Warning, /graphics/krita/3rdparty/ext_heif/0003-arm-asm-primitives.patch is written in an unsupported language. File is not indexed.
0001 Fixes "arm assembly fail to compile on 1.8" 0002 0003 Downloaded from upstream bug report: 0004 https://bitbucket.org/multicoreware/x265/issues/406 0005 0006 Signed-off-by: Bernd Kuhls <bernd.kuhls@t-online.de> 0007 0008 --- ./source/common/arm/asm-primitives.cpp.orig 2018-05-21 02:33:10.000000000 -0600 0009 +++ ./source/common/arm/asm-primitives.cpp 2018-05-28 20:38:37.302378303 -0600 0010 @@ -48,77 +48,77 @@ void setupAssemblyPrimitives(EncoderPrim 0011 p.ssim_4x4x2_core = PFX(ssim_4x4x2_core_neon); 0012 0013 // addAvg 0014 - p.pu[LUMA_4x4].addAvg = PFX(addAvg_4x4_neon); 0015 - p.pu[LUMA_4x8].addAvg = PFX(addAvg_4x8_neon); 0016 - p.pu[LUMA_4x16].addAvg = PFX(addAvg_4x16_neon); 0017 - p.pu[LUMA_8x4].addAvg = PFX(addAvg_8x4_neon); 0018 - p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_neon); 0019 - p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_neon); 0020 - p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_neon); 0021 - p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_neon); 0022 - p.pu[LUMA_16x4].addAvg = PFX(addAvg_16x4_neon); 0023 - p.pu[LUMA_16x8].addAvg = PFX(addAvg_16x8_neon); 0024 - p.pu[LUMA_16x12].addAvg = PFX(addAvg_16x12_neon); 0025 - p.pu[LUMA_16x16].addAvg = PFX(addAvg_16x16_neon); 0026 - p.pu[LUMA_16x32].addAvg = PFX(addAvg_16x32_neon); 0027 - p.pu[LUMA_16x64].addAvg = PFX(addAvg_16x64_neon); 0028 - p.pu[LUMA_24x32].addAvg = PFX(addAvg_24x32_neon); 0029 - p.pu[LUMA_32x8].addAvg = PFX(addAvg_32x8_neon); 0030 - p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_neon); 0031 - p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_neon); 0032 - p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_neon); 0033 - p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_neon); 0034 - p.pu[LUMA_48x64].addAvg = PFX(addAvg_48x64_neon); 0035 - p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_neon); 0036 - p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_neon); 0037 - p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_neon); 0038 - p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_neon); 0039 + p.pu[LUMA_4x4].addAvg[ALIGNED] = PFX(addAvg_4x4_neon); 0040 + p.pu[LUMA_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon); 0041 + p.pu[LUMA_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon); 0042 + p.pu[LUMA_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon); 0043 + p.pu[LUMA_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon); 0044 + p.pu[LUMA_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon); 0045 + p.pu[LUMA_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon); 0046 + p.pu[LUMA_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon); 0047 + p.pu[LUMA_16x4].addAvg[ALIGNED] = PFX(addAvg_16x4_neon); 0048 + p.pu[LUMA_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon); 0049 + p.pu[LUMA_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon); 0050 + p.pu[LUMA_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon); 0051 + p.pu[LUMA_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon); 0052 + p.pu[LUMA_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon); 0053 + p.pu[LUMA_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon); 0054 + p.pu[LUMA_32x8].addAvg[ALIGNED] = PFX(addAvg_32x8_neon); 0055 + p.pu[LUMA_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon); 0056 + p.pu[LUMA_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon); 0057 + p.pu[LUMA_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon); 0058 + p.pu[LUMA_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon); 0059 + p.pu[LUMA_48x64].addAvg[ALIGNED] = PFX(addAvg_48x64_neon); 0060 + p.pu[LUMA_64x16].addAvg[ALIGNED] = PFX(addAvg_64x16_neon); 0061 + p.pu[LUMA_64x32].addAvg[ALIGNED] = PFX(addAvg_64x32_neon); 0062 + p.pu[LUMA_64x48].addAvg[ALIGNED] = PFX(addAvg_64x48_neon); 0063 + p.pu[LUMA_64x64].addAvg[ALIGNED] = PFX(addAvg_64x64_neon); 0064 