@scrypted/libav 1.0.181 → 1.0.183
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/_openh264install/include/wels/codec_api.h +0 -592
- package/_openh264install/include/wels/codec_app_def.h +0 -820
- package/_openh264install/include/wels/codec_def.h +0 -216
- package/_openh264install/include/wels/codec_ver.h +0 -15
- package/_openh264install/lib/libopenh264.a +0 -0
- package/_openh264install/lib/pkgconfig/openh264.pc +0 -10
- package/openh264/.reviewboardrc +0 -3
- package/openh264/.travis.yml +0 -42
- package/openh264/CODING_STYLE +0 -8
- package/openh264/CONTRIBUTORS +0 -55
- package/openh264/LICENSE +0 -23
- package/openh264/Makefile +0 -408
- package/openh264/README.md +0 -197
- package/openh264/RELEASES +0 -467
- package/openh264/SECURITY.md +0 -16
- package/openh264/autotest/codec_comparision/run_MultiCodecComparision.sh +0 -431
- package/openh264/autotest/performanceTest/android/run_AutoTest_android.sh +0 -180
- package/openh264/autotest/performanceTest/ios/fruitstrap +0 -0
- package/openh264/autotest/performanceTest/ios/iFileTransfer +0 -0
- package/openh264/autotest/performanceTest/ios/run_AutoTest_ios.sh +0 -193
- package/openh264/autotest/performanceTest/ios/uiascript.js +0 -19
- package/openh264/autotest/performanceTest/parsePerfData.sh +0 -107
- package/openh264/autotest/performanceTest/run_perfTest.sh +0 -122
- package/openh264/autotest/unitTest/android/run_AutoTest_android.sh +0 -112
- package/openh264/autotest/unitTest/ios/run_AutoTest_ios.sh +0 -167
- package/openh264/autotest/unitTest/run_ParseUTxml.sh +0 -70
- package/openh264/autotest/unitTest/run_unitTest.sh +0 -98
- package/openh264/build.gradle +0 -30
- package/openh264/code-coverage.sh +0 -6
- package/openh264/codec/api/meson.build +0 -1
- package/openh264/codec/api/wels/codec_api.h +0 -592
- package/openh264/codec/api/wels/codec_app_def.h +0 -820
- package/openh264/codec/api/wels/codec_def.h +0 -216
- package/openh264/codec/api/wels/codec_ver.h +0 -15
- package/openh264/codec/api/wels/meson.build +0 -9
- package/openh264/codec/common/arm/arm_arch_common_macro.S +0 -83
- package/openh264/codec/common/arm/copy_mb_neon.S +0 -161
- package/openh264/codec/common/arm/deblocking_neon.S +0 -857
- package/openh264/codec/common/arm/expand_picture_neon.S +0 -154
- package/openh264/codec/common/arm/intra_pred_common_neon.S +0 -82
- package/openh264/codec/common/arm/mc_neon.S +0 -2283
- package/openh264/codec/common/arm64/arm_aarch64_common.h +0 -85
- package/openh264/codec/common/arm64/arm_arch64_common_macro.S +0 -79
- package/openh264/codec/common/arm64/copy_mb_aarch64_neon.S +0 -202
- package/openh264/codec/common/arm64/deblocking_aarch64_neon.S +0 -852
- package/openh264/codec/common/arm64/expand_picture_aarch64_neon.S +0 -150
- package/openh264/codec/common/arm64/intra_pred_common_aarch64_neon.S +0 -58
- package/openh264/codec/common/arm64/mc_aarch64_neon.S +0 -2614
- package/openh264/codec/common/generate_version.sh +0 -28
- package/openh264/codec/common/inc/WelsList.h +0 -287
- package/openh264/codec/common/inc/WelsLock.h +0 -97
- package/openh264/codec/common/inc/WelsTask.h +0 -75
- package/openh264/codec/common/inc/WelsTaskThread.h +0 -83
- package/openh264/codec/common/inc/WelsThread.h +0 -106
- package/openh264/codec/common/inc/WelsThreadLib.h +0 -151
- package/openh264/codec/common/inc/WelsThreadPool.h +0 -124
- package/openh264/codec/common/inc/asmdefs_mmi.h +0 -345
- package/openh264/codec/common/inc/copy_mb.h +0 -103
- package/openh264/codec/common/inc/cpu.h +0 -81
- package/openh264/codec/common/inc/cpu_core.h +0 -103
- package/openh264/codec/common/inc/crt_util_safe_x.h +0 -104
- package/openh264/codec/common/inc/deblocking_common.h +0 -124
- package/openh264/codec/common/inc/expand_pic.h +0 -107
- package/openh264/codec/common/inc/golomb_common.h +0 -166
- package/openh264/codec/common/inc/intra_pred_common.h +0 -86
- package/openh264/codec/common/inc/loongson_intrinsics.h +0 -1919
- package/openh264/codec/common/inc/ls_defines.h +0 -130
- package/openh264/codec/common/inc/macros.h +0 -329
- package/openh264/codec/common/inc/mc.h +0 -407
- package/openh264/codec/common/inc/measure_time.h +0 -88
- package/openh264/codec/common/inc/memory_align.h +0 -116
- package/openh264/codec/common/inc/msa_macros.h +0 -2393
