8bpc film grain synthesis Zen 2

taskset -c 4-7,12-15 /home/bluezakm/dav1d_clang_O3_LTO_march-native/build/tests/checkasm  --test=filmgrain_8bpc --bench
checkasm: using random seed 3801148340
SSSE3:
 - filmgrain_8bpc.gen_grain_y  [OK]
 - filmgrain_8bpc.gen_grain_uv [OK]
 - filmgrain_8bpc.fgy_32x32xn  [OK]
 - filmgrain_8bpc.fguv_32x32xn [OK]
AVX2:
 - filmgrain_8bpc.gen_grain_y  [OK]
 - filmgrain_8bpc.gen_grain_uv [OK]
 - filmgrain_8bpc.fgy_32x32xn  [OK]
 - filmgrain_8bpc.fguv_32x32xn [OK]
checkasm: all 46 tests passed
nop: 51.2
fguv_32x32xn_8bpc_420_csfl0_c: 6990.9
fguv_32x32xn_8bpc_420_csfl0_ssse3: 986.7
fguv_32x32xn_8bpc_420_csfl0_avx2: 1267.4
fguv_32x32xn_8bpc_420_csfl1_c: 5156.6
fguv_32x32xn_8bpc_420_csfl1_ssse3: 701.2
fguv_32x32xn_8bpc_420_csfl1_avx2: 1096.6
fguv_32x32xn_8bpc_422_csfl0_c: 7169.1
fguv_32x32xn_8bpc_422_csfl0_ssse3: 990.3
fguv_32x32xn_8bpc_422_csfl0_avx2: 1258.6
fguv_32x32xn_8bpc_422_csfl1_c: 5281.4
fguv_32x32xn_8bpc_422_csfl1_ssse3: 708.0
fguv_32x32xn_8bpc_422_csfl1_avx2: 1080.8
fguv_32x32xn_8bpc_444_csfl0_c: 6404.1
fguv_32x32xn_8bpc_444_csfl0_ssse3: 764.2
fguv_32x32xn_8bpc_444_csfl0_avx2: 1054.5
fguv_32x32xn_8bpc_444_csfl1_c: 4204.9
fguv_32x32xn_8bpc_444_csfl1_ssse3: 614.7
fguv_32x32xn_8bpc_444_csfl1_avx2: 986.7
fgy_32x32xn_8bpc_c: 12041.6
fgy_32x32xn_8bpc_ssse3: 2440.5
fgy_32x32xn_8bpc_avx2: 3872.1
gen_grain_uv_ar0_8bpc_420_c: 35895.2
gen_grain_uv_ar0_8bpc_420_ssse3: 6172.7
gen_grain_uv_ar0_8bpc_420_avx2: 6816.4
gen_grain_uv_ar0_8bpc_422_c: 63727.5
gen_grain_uv_ar0_8bpc_422_ssse3: 11579.2
gen_grain_uv_ar0_8bpc_422_avx2: 12843.0
gen_grain_uv_ar0_8bpc_444_c: 121788.7
gen_grain_uv_ar0_8bpc_444_ssse3: 20816.7
gen_grain_uv_ar0_8bpc_444_avx2: 23342.3
gen_grain_uv_ar1_8bpc_420_c: 48306.5
gen_grain_uv_ar1_8bpc_420_ssse3: 16082.4
gen_grain_uv_ar1_8bpc_420_avx2: 17156.2
gen_grain_uv_ar1_8bpc_422_c: 88317.3
gen_grain_uv_ar1_8bpc_422_ssse3: 31824.7
gen_grain_uv_ar1_8bpc_422_avx2: 33498.4
gen_grain_uv_ar1_8bpc_444_c: 165709.8
gen_grain_uv_ar1_8bpc_444_ssse3: 63030.8
gen_grain_uv_ar1_8bpc_444_avx2: 66023.4
gen_grain_uv_ar2_8bpc_420_c: 69231.4
gen_grain_uv_ar2_8bpc_420_ssse3: 23650.7
gen_grain_uv_ar2_8bpc_420_avx2: 23177.0
gen_grain_uv_ar2_8bpc_422_c: 129804.7
gen_grain_uv_ar2_8bpc_422_ssse3: 46725.8
gen_grain_uv_ar2_8bpc_422_avx2: 45623.2
gen_grain_uv_ar2_8bpc_444_c: 248778.9
gen_grain_uv_ar2_8bpc_444_ssse3: 91129.6
gen_grain_uv_ar2_8bpc_444_avx2: 89971.5
gen_grain_uv_ar3_8bpc_420_c: 99806.4
gen_grain_uv_ar3_8bpc_420_ssse3: 26463.0
gen_grain_uv_ar3_8bpc_420_avx2: 25874.9
gen_grain_uv_ar3_8bpc_422_c: 191304.2
gen_grain_uv_ar3_8bpc_422_ssse3: 52296.1
gen_grain_uv_ar3_8bpc_422_avx2: 50920.4
gen_grain_uv_ar3_8bpc_444_c: 373250.7
gen_grain_uv_ar3_8bpc_444_ssse3: 102954.3
gen_grain_uv_ar3_8bpc_444_avx2: 99264.7
gen_grain_y_ar0_8bpc_c: 95058.7
gen_grain_y_ar0_8bpc_ssse3: 18087.1
gen_grain_y_ar0_8bpc_avx2: 22459.1
gen_grain_y_ar1_8bpc_c: 141876.6
gen_grain_y_ar1_8bpc_ssse3: 61732.0
gen_grain_y_ar1_8bpc_avx2: 66239.4
gen_grain_y_ar2_8bpc_c: 220122.5
gen_grain_y_ar2_8bpc_ssse3: 94537.1
gen_grain_y_ar2_8bpc_avx2: 93278.1
gen_grain_y_ar3_8bpc_c: 342873.4
gen_grain_y_ar3_8bpc_ssse3: 101492.4
gen_grain_y_ar3_8bpc_avx2: 99977.9