottom first + if (!(edges & CDEF_HAVE_TOP)) { + l0 = fill; + l1 = fill; + y_start = 0; + } else { + l0 = u8h_to_u16(vec_vsx_ld(0, top + 0 * src_stride - 2)); + l1 = u8h_to_u16(vec_vsx_ld(0, top + 1 * src_stride - 2)); + } + + vec_st(l0, 0, tmp - 2 * 8); + vec_st(l1, 0, tmp - 1 * 8); + + if (!(edges & CDEF_HAVE_BOTTOM)) { + l0 = fill; + l1 = fill; + y_end -= 2; + } else { + l0 = u8h_to_u16(vec_vsx_ld(0, bottom + 0 * src_stride - 2)); + l1 = u8h_to_u16(vec_vsx_ld(0, bottom + 1 * src_stride - 2)); + } + + vec_st(l0, 0, tmp + (h + 0) * 8); + vec_st(l1, 0, tmp + (h + 1) * 8); + + int y_with_left_edge = 0; + if (!(edges & CDEF_HAVE_LEFT)) { + u16x8 l = u8h_to_u16(vec_vsx_ld(0, src)); + vec_vsx_st(l, 0, tmp + 2); + + y_with_left_edge = 1; + } + + for (int y = y_with_left_edge; y < h; y++) { + u16x8 l = u8h_to_u16(vec_vsx_ld(0, src - 2 + y * src_stride)); + vec_st(l, 0, tmp + y * 8); + } + + if (!(edges & CDEF_HAVE_LEFT)) { + for (int y = y_start; y < y_end; y++) { + tmp[y * 8] = INT16_MAX; + tmp[1 + y * 8] = INT16_MAX; + } + } else { + for (int y = 0; y < h; y++) { + tmp[y * 8] = left[y][0]; + tmp[1 + y * 8] = left[y][1]; + } + } + if (!(edges & CDEF_HAVE_RIGHT)) { + for (int y = y_start; y < y_end; y++) { + tmp[- 2 + (y + 1) * 8] = INT16_MAX; + tmp[- 1 + (y + 1) * 8] = INT16_MAX; + } + } +} + +static inline void copy8xN(uint16_t *tmp, const ptrdiff_t tmp_stride, + const uint8_t *src, const ptrdiff_t src_stride, + const uint8_t (*left)[2], const uint8_t *const top, + const uint8_t *const bottom, const int w, const int h, + const enum CdefEdgeFlags edges) +{ + const u16x8 fill = vec_splats((uint16_t)INT16_MAX); + + u16x8 l0h, l0l; + u16x8 l1h, l1l; + + int y_start = -2, y_end = h + 2; + + // Copy top and bottom first + if (!(edges & CDEF_HAVE_TOP)) { + l0h = fill; + l0l = fill; + l1h = fill; + l1l = fill; + y_start = 0; + } else { + u8x16 l0 = vec_vsx_ld(0, top + 0 * src_stride - 2); + u8x16 l1 = vec_vsx_ld(0, top + 1 * src_stride - 2); + l0h = u8h_to_u16(l0); + l0l = u8l_to_u16(l0); + l1h = u8h_to_u16(l1); + l1l = u8l_to_u16(l1); + } + + vec_st(l0h, 0, tmp - 4 * 8); + vec_st(l0l, 0, tmp - 3 * 8); + vec_st(l1h, 0, tmp - 2 * 8); + vec_st(l1l, 0, tmp - 1 * 8); + + if (!(edges & CDEF_HAVE_BOTTOM)) { + l0h = fill; + l0l = fill; + l1h = fill; + l1l = fill; + y_end -= 2; + } else { + u8x16 l0 = vec_vsx_ld(0, bottom + 0 * src_stride - 2); + u8x16 l1 = vec_vsx_ld(0, bottom + 1 * src_stride - 2); + l0h = u8h_to_u16(l0); + l0l = u8l_to_u16(l0); + l1h = u8h_to_u16(l1); + l1l = u8l_to_u16(l1); + } + + vec_st(l0h, 0, tmp + (h + 0) * 16); + vec_st(l0l, 0, tmp + (h + 0) * 16 + 8); + vec_st(l1h, 0, tmp + (h + 1) * 16); + vec_st(l1l, 0, tmp + (h + 1) * 16 + 8); + + int y_with_left_edge = 0; + if (!(edges & CDEF_HAVE_LEFT)) { + u8x16 l = vec_vsx_ld(0, src); + u16x8 lh = u8h_to_u16(l); + u16x8 ll = u8l_to_u16(l); + vec_vsx_st(lh, 0, tmp + 2); + vec_vsx_st(ll, 0, tmp + 8 + 2); + + y_with_left_edge = 1; + } + + for (int y = y_with_left_edge; y < h; y++) { + u8x16 l = vec_vsx_ld(0, src - 2 + y * src_stride); + u16x8 lh = u8h_to_u16(l); + u16x8 ll = u8l_to_u16(l); + vec_st(lh, 0, tmp + y * 16); + vec_st(ll, 0, tmp + 8 + y * 16); + } + + if (!