- package/openh264/codec/common/inc/sad_common.h +0 -139
- package/openh264/codec/common/inc/typedefs.h +0 -86
- package/openh264/codec/common/inc/utils.h +0 -95
- package/openh264/codec/common/inc/version.h +0 -10
- package/openh264/codec/common/inc/version_gen.h.template +0 -6
- package/openh264/codec/common/inc/welsCodecTrace.h +0 -64
- package/openh264/codec/common/inc/wels_common_defs.h +0 -373
- package/openh264/codec/common/inc/wels_const_common.h +0 -63
- package/openh264/codec/common/loongarch/copy_mb_lsx.c +0 -194
- package/openh264/codec/common/loongarch/deblock_lsx.c +0 -1489
- package/openh264/codec/common/loongarch/intra_pred_com_lasx.c +0 -112
- package/openh264/codec/common/loongarch/intra_pred_com_lsx.c +0 -91
- package/openh264/codec/common/loongarch/mc_chroma_lsx.c +0 -192
- package/openh264/codec/common/loongarch/mc_horver_lsx.c +0 -654
- package/openh264/codec/common/loongarch/satd_sad_lasx.c +0 -568
- package/openh264/codec/common/meson.build +0 -84
- package/openh264/codec/common/mips/copy_mb_mmi.c +0 -477
- package/openh264/codec/common/mips/copy_mb_msa.c +0 -80
- package/openh264/codec/common/mips/deblock_mmi.c +0 -2826
- package/openh264/codec/common/mips/deblock_msa.c +0 -1024
- package/openh264/codec/common/mips/expand_picture_mmi.c +0 -673
- package/openh264/codec/common/mips/intra_pred_com_mmi.c +0 -548
- package/openh264/codec/common/mips/satd_sad_mmi.c +0 -2154
- package/openh264/codec/common/targets.mk +0 -136
- package/openh264/codec/common/x86/asm_inc.asm +0 -743
- package/openh264/codec/common/x86/cpuid.asm +0 -263
- package/openh264/codec/common/x86/dct.asm +0 -1036
- package/openh264/codec/common/x86/deblock.asm +0 -848
- package/openh264/codec/common/x86/expand_picture.asm +0 -728
- package/openh264/codec/common/x86/intra_pred_com.asm +0 -117
- package/openh264/codec/common/x86/mb_copy.asm +0 -615
- package/openh264/codec/common/x86/mc_chroma.asm +0 -313
- package/openh264/codec/common/x86/mc_luma.asm +0 -4490
- package/openh264/codec/common/x86/satd_sad.asm +0 -2734
- package/openh264/codec/common/x86/vaa.asm +0 -411
- package/openh264/codec/console/common/inc/read_config.h +0 -68
- package/openh264/codec/console/common/meson.build +0 -12
- package/openh264/codec/console/common/targets.mk +0 -20
- package/openh264/codec/console/dec/inc/d3d9_utils.h +0 -148
- package/openh264/codec/console/dec/meson.build +0 -15
- package/openh264/codec/console/dec/targets.mk +0 -20
- package/openh264/codec/console/enc/meson.build +0 -9
- package/openh264/codec/console/enc/targets.mk +0 -19
- package/openh264/codec/console/meson.build +0 -3
- package/openh264/codec/decoder/core/arm/block_add_neon.S +0 -129
- package/openh264/codec/decoder/core/arm/intra_pred_neon.S +0 -635
- package/openh264/codec/decoder/core/arm64/block_add_aarch64_neon.S +0 -131
- package/openh264/codec/decoder/core/arm64/intra_pred_aarch64_neon.S +0 -524
- package/openh264/codec/decoder/core/inc/au_parser.h +0 -199
- package/openh264/codec/decoder/core/inc/bit_stream.h +0 -62
- package/openh264/codec/decoder/core/inc/cabac_decoder.h +0 -111
- package/openh264/codec/decoder/core/inc/deblocking.h +0 -118
- package/openh264/codec/decoder/core/inc/dec_frame.h +0 -156
- package/openh264/codec/decoder/core/inc/dec_golomb.h +0 -344
- package/openh264/codec/decoder/core/inc/decode_mb_aux.h +0 -83
- package/openh264/codec/decoder/core/inc/decode_slice.h +0 -104
- package/openh264/codec/decoder/core/inc/decoder.h +0 -186
- package/openh264/codec/decoder/core/inc/decoder_context.h +0 -590
- package/openh264/codec/decoder/core/inc/decoder_core.h +0 -206
- package/openh264/codec/decoder/core/inc/error_code.h +0 -222
- package/openh264/codec/decoder/core/inc/error_concealment.h +0 -68
- package/openh264/codec/decoder/core/inc/fmo.h +0 -114
- package/openh264/codec/decoder/core/inc/get_intra_predictor.h +0 -191
- package/openh264/codec/decoder/core/inc/manage_dec_ref.h +0 -62
- package/openh264/codec/decoder/core/inc/mb_cache.h +0 -80
- package/openh264/codec/decoder/core/inc/memmgr_nal_unit.h +0 -66
- package/openh264/codec/decoder/core/inc/mv_pred.h +0 -195
- package/openh264/codec/decoder/core/inc/nal_prefix.h +0 -56
- package/openh264/codec/decoder/core/inc/nalu.h +0 -79
- package/openh264/codec/decoder/core/inc/parameter_sets.h +0 -217
