max/dav1d
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

x86: Add 8-bit film grain AVX-512 (Ice Lake) asm

Browse Source
This commit is contained in:
Henrik Gramner
2022-02-17 23:21:18 +01:00
committed by Henrik Gramner
parent ba07371c2e
commit 0028e8ea57
10 changed files with 1172 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/checkasm/filmgrain.c
+4 -4
View File
@@ -155,6 +155,7 @@ static void check_fgy_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
if (check_func(dsp->fgy_32x32xn, "fgy_32x32xn_%dbpc", BITDEPTH)) {
ALIGN_STK_16(Dav1dFilmGrainData, fg_data, 16,);
ALIGN_STK_64(uint8_t, scaling, SCALING_SIZE,);
fg_data[0].seed = rnd() & 0xFFFF;
#if BITDEPTH == 16
@@ -163,7 +164,6 @@ static void check_fgy_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
const int bitdepth_max = 0xff;
#endif
uint8_t scaling[SCALING_SIZE];
entry grain_lut[GRAIN_HEIGHT + 1][GRAIN_WIDTH];
fg_data[0].grain_scale_shift = rnd() & 3;
fg_data[0].ar_coeff_shift = (rnd() & 3) + 6;
@@ -267,6 +267,7 @@ static void check_fguv_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
BITDEPTH, ss_name[layout_idx], csfl))
{
ALIGN_STK_16(Dav1dFilmGrainData, fg_data, 1,);
ALIGN_STK_64(uint8_t, scaling, SCALING_SIZE,);
fg_data[0].seed = rnd() & 0xFFFF;
@@ -278,7 +279,6 @@ static void check_fguv_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
const int uv_pl = rnd() & 1;
const int is_identity = rnd() & 1;
uint8_t scaling[SCALING_SIZE];
entry grain_lut[2][GRAIN_HEIGHT + 1][GRAIN_WIDTH];
fg_data[0].grain_scale_shift = rnd() & 3;
fg_data[0].ar_coeff_shift = (rnd() & 3) + 6;
@@ -368,7 +368,7 @@ static void check_fguv_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
checkasm_check_pixel_padded_align(c_dst, stride,
a_dst, stride,
w, h, "dst",
32 >> ss_x, 2);
32 >> ss_x, 4);
}
}
@@ -380,7 +380,7 @@ static void check_fguv_sbrow(const Dav1dFilmGrainDSPContext *const dsp) {
luma_src[y * PXSTRIDE(lstride) + x] &= bitdepth_max;
}
}
bench_new(a_dst, src, stride, fg_data, 32, scaling, grain_lut[1], 16,
bench_new(a_dst, src, stride, fg_data, 64 >> ss_x, scaling, grain_lut[1], 32 >> ss_y,
1, luma_src, lstride, uv_pl, is_identity HIGHBD_TAIL_SUFFIX);
}
}