(edges & CDEF_HAVE_LEFT)) { + for (int y = y_start; y < y_end; y++) { + tmp[y * 16] = INT16_MAX; + tmp[1 + y * 16] = INT16_MAX; + } + } else { + for (int y = 0; y < h; y++) { + tmp[y * 16] = left[y][0]; + tmp[1 + y * 16] = left[y][1]; + } + } + if (!(edges & CDEF_HAVE_RIGHT)) { + for (int y = y_start; y < y_end; y++) { + tmp[- 6 + (y + 1) * 16] = INT16_MAX; + tmp[- 5 + (y + 1) * 16] = INT16_MAX; + } + } +} + +static inline i16x8 max_mask(i16x8 a, i16x8 b) { + const i16x8 I16X8_INT16_MAX = vec_splats((int16_t)INT16_MAX); + + const b16x8 mask = vec_cmpeq(a, I16X8_INT16_MAX); + + const i16x8 val = vec_sel(a, b, mask); + + return vec_max(val, b); +} + +#define LOAD_PIX(addr) \ + const i16x8 px = (i16x8)vec_vsx_ld(0, addr); \ + i16x8 max = px; \ + i16x8 min = px; \ + i16x8 sum = vec_splat_s16(0); + +#define LOAD_PIX4(addr) \ + const i16x8 a = (i16x8)vec_vsx_ld(0, addr); \ + const i16x8 b = (i16x8)vec_vsx_ld(0, addr + tmp_stride); \ + const i16x8 px = vec_xxpermdi(a, b, 0); \ + i16x8 max = px; \ + i16x8 min = px; \ + i16x8 sum = vec_splat_s16(0); + +#define LOAD_DIR(p, addr, o0, o1) \ + const i16x8 p ## 0 = (i16x8)vec_vsx_ld(0, addr + o0); \ + const i16x8 p ## 1 = (i16x8)vec_vsx_ld(0, addr - o0); \ + const i16x8 p ## 2 = (i16x8)vec_vsx_ld(0, addr + o1); \ + const i16x8 p ## 3 = (i16x8)vec_vsx_ld(0, addr - o1); + +#define LOAD_DIR4(p, addr, o0, o1) \ + LOAD_DIR(p ## a, addr, o0, o1) \ + LOAD_DIR(p ## b, addr + tmp_stride, o0, o1) \ + const i16x8 p ## 0 = vec_xxpermdi(p ## a ## 0, p ## b ## 0, 0); \ + const i16x8 p ## 1 = vec_xxpermdi(p ## a ## 1, p ## b ## 1, 0); \ + const i16x8 p ## 2 = vec_xxpermdi(p ## a ## 2, p ## b ## 2, 0); \ + const i16x8 p ## 3 = vec_xxpermdi(p ## a ## 3, p ## b ## 3, 0); + +#define CONSTRAIN(p, strength) \ + const i16x8 p ## _d0 = vec_sub(p ## 0, px); \ + const i16x8 p ## _d1 = vec_sub(p ## 1, px); \ + const i16x8 p ## _d2 = vec_sub(p ## 2, px); \ + const i16x8 p ## _d3 = vec_sub(p ## 3, px); \ +\ + i16x8 p ## _c0 = vconstrain(p ## _d0, strength, damping); \ + i16x8 p ## _c1 = vconstrain(p ## _d1, strength, damping); \ + i16x8 p ## _c2 = vconstrain(p ## _d2, strength, damping); \ + i16x8 p ## _c3 = vconstrain(p ## _d3, strength, damping); + +#define MIN_MAX(p) \ + max = max_mask(p ## 0, max); \ + min = vec_min(p ## 0, min); \ + max = max_mask(p ## 1, max); \ + min = vec_min(p ## 1, min); \ + max = max_mask(p ## 2, max); \ + min = vec_min(p ## 2, min); \ + max = max_mask(p ## 3, max); \ + min = vec_min(p ## 3, min); + +#define PRI_0(p) \ + p ## _c0 = vec_add(vec_sl(p ## _c0, vec_splat_u16(1)), vec_sl(p ## _c0, vec_splats(tap_even))); \ + p ## _c1 = vec_add(vec_sl(p ## _c1, vec_splat_u16(1)), vec_sl(p ## _c1, vec_splats(tap_even))); + +#define PRI_1(p) \ + p ## _c2 = vec_sub(vec_sl(p ## _c2, vec_splat_u16(2)), vec_sl(p ## _c2, vec_splats(tap_even))); \ + p ## _c3 = vec_sub(vec_sl(p ## _c3, vec_splat_u16(2)), vec_sl(p ## _c3, vec_splats(tap_even))); + +#define SEC_0(p) \ + p ## _c0 = vec_sl(p ## _c0, vec_splat_u16(1)); \ + p ## _c1 = vec_sl(p ## _c1, vec_splat_u16(1)); \ + p ## _c2 = vec_sl(p ## _c2, vec_splat_u16(1)); \ + p ## _c3 = vec_sl(p ## _c3, vec_splat_u16(1)); + +#define UPDATE_SUM(p) \ + const i16x8 p ## sum0 = vec_add(p ## _c0, p ## _c1); \ + const i16x8 p ## sum1 = vec_add(p ## _c2, p ## _c3); \ + sum = vec_add(sum, p ## sum0); \ + sum = vec_add(sum, p ## sum1); + +static inline void +filter_4xN(pixel *dst, const ptrdiff_t dst_stride, + const pixel (*left)[2], const pixel *const top, + const pixel *const bottom, const int w, const int h, + const int pri_strength, const int sec_strength, const int dir, + const int damping, const enum CdefEdgeFlags edges, + const ptrdiff_t tmp_stride, uint16_t *tmp) +{ + const int8_t cdef_directions[8 /* dir */][2 /* pass */] = { + { -1 * tmp_stride + 1, -2 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, -1 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, 0 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, 1 * tmp_stride + 2 }, + { 1 * tmp_stride + 1, 2 * tmp_stride + 2 }, + { 1 * tmp_stride + 0, 2 * tmp_stride + 1 }, + { 1 * tmp_stride + 0, 2 * tmp_stride + 0 }, + { 1 * tmp_stride + 0, 2 * tmp_stride - 1 } + }; + const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8; + const uint16_t tap_even = !((pri_strength >> bitdepth_min_8) & 1); + const int off1 = cdef_directions[dir][0]; + const int off1_1 = cdef_directions[dir][1]; + + const int off2 = cdef_directions[(dir + 2) & 7][0]; + const int off3 = cdef_directions[(dir + 6) & 7][0]; + + const int off2_1 = cdef_directions[(dir + 2) & 7][1]; + const int off3_1 = cdef_directions[(dir + 6) & 7][1]; + + copy4xN(tmp - 2, tmp_stride, dst, dst_stride, left, top, bottom, w, h, edges); + + for (int y = 0; y < h / 2; y++) { + LOAD_PIX4(tmp) + + // Primary pass + LOAD_DIR4(p, tmp, off1, off1_1) + + CONSTRAIN(p, pri_strength) + + MIN_MAX(p) + + PRI_0(p) + PRI_1(p) + + UPDATE_SUM(p) + + // Secondary pass 1 + LOAD_DIR4(s, tmp, off2, off3) + + CONSTRAIN(s, sec_strength) + + MIN_MAX(s) + + SEC_0(s) + + UPDATE_SUM(s) + + // Secondary pass 2 + LOAD_DIR4(s2, tmp, off2_1, off3_1) + + CONSTRAIN(s2, sec_strength) + + MIN_MAX(s2) + + UPDATE_SUM(s2) + + // Store + i16x8 bias = vec_and((i16x8)vec_cmplt(sum, vec_splat_s16(0)), vec_splat_s16(1)); + bias = vec_sub(vec_splat_s16(8), bias); + i16x8 unclamped = vec_add(px, vec_sra(vec_add(sum, bias), vec_splat_u16(4))); + i16x8 vdst = vec_max(vec_min(unclamped, max), min); + + dst[0] = vdst[0]; + dst[1] = vdst[1]; + dst[2] = vdst[2]; + dst[3] = vdst[3]; + + tmp += tmp_stride; + dst += PXSTRIDE(dst_stride); + dst[0] = vdst[4]; + dst[1] = vdst[5]; + dst[2] = vdst[6]; + dst[3] = vdst[7]; + + tmp += tmp_stride; + dst += PXSTRIDE(dst_stride); + } +} + +static