- package/openh264/codec/decoder/core/inc/parse_mb_syn_cabac.h +0 -89
- package/openh264/codec/decoder/core/inc/parse_mb_syn_cavlc.h +0 -137
- package/openh264/codec/decoder/core/inc/pic_queue.h +0 -62
- package/openh264/codec/decoder/core/inc/picture.h +0 -112
- package/openh264/codec/decoder/core/inc/rec_mb.h +0 -104
- package/openh264/codec/decoder/core/inc/slice.h +0 -208
- package/openh264/codec/decoder/core/inc/vlc_decoder.h +0 -177
- package/openh264/codec/decoder/core/inc/wels_common_basis.h +0 -317
- package/openh264/codec/decoder/core/inc/wels_const.h +0 -65
- package/openh264/codec/decoder/core/inc/wels_decoder_thread.h +0 -170
- package/openh264/codec/decoder/core/loongarch/mb_aux_lsx.c +0 -296
- package/openh264/codec/decoder/core/mips/dct_mmi.c +0 -786
- package/openh264/codec/decoder/core/x86/dct.asm +0 -72
- package/openh264/codec/decoder/core/x86/intra_pred.asm +0 -1456
- package/openh264/codec/decoder/meson.build +0 -69
- package/openh264/codec/decoder/plus/inc/welsDecoderExt.h +0 -170
- package/openh264/codec/decoder/targets.mk +0 -115
- package/openh264/codec/encoder/core/arm/intra_pred_neon.S +0 -589
- package/openh264/codec/encoder/core/arm/intra_pred_sad_3_opt_neon.S +0 -735
- package/openh264/codec/encoder/core/arm/memory_neon.S +0 -62
- package/openh264/codec/encoder/core/arm/pixel_neon.S +0 -882
- package/openh264/codec/encoder/core/arm/reconstruct_neon.S +0 -903
- package/openh264/codec/encoder/core/arm/svc_motion_estimation.S +0 -366
- package/openh264/codec/encoder/core/arm64/intra_pred_aarch64_neon.S +0 -505
- package/openh264/codec/encoder/core/arm64/intra_pred_sad_3_opt_aarch64_neon.S +0 -611
- package/openh264/codec/encoder/core/arm64/memory_aarch64_neon.S +0 -63
- package/openh264/codec/encoder/core/arm64/pixel_aarch64_neon.S +0 -621
- package/openh264/codec/encoder/core/arm64/reconstruct_aarch64_neon.S +0 -717
- package/openh264/codec/encoder/core/arm64/svc_motion_estimation_aarch64_neon.S +0 -350
- package/openh264/codec/encoder/core/inc/as264_common.h +0 -138
- package/openh264/codec/encoder/core/inc/au_set.h +0 -152
- package/openh264/codec/encoder/core/inc/deblocking.h +0 -92
- package/openh264/codec/encoder/core/inc/decode_mb_aux.h +0 -112
- package/openh264/codec/encoder/core/inc/dq_map.h +0 -58
- package/openh264/codec/encoder/core/inc/encode_mb_aux.h +0 -199
- package/openh264/codec/encoder/core/inc/encoder.h +0 -148
- package/openh264/codec/encoder/core/inc/encoder_context.h +0 -240
- package/openh264/codec/encoder/core/inc/extern.h +0 -124
- package/openh264/codec/encoder/core/inc/get_intra_predictor.h +0 -186
- package/openh264/codec/encoder/core/inc/mb_cache.h +0 -141
- package/openh264/codec/encoder/core/inc/md.h +0 -181
- package/openh264/codec/encoder/core/inc/mt_defs.h +0 -91
- package/openh264/codec/encoder/core/inc/mv_pred.h +0 -191
- package/openh264/codec/encoder/core/inc/nal_encap.h +0 -143
- package/openh264/codec/encoder/core/inc/param_svc.h +0 -576
- package/openh264/codec/encoder/core/inc/parameter_sets.h +0 -181
- package/openh264/codec/encoder/core/inc/paraset_strategy.h +0 -310
- package/openh264/codec/encoder/core/inc/picture.h +0 -137
- package/openh264/codec/encoder/core/inc/picture_handle.h +0 -66
- package/openh264/codec/encoder/core/inc/rc.h +0 -284
- package/openh264/codec/encoder/core/inc/ref_list_mgr_svc.h +0 -148
- package/openh264/codec/encoder/core/inc/sample.h +0 -154
- package/openh264/codec/encoder/core/inc/set_mb_syn_cabac.h +0 -127
- package/openh264/codec/encoder/core/inc/set_mb_syn_cavlc.h +0 -92
- package/openh264/codec/encoder/core/inc/slice.h +0 -212
- package/openh264/codec/encoder/core/inc/slice_multi_threading.h +0 -97
- package/openh264/codec/encoder/core/inc/stat.h +0 -102
- package/openh264/codec/encoder/core/inc/svc_base_layer_md.h +0 -102
- package/openh264/codec/encoder/core/inc/svc_enc_frame.h +0 -136
- package/openh264/codec/encoder/core/inc/svc_enc_golomb.h +0 -123
- package/openh264/codec/encoder/core/inc/svc_enc_macroblock.h +0 -82
- package/openh264/codec/encoder/core/inc/svc_enc_slice_segment.h +0 -212
- package/openh264/codec/encoder/core/inc/svc_encode_mb.h +0 -63
- package/openh264/codec/encoder/core/inc/svc_encode_slice.h +0 -184
- package/openh264/codec/encoder/core/inc/svc_mode_decision.h +0 -94
