16bpc film grain synthesis Zen 2

taskset -c 4-7,12-15 /home/bluezakm/dav1d_clang_O3_LTO_march-native/build/tests/checkasm  --test=filmgrain_16bpc --bench
checkasm: using random seed 3757977363
SSSE3:
 - filmgrain_16bpc.gen_grain_y  [OK]
 - filmgrain_16bpc.gen_grain_uv [OK]
 - filmgrain_16bpc.fgy_32x32xn  [OK]
 - filmgrain_16bpc.fguv_32x32xn [OK]
AVX2:
 - filmgrain_16bpc.gen_grain_y  [OK]
 - filmgrain_16bpc.gen_grain_uv [OK]
 - filmgrain_16bpc.fgy_32x32xn  [OK]
 - filmgrain_16bpc.fguv_32x32xn [OK]
checkasm: all 46 tests passed
nop: 49.2
fguv_32x32xn_16bpc_420_csfl0_c: 6628.1
fguv_32x32xn_16bpc_420_csfl0_ssse3: 941.5
fguv_32x32xn_16bpc_420_csfl0_avx2: 1246.0
fguv_32x32xn_16bpc_420_csfl1_c: 5293.8
fguv_32x32xn_16bpc_420_csfl1_ssse3: 705.8
fguv_32x32xn_16bpc_420_csfl1_avx2: 1106.5
fguv_32x32xn_16bpc_422_csfl0_c: 6882.9
fguv_32x32xn_16bpc_422_csfl0_ssse3: 959.6
fguv_32x32xn_16bpc_422_csfl0_avx2: 1259.1
fguv_32x32xn_16bpc_422_csfl1_c: 5402.1
fguv_32x32xn_16bpc_422_csfl1_ssse3: 733.5
fguv_32x32xn_16bpc_422_csfl1_avx2: 1127.8
fguv_32x32xn_16bpc_444_csfl0_c: 5982.6
fguv_32x32xn_16bpc_444_csfl0_ssse3: 839.1
fguv_32x32xn_16bpc_444_csfl0_avx2: 1091.1
fguv_32x32xn_16bpc_444_csfl1_c: 4610.3
fguv_32x32xn_16bpc_444_csfl1_ssse3: 598.3
fguv_32x32xn_16bpc_444_csfl1_avx2: 996.3
fgy_32x32xn_16bpc_c: 10831.6
fgy_32x32xn_16bpc_ssse3: 2329.3
fgy_32x32xn_16bpc_avx2: 3923.3
gen_grain_uv_ar0_16bpc_420_c: 35745.8
gen_grain_uv_ar0_16bpc_420_ssse3: 6705.8
gen_grain_uv_ar0_16bpc_420_avx2: 7156.5
gen_grain_uv_ar0_16bpc_422_c: 63680.5
gen_grain_uv_ar0_16bpc_422_ssse3: 12617.6
gen_grain_uv_ar0_16bpc_422_avx2: 13575.8
gen_grain_uv_ar0_16bpc_444_c: 117817.0
gen_grain_uv_ar0_16bpc_444_ssse3: 22231.5
gen_grain_uv_ar0_16bpc_444_avx2: 25344.1
gen_grain_uv_ar1_16bpc_420_c: 48458.4
gen_grain_uv_ar1_16bpc_420_ssse3: 18311.5
gen_grain_uv_ar1_16bpc_420_avx2: 19256.2
gen_grain_uv_ar1_16bpc_422_c: 88209.0
gen_grain_uv_ar1_16bpc_422_ssse3: 35909.8
gen_grain_uv_ar1_16bpc_422_avx2: 38474.5
gen_grain_uv_ar1_16bpc_444_c: 162797.9
gen_grain_uv_ar1_16bpc_444_ssse3: 72423.3
gen_grain_uv_ar1_16bpc_444_avx2: 77465.8
gen_grain_uv_ar2_16bpc_420_c: 68789.7
gen_grain_uv_ar2_16bpc_420_ssse3: 24026.4
gen_grain_uv_ar2_16bpc_420_avx2: 22795.5
gen_grain_uv_ar2_16bpc_422_c: 128814.9
gen_grain_uv_ar2_16bpc_422_ssse3: 47320.4
gen_grain_uv_ar2_16bpc_422_avx2: 45516.8
gen_grain_uv_ar2_16bpc_444_c: 245827.4
gen_grain_uv_ar2_16bpc_444_ssse3: 92946.7
gen_grain_uv_ar2_16bpc_444_avx2: 90109.3
gen_grain_uv_ar3_16bpc_420_c: 100092.0
gen_grain_uv_ar3_16bpc_420_ssse3: 25272.5
gen_grain_uv_ar3_16bpc_420_avx2: 24469.3
gen_grain_uv_ar3_16bpc_422_c: 189530.1
gen_grain_uv_ar3_16bpc_422_ssse3: 50015.8
gen_grain_uv_ar3_16bpc_422_avx2: 48981.4
gen_grain_uv_ar3_16bpc_444_c: 373360.0
gen_grain_uv_ar3_16bpc_444_ssse3: 98375.8
gen_grain_uv_ar3_16bpc_444_avx2: 95823.3
gen_grain_y_ar0_16bpc_c: 94838.8
gen_grain_y_ar0_16bpc_ssse3: 19753.2
gen_grain_y_ar0_16bpc_avx2: 22482.4
gen_grain_y_ar1_16bpc_c: 135236.3
gen_grain_y_ar1_16bpc_ssse3: 72171.0
gen_grain_y_ar1_16bpc_avx2: 75110.8
gen_grain_y_ar2_16bpc_c: 214918.9
gen_grain_y_ar2_16bpc_ssse3: 93905.4
gen_grain_y_ar2_16bpc_avx2: 86820.7
gen_grain_y_ar3_16bpc_c: 335112.8
gen_grain_y_ar3_16bpc_ssse3: 98511.1
gen_grain_y_ar3_16bpc_avx2: 94255.9