inline void +filter_8xN(pixel *dst, const ptrdiff_t dst_stride, + const pixel (*left)[2], const pixel *const top, + const pixel *const bottom, const int w, const int h, + const int pri_strength, const int sec_strength, const int dir, + const int damping, const enum CdefEdgeFlags edges, + const ptrdiff_t tmp_stride, uint16_t *tmp) +{ + const int8_t cdef_directions[8 /* dir */][2 /* pass */] = { + { -1 * tmp_stride + 1, -2 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, -1 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, 0 * tmp_stride + 2 }, + { 0 * tmp_stride + 1, 1 * tmp_stride + 2 }, + { 1 * tmp_stride + 1, 2 * tmp_stride + 2 }, + { 1 * tmp_stride + 0, 2 * tmp_stride + 1 }, + { 1 * tmp_stride + 0, 2 * tmp_stride + 0 }, + { 1 * tmp_stride + 0, 2 * tmp_stride - 1 } + }; + const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8; + + + const uint16_t tap_even = !((pri_strength >> bitdepth_min_8) & 1); + const int off1 = cdef_directions[dir][0]; + const int off1_1 = cdef_directions[dir][1]; + + const int off2 = cdef_directions[(dir + 2) & 7][0]; + const int off3 = cdef_directions[(dir + 6) & 7][0]; + + const int off2_1 = cdef_directions[(dir + 2) & 7][1]; + const int off3_1 = cdef_directions[(dir + 6) & 7][1]; + + copy8xN(tmp - 2, tmp_stride, dst, dst_stride, left, top, bottom, w, h, edges); + + for (int y = 0; y < h; y++) { + LOAD_PIX(tmp) + + // Primary pass + LOAD_DIR(p, tmp, off1, off1_1) + + CONSTRAIN(p, pri_strength) + + MIN_MAX(p) + + PRI_0(p) + PRI_1(p) + + UPDATE_SUM(p) + + // Secondary pass 1 + LOAD_DIR(s, tmp, off2, off3) + + CONSTRAIN(s, sec_strength) + + MIN_MAX(s) + + SEC_0(s) + + UPDATE_SUM(s) + + // Secondary pass 2 + LOAD_DIR(s2, tmp, off2_1, off3_1) + + CONSTRAIN(s2, sec_strength) + + MIN_MAX(s2) + + UPDATE_SUM(s2) + + // Store + i16x8 bias = vec_and((i16x8)vec_cmplt(sum, vec_splat_s16(0)), vec_splat_s16(1)); + bias = vec_sub(vec_splat_s16(8), bias); + i16x8 unclamped = vec_add(px, vec_sra(vec_add(sum, bias), vec_splat_u16(4))); + i16x8 vdst = vec_max(vec_min(unclamped, max), min); + + dst[0] = vdst[0]; + dst[1] = vdst[1]; + dst[2] = vdst[2]; + dst[3] = vdst[3]; + dst[4] = vdst[4]; + dst[5] = vdst[5]; + dst[6] = vdst[6]; + dst[7] = vdst[7]; + + tmp += tmp_stride; + dst += PXSTRIDE(dst_stride); + } + +} + +#define cdef_fn(w, h, tmp_stride) \ +void dav1d_cdef_filter_##w##x##h##_vsx(pixel *const dst, \ + const ptrdiff_t dst_stride, \ + const pixel (*left)[2], \ + const pixel *const top, \ + const pixel *const bottom, \ + const int pri_strength, \ + const int sec_strength, \ + const int dir, \ + const int damping, \ + const enum CdefEdgeFlags edges) \ +{ \ + ALIGN_STK_16(uint16_t, tmp_buf, 12 * tmp_stride + 8,); \ + uint16_t *tmp = tmp_buf + 2 * tmp_stride + 2; \ + filter_##w##xN(dst, dst_stride, left, top, bottom, w, h, pri_strength, \ + sec_strength, dir, damping, edges, tmp_stride, tmp); \ +} + +cdef_fn(4, 4, 8); +cdef_fn(4, 8, 8); +cdef_fn(8, 8, 16); +#endif -- cgit v1.2.1