- package/openh264/codec/encoder/core/inc/svc_motion_estimate.h +0 -369
- package/openh264/codec/encoder/core/inc/svc_set_mb_syn.h +0 -70
- package/openh264/codec/encoder/core/inc/svc_set_mb_syn_cavlc.h +0 -67
- package/openh264/codec/encoder/core/inc/vlc_encoder.h +0 -92
- package/openh264/codec/encoder/core/inc/wels_common_basis.h +0 -160
- package/openh264/codec/encoder/core/inc/wels_const.h +0 -174
- package/openh264/codec/encoder/core/inc/wels_func_ptr_def.h +0 -300
- package/openh264/codec/encoder/core/inc/wels_preprocess.h +0 -251
- package/openh264/codec/encoder/core/inc/wels_task_base.h +0 -79
- package/openh264/codec/encoder/core/inc/wels_task_encoder.h +0 -142
- package/openh264/codec/encoder/core/inc/wels_task_management.h +0 -135
- package/openh264/codec/encoder/core/inc/wels_transpose_matrix.h +0 -58
- package/openh264/codec/encoder/core/loongarch/dct_lasx.c +0 -452
- package/openh264/codec/encoder/core/loongarch/get_intra_predictor_lsx.c +0 -166
- package/openh264/codec/encoder/core/loongarch/quant_lsx.c +0 -153
- package/openh264/codec/encoder/core/loongarch/sample_lasx.c +0 -87
- package/openh264/codec/encoder/core/loongarch/svc_motion_estimate_lsx.c +0 -104
- package/openh264/codec/encoder/core/mips/dct_mmi.c +0 -529
- package/openh264/codec/encoder/core/mips/quant_mmi.c +0 -553
- package/openh264/codec/encoder/core/mips/score_mmi.c +0 -324
- package/openh264/codec/encoder/core/x86/coeff.asm +0 -695
- package/openh264/codec/encoder/core/x86/dct.asm +0 -108
- package/openh264/codec/encoder/core/x86/intra_pred.asm +0 -1129
- package/openh264/codec/encoder/core/x86/matrix_transpose.asm +0 -395
- package/openh264/codec/encoder/core/x86/memzero.asm +0 -132
- package/openh264/codec/encoder/core/x86/quant.asm +0 -507
- package/openh264/codec/encoder/core/x86/sample_sc.asm +0 -1829
- package/openh264/codec/encoder/core/x86/score.asm +0 -363
- package/openh264/codec/encoder/meson.build +0 -98
- package/openh264/codec/encoder/plus/inc/welsEncoderExt.h +0 -129
- package/openh264/codec/encoder/targets.mk +0 -146
- package/openh264/codec/meson.build +0 -9
- package/openh264/codec/processing/interface/IWelsVP.h +0 -310
- package/openh264/codec/processing/meson.build +0 -71
- package/openh264/codec/processing/targets.mk +0 -117
- package/openh264/docs/doxygen/Doxyfile +0 -1865
- package/openh264/docs/doxygen/Home.rest +0 -25
- package/openh264/docs/doxygen/ISVCDecoder.rest +0 -19
- package/openh264/docs/doxygen/ISVCEncoder.rest +0 -16
- package/openh264/docs/doxygen/UsageExampleForDecoder.rest +0 -10
- package/openh264/docs/doxygen/UsageExampleForEncoder.rest +0 -15
- package/openh264/docs/doxygen2rst.py +0 -619
- package/openh264/gmpopenh264.info +0 -4
- package/openh264/gradlew +0 -172
- package/openh264/meson.build +0 -218
- package/openh264/meson_options.txt +0 -1
- package/openh264/module/RefCounted.h +0 -141
- package/openh264/module/gmp-openh264.cpp +0 -1196
- package/openh264/module/targets.mk +0 -14
- package/openh264/module/task_utils.h +0 -38
- package/openh264/module/task_utils.py +0 -207
- package/openh264/module/task_utils_generated.h +0 -2087
- package/openh264/openh264.def +0 -7
- package/openh264/openh264.pc.in +0 -10
- package/openh264/openh264.rc +0 -61
- package/openh264/openh264.rc.template +0 -61
- package/openh264/res/Adobe_PDF_sample_a_1024x768_50Frms.264 +0 -0
- package/openh264/res/BA1_FT_C.264 +0 -0
- package/openh264/res/BA1_Sony_D.jsv +0 -0
- package/openh264/res/BAMQ1_JVC_C.264 +0 -0
- package/openh264/res/BAMQ2_JVC_C.264 +0 -0
- package/openh264/res/BANM_MW_D.264 +0 -0
- package/openh264/res/BASQP1_Sony_C.jsv +0 -0
- package/openh264/res/BA_MW_D.264 +0 -0
- package/openh264/res/BA_MW_D_IDR_LOST.264 +0 -0
- package/openh264/res/BA_MW_D_P_LOST.264 +0 -0
- package/openh264/res/CI1_FT_B.264 +0 -0
- package/openh264/res/CI_MW_D.264 +0 -0
- package/openh264/res/CVFC1_Sony_C.jsv +0 -0
- package/openh264/res/CVPCMNL1_SVA_C.264 +0 -0
- package/openh264/res/CiscoVT2people_160x96_6fps.yuv +0 -1
- package/openh264/res/CiscoVT2people_320x192_12fps.yuv +0 -1
- package/openh264/res/Cisco_Absolute_Power_1280x720_30fps.yuv +0 -3607
- package/openh264/res/Cisco_Adobe_PDF_sample_a_1024x768_CAVLC_Bframe_9.264 +0 -0
- package/openh264/res/Cisco_Men_whisper_640x320_CABAC_Bframe_9.264 +0 -0