0065 // chroma addAvg 0066 - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg = PFX(addAvg_4x2_neon); 0067 - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg = PFX(addAvg_4x4_neon); 0068 - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg = PFX(addAvg_4x8_neon); 0069 - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg = PFX(addAvg_4x16_neon); 0070 - p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg = PFX(addAvg_6x8_neon); 0071 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg = PFX(addAvg_8x2_neon); 0072 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_neon); 0073 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg = PFX(addAvg_8x6_neon); 0074 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg = PFX(addAvg_8x8_neon); 0075 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg = PFX(addAvg_8x16_neon); 0076 - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg = PFX(addAvg_8x32_neon); 0077 - p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_neon); 0078 - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg = PFX(addAvg_16x4_neon); 0079 - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg = PFX(addAvg_16x8_neon); 0080 - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg = PFX(addAvg_16x12_neon); 0081 - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg = PFX(addAvg_16x16_neon); 0082 - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg = PFX(addAvg_16x32_neon); 0083 - p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg = PFX(addAvg_24x32_neon); 0084 - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg = PFX(addAvg_32x8_neon); 0085 - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_neon); 0086 - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg = PFX(addAvg_32x24_neon); 0087 - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg = PFX(addAvg_32x32_neon); 0088 - 0089 - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg = PFX(addAvg_4x8_neon); 0090 - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg = PFX(addAvg_4x16_neon); 0091 - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg = PFX(addAvg_4x32_neon); 0092 - p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg = PFX(addAvg_6x16_neon); 0093 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg = PFX(addAvg_8x4_neon); 0094 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg = PFX(addAvg_8x8_neon); 0095 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg = PFX(addAvg_8x12_neon); 0096 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg = PFX(addAvg_8x16_neon); 0097 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg = PFX(addAvg_8x32_neon); 0098 - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg = PFX(addAvg_8x64_neon); 0099 - p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_neon); 0100 - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg = PFX(addAvg_16x8_neon); 0101 - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg = PFX(addAvg_16x16_neon); 0102 - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg = PFX(addAvg_16x24_neon); 0103 - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg = PFX(addAvg_16x32_neon); 0104 - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg = PFX(addAvg_16x64_neon); 0105 - p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg = PFX(addAvg_24x64_neon); 0106 - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_neon); 0107 - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_neon); 0108 - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_neon); 0109 - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_neon); 0110 + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg[ALIGNED] = PFX(addAvg_4x2_neon); 0111 + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg[ALIGNED] = PFX(addAvg_4x4_neon); 0112 + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon); 0113 + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon); 0114 + p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg[ALIGNED] = PFX(addAvg_6x8_neon); 0115 