- package/openh264/res/Cisco_Men_whisper_640x320_CAVLC_Bframe_9.264 +0 -0
- package/openh264/res/Error_I_P.264 +0 -0
- package/openh264/res/LS_SVA_D.264 +0 -0
- package/openh264/res/MIDR_MW_D.264 +0 -0
- package/openh264/res/MPS_MW_A.264 +0 -0
- package/openh264/res/MR1_BT_A.h264 +0 -0
- package/openh264/res/MR1_MW_A.264 +0 -0
- package/openh264/res/MR2_MW_A.264 +0 -0
- package/openh264/res/MR2_TANDBERG_E.264 +0 -0
- package/openh264/res/NL1_Sony_D.jsv +0 -0
- package/openh264/res/NLMQ1_JVC_C.264 +0 -0
- package/openh264/res/NLMQ2_JVC_C.264 +0 -0
- package/openh264/res/NRF_MW_E.264 +0 -0
- package/openh264/res/QCIF_2P_I_allIPCM.264 +0 -0
- package/openh264/res/SVA_BA1_B.264 +0 -0
- package/openh264/res/SVA_BA2_D.264 +0 -0
- package/openh264/res/SVA_Base_B.264 +0 -0
- package/openh264/res/SVA_CL1_E.264 +0 -0
- package/openh264/res/SVA_FM1_E.264 +0 -0
- package/openh264/res/SVA_NL1_B.264 +0 -0
- package/openh264/res/SVA_NL2_E.264 +0 -0
- package/openh264/res/SarVui.264 +0 -0
- package/openh264/res/Static.264 +0 -0
- package/openh264/res/Static_152_100.yuv +0 -1
- package/openh264/res/VID_1280x544_cabac_temporal_direct.264 +0 -0
- package/openh264/res/VID_1280x544_cavlc_temporal_direct.264 +0 -0
- package/openh264/res/VID_1280x720_cabac_temporal_direct.264 +0 -0
- package/openh264/res/VID_1280x720_cavlc_temporal_direct.264 +0 -0
- package/openh264/res/VID_1920x1080_cabac_temporal_direct.264 +0 -0
- package/openh264/res/VID_1920x1080_cavlc_temporal_direct.264 +0 -0
- package/openh264/res/Zhling_1280x720.264 +0 -0
- package/openh264/res/jm_1080p_allslice.264 +0 -0
- package/openh264/res/sps_subsetsps_bothVUI.264 +0 -0
- package/openh264/res/test_cif_I_CABAC_PCM.264 +0 -0
- package/openh264/res/test_cif_I_CABAC_slice.264 +0 -0
- package/openh264/res/test_cif_P_CABAC_slice.264 +0 -0
- package/openh264/res/test_qcif_cabac.264 +0 -0
- package/openh264/res/test_scalinglist_jm.264 +0 -0
- package/openh264/res/test_vd_1d.264 +0 -0
- package/openh264/res/test_vd_rc.264 +0 -0
- package/openh264/run_Test.sh +0 -91
- package/openh264/settings.gradle +0 -4
- package/openh264/subprojects/gtest.wrap +0 -10
- package/openh264/testbin/AutoBuild_Windows_VS2008.bat +0 -59
- package/openh264/testbin/CmdLineExample.sh +0 -7
- package/openh264/testbin/CmdLineReadMe +0 -22
- package/openh264/testbin/layer2.cfg +0 -37
- package/openh264/testbin/layer2_arbitrary_res.cfg +0 -37
- package/openh264/testbin/layer2_vd.cfg +0 -38
- package/openh264/testbin/layer2_vd_rc.cfg +0 -38
- package/openh264/testbin/welsenc.cfg +0 -71
- package/openh264/testbin/welsenc_arbitrary_res.cfg +0 -70
- package/openh264/testbin/welsenc_ios.cfg +0 -71
- package/openh264/testbin/welsenc_vd_1d.cfg +0 -70
- package/openh264/testbin/welsenc_vd_rc.cfg +0 -69
- package/openh264/ut.def +0 -2
|
@@ -1,786 +0,0 @@
|
|
|
1
|
-
/*!
|
|
2
|
-
* \copy
|
|
3
|
-
* Copyright (c) 2009-2018, Cisco Systems
|
|
4
|
-
* All rights reserved.
|
|
5
|
-
*
|
|
6
|
-
* Redistribution and use in source and binary forms, with or without
|
|
7
|
-
* modification, are permitted provided that the following conditions
|
|
8
|
-
* are met:
|
|
9
|
-
*
|
|
10
|
-
* * Redistributions of source code must retain the above copyright
|
|
11
|
-
* notice, this list of conditions and the following disclaimer.
|
|
12
|
-
*
|
|
13
|
-
* * Redistributions in binary form must reproduce the above copyright
|
|
14
|
-
* notice, this list of conditions and the following disclaimer in
|
|
15
|
-
* the documentation and/or other materials provided with the
|
|
16
|
-
* distribution.
|
|
17
|
-
*
|
|
18
|
-
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
19
|
-
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
20
|
-
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
|
|
21
|
-
* FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
|
|
22
|
-
* COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
|
|
23
|
-
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
|
|
24
|
-
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
|
|
25
|
-
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
|
26
|
-
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
27
|
-
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
|
|
28
|
-
* ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
|
|
29
|
-
* POSSIBILITY OF SUCH DAMAGE.