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg[ALIGNED] = PFX(addAvg_8x2_neon); 0116 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon); 0117 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg[ALIGNED] = PFX(addAvg_8x6_neon); 0118 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon); 0119 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon); 0120 + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon); 0121 + p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon); 0122 + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg[ALIGNED] = PFX(addAvg_16x4_neon); 0123 + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon); 0124 + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon); 0125 + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon); 0126 + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon); 0127 + p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon); 0128 + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg[ALIGNED] = PFX(addAvg_32x8_neon); 0129 + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon); 0130 + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon); 0131 + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon); 0132 + 0133 + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon); 0134 + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon); 0135 + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg[ALIGNED] = PFX(addAvg_4x32_neon); 0136 + p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg[ALIGNED] = PFX(addAvg_6x16_neon); 0137 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon); 0138 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon); 0139 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg[ALIGNED] = PFX(addAvg_8x12_neon); 0140 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon); 0141 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon); 0142 + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg[ALIGNED] = PFX(addAvg_8x64_neon); 0143 + p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg[ALIGNED] = PFX(addAvg_12x32_neon); 0144 + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon); 0145 + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon); 0146 + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg[ALIGNED] = PFX(addAvg_16x24_neon); 0147 + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon); 0148 + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon); 0149 + p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg[ALIGNED] = PFX(addAvg_24x64_neon); 0150 + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon); 0151 + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon); 0152 + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg[ALIGNED] = PFX(addAvg_32x48_neon); 0153 + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon); 0154 0155 // quant 0156 p.quant = PFX(quant_neon); 0157 @@ -402,7 +402,7 @@ void setupAssemblyPrimitives(EncoderPrim 0158 p.scale2D_64to32 = PFX(scale2D_64to32_neon); 0159 0160 // scale1D_128to64 0161 - p.scale1D_128to64 = PFX(scale1D_128to64_neon); 0162 + p.scale1D_128to64[ALIGNED] = PFX(scale1D_128to64_neon); 0163 0164 // copy_count 0165 p.cu[BLOCK_4x4].copy_cnt = PFX(copy_cnt_4_neon); 0166 @@ -411,37 +411,37 @@ void setupAssemblyPrimitives(EncoderPrim 0167 p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_neon); 0168 0169 // filterPixelToShort 0170 - p.pu[LUMA_4x4].convert_p2s = PFX(filterPixelToShort_4x4_neon); 0171 - p.pu[LUMA_4x8].convert_p2s = PFX(filterPixelToShort_4x8_neon); 0172 - p.pu[LUMA_4x16].convert_p2s = PFX(filterPixelToShort_4x16_neon); 0173 - p.pu[LUMA_8x4].convert_p2s = PFX(filterPixelToShort_8x4_neon); 0174 - p.pu[LUMA_8x8].convert_p2s = PFX(filterPixelToShort_8x8_neon); 0175 - p.pu[LUMA_8x16].convert_p2s = PFX(filterPixelToShort_8x16_neon); 0176 - p.pu[LUMA_8x32].convert_p2s = PFX(filterPixelToShort_8x32_neon); 0177 - p.pu[LUMA_12x16].convert_p2s = PFX(filterPixelToShort_12x16_neon); 0178 - p.pu[LUMA_16x4].convert_p2s = PFX(filterPixelToShort_16x4_neon); 0179 - p.pu[LUMA_16x8].convert_p2s = PFX(filterPixelToShort_16x8_neon); 0180 - p.pu[LUMA_16x12].convert_p2s = PFX(filterPixelToShort_16x12_neon); 0181 - p.pu[LUMA_16x16].convert_p2s = PFX(filterPixelToShort_16x16_neon); 0182 - p.pu[LUMA_16x32].convert_p2s = PFX(filterPixelToShort_16x32_neon); 0183 - p.pu[LUMA_16x64].convert_p2s = PFX(filterPixelToShort_16x64_neon); 0184 - p.pu[LUMA_24x32].convert_p2s = PFX(filterPixelToShort_24x32_neon); 0185 - p.pu[LUMA_32x8].convert_p2s = PFX(filterPixelToShort_32x8_neon); 0186 - p.pu[LUMA_32x16].convert_p2s = PFX(filterPixelToShort_32x16_neon); 0187 - p.pu[LUMA_32x24].convert_p2s = PFX(filterPixelToShort_32x24_neon); 0188 - p.pu[LUMA_32x32].convert_p2s = PFX(filterPixelToShort_32x32_neon); 0189 - p.pu[LUMA_32x64].convert_p2s = PFX(filterPixelToShort_32x64_neon); 0190 - p.pu[LUMA_48x64].convert_p2s = PFX(filterPixelToShort_48x64_neon); 0191 - p.pu[LUMA_64x16].convert_p2s = PFX(filterPixelToShort_64x16_neon); 0192 - p.pu[LUMA_64x32].convert_p2s = PFX(filterPixelToShort_64x32_neon); 0193 - p.pu[LUMA_64x48].convert_p2s = PFX(filterPixelToShort_64x48_neon); 0194 - p.pu[LUMA_64x64].convert_p2s = PFX(filterPixelToShort_64x64_neon); 0195 + p.pu[LUMA_4x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x4_neon); 0196 + p.pu[LUMA_4x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x8_neon); 0197 + p.pu[LUMA_4x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x16_neon); 0198 + p.pu[LUMA_8x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x4_neon); 0199 + p.pu[LUMA_8x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x8_neon); 0200 + p.pu[LUMA_8x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x16_neon); 0201 + p.pu[LUMA_8x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x32_neon); 0202 + p.pu[LUMA_12x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_12x16_neon); 0203 + p.pu[LUMA_16x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x4_neon); 0204 + p.pu[LUMA_16x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x8_neon); 0205 + p.pu[LUMA_16x12].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x12_neon); 0206 + p.pu[LUMA_16x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x16_neon); 0207 + p.pu[LUMA_16x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x32_neon); 0208 + p.pu[LUMA_16x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x64_neon); 0209 + p.pu[LUMA_24x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_24x32_neon); 0210 + p.pu[LUMA_32x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x8_neon); 0211 + p.pu[LUMA_32x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x16_neon); 0212 + p.pu[LUMA_32x24].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x24_neon); 0213 + p.pu[LUMA_32x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x32_neon); 0214 + p.pu[LUMA_32x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x64_neon); 0215 + p.pu[LUMA_48x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_48x64_neon); 0216 + p.pu[LUMA_64x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x16_neon); 0217 + p.pu[LUMA_64x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x32_neon); 0218 + p.pu[LUMA_64x48].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x48_neon); 0219 + p.pu[LUMA_64x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x64_neon); 0220 0221 // Block_fill 0222 - p.cu[BLOCK_4x4].blockfill_s = PFX(blockfill_s_4x4_neon); 0223 - p.cu[BLOCK_8x8].blockfill_s = PFX(blockfill_s_8x8_neon); 0224 - p.cu[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_neon); 0225 - p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_neon); 0226 + p.cu[BLOCK_4x4].blockfill_s[ALIGNED] = PFX(blockfill_s_4x4_neon); 0227 + p.cu[BLOCK_8x8].blockfill_s[ALIGNED] = PFX(blockfill_s_8x8_neon); 0228 + p.cu[BLOCK_16x16].blockfill_s[ALIGNED] = PFX(blockfill_s_16x16_neon); 0229 + p.cu[BLOCK_32x32].blockfill_s[ALIGNED] = PFX(blockfill_s_32x32_neon); 0230 0231 // Blockcopy_ss 0232 p.cu[BLOCK_4x4].copy_ss = PFX(blockcopy_ss_4x4_neon); 0233 @@ -495,21 +495,21 @@ void setupAssemblyPrimitives(EncoderPrim 0234 p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_neon); 0235 0236 // pixel_add_ps 0237 - p.cu[BLOCK_4x4].add_ps = PFX(pixel_add_ps_4x4_neon); 0238 - p.cu[BLOCK_8x8].add_ps = PFX(pixel_add_ps_8x8_neon); 0239 - p.cu[BLOCK_16x16].add_ps = PFX(pixel_add_ps_16x16_neon); 0240 - p.cu[BLOCK_32x32].add_ps = PFX(pixel_add_ps_32x32_neon); 0241 - p.cu[BLOCK_64x64].add_ps = PFX(pixel_add_ps_64x64_neon); 0242 + p.cu[BLOCK_4x4].add_ps[ALIGNED] = PFX(pixel_add_ps_4x4_neon); 0243 + p.cu[BLOCK_8x8].add_ps[ALIGNED] = PFX(pixel_add_ps_8x8_neon); 0244 + p.cu[BLOCK_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon); 0245 + p.cu[BLOCK_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon); 0246 + p.cu[BLOCK_64x64].add_ps[ALIGNED] = PFX(pixel_add_ps_64x64_neon); 0247 0248 // chroma add_ps 0249 - p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps = PFX(pixel_add_ps_4x4_neon); 0250 - p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps = PFX(pixel_add_ps_8x8_neon); 0251 - p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps = PFX(pixel_add_ps_16x16_neon); 0252 - p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps = PFX(pixel_add_ps_32x32_neon); 0253 - p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps = PFX(pixel_add_ps_4x8_neon); 0254 - p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps = PFX(pixel_add_ps_8x16_neon); 0255 - p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps = PFX(pixel_add_ps_16x32_neon); 0256 - p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps = PFX(pixel_add_ps_32x64_neon); 0257 + p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps[ALIGNED] = PFX(pixel_add_ps_4x4_neon); 0258 + p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps[ALIGNED] = PFX(pixel_add_ps_8x8_neon); 0259 + p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon); 0260 + p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon); 0261 + p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps[ALIGNED] = PFX(pixel_add_ps_4x8_neon); 0262 + p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps[ALIGNED] = PFX(pixel_add_ps_8x16_neon); 0263 + p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps[ALIGNED] = PFX(pixel_add_ps_16x32_neon); 0264 + p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps[ALIGNED] = PFX(pixel_add_ps_32x64_neon); 0265 0266 // cpy2Dto1D_shr 0267 p.cu[BLOCK_4x4].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_4x4_neon); 0268 @@ -518,10 +518,10 @@ void setupAssemblyPrimitives(EncoderPrim 0269 p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon); 0270 0271 // ssd_s 0272 - p.cu[BLOCK_4x4].ssd_s = PFX(pixel_ssd_s_4x4_neon); 0273 - p.cu[BLOCK_8x8].ssd_s = PFX(pixel_ssd_s_8x8_neon); 0274 - p.cu[BLOCK_16x16].ssd_s = PFX(pixel_ssd_s_16x16_neon); 0275 - p.cu[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32x32_neon); 0276 + p.cu[BLOCK_4x4].ssd_s[ALIGNED] = PFX(pixel_ssd_s_4x4_neon); 0277 + p.cu[BLOCK_8x8].ssd_s[ALIGNED] = PFX(pixel_ssd_s_8x8_neon); 0278 + p.cu[BLOCK_16x16].ssd_s[ALIGNED] = PFX(pixel_ssd_s_16x16_neon); 0279 + p.cu[BLOCK_32x32].ssd_s[ALIGNED] = PFX(pixel_ssd_s_32x32_neon); 0280 0281 // sse_ss 0282 p.cu[BLOCK_4x4].sse_ss = PFX(pixel_sse_ss_4x4_neon); 0283 @@ -548,10 +548,10 @@ void setupAssemblyPrimitives(EncoderPrim 0284 p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_neon); 0285 0286 // calc_Residual 0287 - p.cu[BLOCK_4x4].calcresidual = PFX(getResidual4_neon); 0288 - p.cu[BLOCK_8x8].calcresidual = PFX(getResidual8_neon); 0289 - p.cu[BLOCK_16x16].calcresidual = PFX(getResidual16_neon); 0290 - p.cu[BLOCK_32x32].calcresidual = PFX(getResidual32_neon); 0291 + p.cu[BLOCK_4x4].calcresidual[ALIGNED] = PFX(getResidual4_neon); 0292 + p.cu[BLOCK_8x8].calcresidual[ALIGNED] = PFX(getResidual8_neon); 0293 + p.cu[BLOCK_16x16].calcresidual[ALIGNED] = PFX(getResidual16_neon); 0294 + p.cu[BLOCK_32x32].calcresidual[ALIGNED] = PFX(getResidual32_neon); 0295 0296 // sse_pp 0297 p.cu[BLOCK_4x4].sse_pp = PFX(pixel_sse_pp_4x4_neon); 0298 @@ -722,31 +722,31 @@ void