|
|
30
|
-
*
|
|
31
|
-
*
|
|
32
|
-
* \file dct_mmi.c
|
|
33
|
-
*
|
|
34
|
-
* \brief Loongson optimization
|
|
35
|
-
*
|
|
36
|
-
* \date 17/07/2018 Created
|
|
37
|
-
*
|
|
38
|
-
*************************************************************************************
|
|
39
|
-
*/
|
|
40
|
-
#include <stdint.h>
|
|
41
|
-
#include "asmdefs_mmi.h"
|
|
42
|
-
|
|
43
|
-
#define LOAD_2_LEFT_AND_ADD \
|
|
44
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t" \
|
|
45
|
-
"lbu $9, -0x1(%[pPred]) \n\t" \
|
|
46
|
-
PTR_ADDU "$8, $8, $9 \n\t" \
|
|
47
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t" \
|
|
48
|
-
"lbu $9, -0x1(%[pPred]) \n\t" \
|
|
49
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
50
|
-
|
|
51
|
-
unsigned char mmi_dc_0x80[16] __attribute__((aligned(16))) = {
|
|
52
|
-
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
|
|
53
|
-
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
|
|
54
|
-
|
|
55
|
-
short mmi_wd_0x02[8] __attribute__((aligned(16))) = {2, 2, 2, 2, 2, 2, 2, 2};
|
|
56
|
-
short mmi_plane_inc_minus[8]__attribute__((aligned(16))) = {-7, -6, -5, -4, -3, -2, -1, 0};
|
|
57
|
-
short mmi_plane_inc[8]__attribute__((aligned(16))) = {1, 2, 3, 4, 5, 6, 7, 8};
|
|
58
|
-
short mmi_plane_dec[8]__attribute__((aligned(16))) = {8, 7, 6, 5, 4, 3, 2, 1};
|
|
59
|
-
|
|
60
|
-
short mmi_plane_inc_c[4]__attribute__((aligned(16))) = {1, 2, 3, 4};
|
|
61
|
-
short mmi_plane_dec_c[4]__attribute__((aligned(16))) = {4, 3, 2, 1};
|
|
62
|
-
short mmi_plane_mul_b_c[8]__attribute__((aligned(16))) = {-3, -2, -1, 0, 1, 2, 3, 4};
|
|
63
|
-
|
|
64
|
-
unsigned char mmi_01bytes[16]__attribute__((aligned(16))) = {1, 1, 1, 1, 1, 1, 1, 1,
|
|
65
|
-
1, 1, 1, 1, 1, 1, 1, 1};
|
|
66
|
-
|
|
67
|
-
void IdctResAddPred_mmi(uint8_t *pPred, const int32_t kiStride, int16_t *pRs) {
|
|
68
|
-
__asm__ volatile (
|
|
69
|
-
".set arch=loongson3a \n\t"
|
|
70
|
-
"dli $8, 0x1 \n\t"
|
|
71
|
-
"gsldxc1 $f0, 0x0(%[pRs], $0) \n\t"
|
|
72
|
-
"gsldxc1 $f2, 0x8(%[pRs], $0) \n\t"
|
|
73
|
-
"gsldxc1 $f4, 0x10(%[pRs], $0) \n\t"
|
|
74
|
-
"gsldxc1 $f6, 0x18(%[pRs], $0) \n\t"
|
|
75
|
-
"dmtc1 $8, $f14 \n\t"
|
|
76
|
-
|
|
77
|
-
MMI_Trans4x4H_SINGLE($f0, $f2, $f4, $f6, $f8)
|
|
78
|
-
MMI_IDCT_SINGLE($f2, $f4, $f6, $f8, $f0, $f12, $f14)
|
|
79
|
-
MMI_Trans4x4H_SINGLE($f2, $f6, $f0, $f8, $f4)
|
|
80
|
-
MMI_IDCT_SINGLE($f6, $f0, $f8, $f4, $f2, $f12, $f14)
|
|
81
|
-
|
|
82
|
-
"dli $8, 0x20 \n\t"
|
|
83
|
-
"xor $f14, $f14, $f14 \n\t"
|
|
84
|
-
"dmtc1 $8, $f12 \n\t"
|
|
85
|
-
"pshufh $f12, $f12, $f14 \n\t"
|
|
86
|
-
"dli $8, 0x6 \n\t"
|
|
87
|
-
"dmtc1 $8, $f16 \n\t"
|
|
88
|
-
|
|
89
|
-
MMI_StoreDiff4P_SINGLE($f6, $f0, $f12, $f14, %[pPred], %[pPred], $f16)
|
|
90
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
91
|
-
MMI_StoreDiff4P_SINGLE($f8, $f0, $f12, $f14, %[pPred], %[pPred], $f16)
|
|
92
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
93
|
-
MMI_StoreDiff4P_SINGLE($f2, $f0, $f12, $f14, %[pPred], %[pPred], $f16)
|
|
94
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
95
|
-
MMI_StoreDiff4P_SINGLE($f4, $f0, $f12, $f14, %[pPred], %[pPred], $f16)
|
|
96
|
-
: [pPred]"+&r"((unsigned char *)pPred)
|
|
97
|
-
: [pRs]"r"((unsigned char *)pRs), [kiStride]"r"((int)kiStride)
|
|
98
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6", "$f8", "$f10", "$f12",
|
|
99
|
-
"$f14", "$f16"
|
|
100
|
-
);
|
|
101
|
-
}
|
|
102
|
-
|
|
103
|
-
void WelsDecoderI16x16LumaPredDc_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
104
|
-
__asm__ volatile(
|
|
105
|
-
".set arch=loongson3a \n\t"
|
|
106
|
-
"dli $8, 0x5 \n\t"
|
|
107
|
-
"gsldxc1 $f10, 0x0(%[mmi_01bytes], $0) \n\t"
|
|
108
|
-
"dmtc1 $8, $f8 \n\t"
|
|
109
|
-
|
|
110
|
-
"move $10, %[pPred] \n\t"
|
|
111
|
-
PTR_SUBU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
112
|
-
"gslqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
113
|
-
"xor $f4, $f4, $f4 \n\t"
|
|
114
|
-
"pasubub $f0, $f0, $f4 \n\t"
|
|
115
|
-
"pasubub $f2, $f2, $f4 \n\t"
|
|
116
|
-
"biadd $f0, $f0 \n\t"
|
|
117
|
-
"biadd $f2, $f2 \n\t"
|
|
118
|
-
"paddh $f0, $f0, $f2 \n\t"
|
|
119
|
-
|
|
120
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
121
|
-
"lbu $8, -0x1(%[pPred]) \n\t"
|
|
122
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
123
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
124
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
125
|
-
|
|
126
|
-
LOAD_2_LEFT_AND_ADD
|
|
127
|
-
LOAD_2_LEFT_AND_ADD
|
|
128
|
-
LOAD_2_LEFT_AND_ADD
|
|
129
|
-
LOAD_2_LEFT_AND_ADD
|
|
130
|
-
LOAD_2_LEFT_AND_ADD
|
|
131
|
-
LOAD_2_LEFT_AND_ADD
|
|
132
|
-
LOAD_2_LEFT_AND_ADD
|
|
133
|
-
|
|
134
|
-
PTR_ADDIU "$8, $8, 0x10 \n\t"
|
|
135
|
-
"dmtc1 $8, $f4 \n\t"
|
|
136
|
-
"paddh $f0, $f0, $f4 \n\t"
|
|
137
|
-
"psrlw $f0, $f0, $f8 \n\t"
|
|
138
|
-
"pmuluw $f0, $f0, $f10 \n\t"
|
|
139
|
-
"punpcklwd $f0, $f0, $f0 \n\t"
|
|
140
|
-
"mov.d $f2, $f0 \n\t"
|
|
141
|
-
|
|
142
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
143
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
144
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
145
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
146
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
147
|
-
|
|
148
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
149
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
150
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
151
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
152
|
-
|
|
153
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
154
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
155
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
156
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
157
|
-
|
|
158
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
159
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
160
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
161
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
162
|
-
|
|
163
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
164
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
165
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
166
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
167
|
-
|
|
168
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
169
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
170
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
171
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
172
|
-
|
|
173
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
174
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
175
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
176
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
177
|
-
|
|
178
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
179
|
-
"gssqc1 $f2, $f0, 0x0($10) \n\t"
|
|
180
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
181
|
-
: [kiStride] "r"((int)kiStride),
|
|
182
|
-
[mmi_01bytes] "r"((unsigned char *)mmi_01bytes)
|
|
183
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6", "$f8", "$f10"
|
|
184
|
-
);
|
|
185
|
-
}
|
|
186
|
-
|
|
187
|
-
void WelsDecoderI16x16LumaPredPlane_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
188
|
-
BACKUP_REG;
|
|
189
|
-
__asm__ volatile (
|
|
190
|
-
".set arch=loongson3a \n\t"
|
|
191
|
-
"move $10, %[pPred] \n\t"
|
|
192
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x1 \n\t"
|
|
193
|
-
PTR_SUBU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
194
|
-
|
|
195
|
-
"gsldlc1 $f0, 0x7(%[pPred]) \n\t"
|
|
196
|
-
"xor $f28, $f28, $f28 \n\t"
|
|
197
|
-
"gsldrc1 $f0, 0x0(%[pPred]) \n\t"
|
|
198
|
-
"gslqc1 $f22, $f20, 0x0(%[mmi_plane_dec]) \n\t"
|
|
199
|
-
"punpckhbh $f2, $f0, $f28 \n\t"
|
|
200
|
-
"punpcklbh $f0, $f0, $f28 \n\t"
|
|
201
|
-
"pmullh $f0, $f0, $f20 \n\t"
|
|
202
|
-
"gsldlc1 $f4, 0x10(%[pPred]) \n\t"
|
|
203
|
-
"pmullh $f2, $f2, $f22 \n\t"
|
|
204
|
-
"gsldrc1 $f4, 0x9(%[pPred]) \n\t"
|
|
205
|
-
"gslqc1 $f26, $f24, 0x0(%[mmi_plane_inc]) \n\t"
|
|
206
|
-
"punpckhbh $f6, $f4, $f28 \n\t"
|
|
207
|
-
"punpcklbh $f4, $f4, $f28 \n\t"
|
|
208
|
-
"pmullh $f4, $f4, $f24 \n\t"
|
|
209