setupAssemblyPrimitives(EncoderPrim 0299 p.pu[LUMA_64x64].sad_x4 = PFX(sad_x4_64x64_neon); 0300 0301 // pixel_avg_pp 0302 - p.pu[LUMA_4x4].pixelavg_pp = PFX(pixel_avg_pp_4x4_neon); 0303 - p.pu[LUMA_4x8].pixelavg_pp = PFX(pixel_avg_pp_4x8_neon); 0304 - p.pu[LUMA_4x16].pixelavg_pp = PFX(pixel_avg_pp_4x16_neon); 0305 - p.pu[LUMA_8x4].pixelavg_pp = PFX(pixel_avg_pp_8x4_neon); 0306 - p.pu[LUMA_8x8].pixelavg_pp = PFX(pixel_avg_pp_8x8_neon); 0307 - p.pu[LUMA_8x16].pixelavg_pp = PFX(pixel_avg_pp_8x16_neon); 0308 - p.pu[LUMA_8x32].pixelavg_pp = PFX(pixel_avg_pp_8x32_neon); 0309 - p.pu[LUMA_12x16].pixelavg_pp = PFX(pixel_avg_pp_12x16_neon); 0310 - p.pu[LUMA_16x4].pixelavg_pp = PFX(pixel_avg_pp_16x4_neon); 0311 - p.pu[LUMA_16x8].pixelavg_pp = PFX(pixel_avg_pp_16x8_neon); 0312 - p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_pp_16x12_neon); 0313 - p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_pp_16x16_neon); 0314 - p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_pp_16x32_neon); 0315 - p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_pp_16x64_neon); 0316 - p.pu[LUMA_24x32].pixelavg_pp = PFX(pixel_avg_pp_24x32_neon); 0317 - p.pu[LUMA_32x8].pixelavg_pp = PFX(pixel_avg_pp_32x8_neon); 0318 - p.pu[LUMA_32x16].pixelavg_pp = PFX(pixel_avg_pp_32x16_neon); 0319 - p.pu[LUMA_32x24].pixelavg_pp = PFX(pixel_avg_pp_32x24_neon); 0320 - p.pu[LUMA_32x32].pixelavg_pp = PFX(pixel_avg_pp_32x32_neon); 0321 - p.pu[LUMA_32x64].pixelavg_pp = PFX(pixel_avg_pp_32x64_neon); 0322 - p.pu[LUMA_48x64].pixelavg_pp = PFX(pixel_avg_pp_48x64_neon); 0323 - p.pu[LUMA_64x16].pixelavg_pp = PFX(pixel_avg_pp_64x16_neon); 0324 - p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_pp_64x32_neon); 0325 - p.pu[LUMA_64x48].pixelavg_pp = PFX(pixel_avg_pp_64x48_neon); 0326 - p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_pp_64x64_neon); 0327 + p.pu[LUMA_4x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x4_neon); 0328 + p.pu[LUMA_4x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x8_neon); 0329 + p.pu[LUMA_4x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x16_neon); 0330 + p.pu[LUMA_8x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x4_neon); 0331 + p.pu[LUMA_8x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x8_neon); 0332 + p.pu[LUMA_8x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x16_neon); 0333 + p.pu[LUMA_8x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x32_neon); 0334 + p.pu[LUMA_12x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_12x16_neon); 0335 + p.pu[LUMA_16x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x4_neon); 0336 + p.pu[LUMA_16x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x8_neon); 0337 + p.pu[LUMA_16x12].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x12_neon); 0338 + p.pu[LUMA_16x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x16_neon); 0339 + p.pu[LUMA_16x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x32_neon); 0340 + p.pu[LUMA_16x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x64_neon); 0341 + p.pu[LUMA_24x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_24x32_neon); 0342 + p.pu[LUMA_32x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x8_neon); 0343 + p.pu[LUMA_32x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x16_neon); 0344 + p.pu[LUMA_32x24].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x24_neon); 0345 + p.pu[LUMA_32x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x32_neon); 0346 + p.pu[LUMA_32x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x64_neon); 0347 + p.pu[LUMA_48x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_48x64_neon); 0348 + p.pu[LUMA_64x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x16_neon); 0349 + p.pu[LUMA_64x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x32_neon); 0350 + p.pu[LUMA_64x48].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x48_neon); 0351 + p.pu[LUMA_64x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x64_neon); 0352 0353 // planecopy 0354 p.planecopy_cp = PFX(pixel_planecopy_cp_neon);