|
-
"pmullh $f6, $f6, $f26 \n\t"
|
|
210
|
-
"psubh $f4, $f4, $f0 \n\t"
|
|
211
|
-
"psubh $f6, $f6, $f2 \n\t"
|
|
212
|
-
|
|
213
|
-
SUMH_HORIZON($f4, $f6, $f0, $f2, $f8)
|
|
214
|
-
"dmfc1 $8, $f4 \n\t"
|
|
215
|
-
"seh $8, $8 \n\t"
|
|
216
|
-
"mul $8, $8, 0x5 \n\t"
|
|
217
|
-
PTR_ADDIU "$8, $8, 0x20 \n\t"
|
|
218
|
-
"sra $8, $8, 0x6 \n\t"
|
|
219
|
-
MMI_Copy8Times($f4, $f6, $f28, $8)
|
|
220
|
-
|
|
221
|
-
"lbu $9, 0x10(%[pPred]) \n\t"
|
|
222
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x3 \n\t"
|
|
223
|
-
LOAD_COLUMN($f0, $f2, $f8, $f10, $f12, $f14, $f16, $f18, %[pPred],
|
|
224
|
-
%[kiStride], $11)
|
|
225
|
-
|
|
226
|
-
PTR_ADDIU "%[pPred], %[pPred], 0x3 \n\t"
|
|
227
|
-
"dsll $11, %[kiStride], 0x3 \n\t"
|
|
228
|
-
PTR_ADDU "$11, $11, %[pPred] \n\t"
|
|
229
|
-
"lbu $8, 0x0($11) \n\t"
|
|
230
|
-
PTR_ADDU "$9, $9, $8 \n\t"
|
|
231
|
-
"dsll $9, $9, 0x4 \n\t"
|
|
232
|
-
|
|
233
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x3 \n\t"
|
|
234
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
235
|
-
LOAD_COLUMN($f28, $f30, $f8, $f10, $f12, $f14, $f16, $f18, %[pPred],
|
|
236
|
-
%[kiStride], $11)
|
|
237
|
-
|
|
238
|
-
"xor $f16, $f16, $f16 \n\t"
|
|
239
|
-
"punpcklbh $f0, $f2, $f16 \n\t"
|
|
240
|
-
"punpckhbh $f2, $f2, $f16 \n\t"
|
|
241
|
-
"pmullh $f0, $f0, $f20 \n\t"
|
|
242
|
-
"pmullh $f2, $f2, $f22 \n\t"
|
|
243
|
-
"punpcklbh $f28, $f30, $f16 \n\t"
|
|
244
|
-
"punpckhbh $f30, $f30, $f16 \n\t"
|
|
245
|
-
"pmullh $f28, $f28, $f24 \n\t"
|
|
246
|
-
"pmullh $f30, $f30, $f26 \n\t"
|
|
247
|
-
"psubh $f28, $f28, $f0 \n\t"
|
|
248
|
-
"psubh $f30, $f30, $f2 \n\t"
|
|
249
|
-
|
|
250
|
-
"xor $f8, $f8, $f8 \n\t"
|
|
251
|
-
|
|
252
|
-
SUMH_HORIZON($f28, $f30, $f0, $f2, $f8)
|
|
253
|
-
"dmfc1 $8, $f28 \n\t"
|
|
254
|
-
"seh $8, $8 \n\t"
|
|
255
|
-
|
|
256
|
-
"mul $8, $8, 0x5 \n\t"
|
|
257
|
-
PTR_ADDIU "$8, $8, 0x20 \n\t"
|
|
258
|
-
"sra $8, $8, 0x6 \n\t"
|
|
259
|
-
MMI_Copy8Times($f16, $f18, $f8, $8)
|
|
260
|
-
|
|
261
|
-
"move %[pPred], $10 \n\t"
|
|
262
|
-
PTR_ADDIU "$9, $9, 0x10 \n\t"
|
|
263
|
-
"mul $8, $8, -0x7 \n\t"
|
|
264
|
-
PTR_ADDU "$9, $9, $8 \n\t"
|
|
265
|
-
MMI_Copy8Times($f0, $f2, $f8, $9)
|
|
266
|
-
|
|
267
|
-
"xor $8, $8, $8 \n\t"
|
|
268
|
-
"gslqc1 $f22, $f20, 0x0(%[mmi_plane_inc_minus]) \n\t"
|
|
269
|
-
|
|
270
|
-
"dli $11, 0x5 \n\t"
|
|
271
|
-
"dmtc1 $11, $f30 \n\t"
|
|
272
|
-
"1: \n\t"
|
|
273
|
-
"pmullh $f8, $f4, $f20 \n\t"
|
|
274
|
-
"pmullh $f10, $f6, $f22 \n\t"
|
|
275
|
-
"paddh $f8, $f8, $f0 \n\t"
|
|
276
|
-
"paddh $f10, $f10, $f2 \n\t"
|
|
277
|
-
"psrah $f8, $f8, $f30 \n\t"
|
|
278
|
-
"psrah $f10, $f10, $f30 \n\t"
|
|
279
|
-
"pmullh $f12, $f4, $f24 \n\t"
|
|
280
|
-
"pmullh $f14, $f6, $f26 \n\t"
|
|
281
|
-
"paddh $f12, $f12, $f0 \n\t"
|
|
282
|
-
"paddh $f14, $f14, $f2 \n\t"
|
|
283
|
-
"psrah $f12, $f12, $f30 \n\t"
|
|
284
|
-
"psrah $f14, $f14, $f30 \n\t"
|
|
285
|
-
"packushb $f8, $f8, $f10 \n\t"
|
|
286
|
-
"packushb $f10, $f12, $f14 \n\t"
|
|
287
|
-
"gssqc1 $f10, $f8, 0x0(%[pPred]) \n\t"
|
|
288
|
-
"paddh $f0, $f0, $f16 \n\t"
|
|
289
|
-
"paddh $f2, $f2, $f18 \n\t"
|
|
290
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
291
|
-
PTR_ADDIU "$8, $8, 0x1 \n\t"
|
|
292
|
-
PTR_ADDIU "$11, $8, -0x10 \n\t"
|
|
293
|
-
"bnez $11, 1b \n\t"
|
|
294
|
-
"nop \n\t"
|
|
295
|
-
: [pPred]"+&r"((unsigned char *)pPred)
|
|
296
|
-
: [kiStride]"r"((int)kiStride), [mmi_plane_inc_minus]"r"(mmi_plane_inc_minus),
|
|
297
|
-
[mmi_plane_inc]"r"(mmi_plane_inc), [mmi_plane_dec]"r"(mmi_plane_dec)
|
|
298
|
-
: "memory", "$8", "$9", "$10", "$11", "$f0", "$f2", "$f4", "$f6", "$f8",
|
|
299
|
-
"$f10", "$f12", "$f14", "$f16", "$f18", "$f20", "$f22", "$f24", "$f26",
|
|
300
|
-
"$f28", "$f30"
|
|
301
|
-
);
|
|
302
|
-
RECOVER_REG;
|
|
303
|
-
}
|
|
304
|
-
|
|
305
|
-
#define COPY_16_TIMES(r0, f0, f2, f4, f6, f8) \
|
|
306
|
-
"gslqc1 "#f2", "#f0", -0x10("#r0") \n\t" \
|
|
307
|
-
"dsrl "#f0", "#f2", "#f4" \n\t" \
|
|
308
|
-
"pmuluw "#f0", "#f0", "#f6" \n\t" \
|
|
309
|
-
"punpcklwd "#f0", "#f0", "#f0" \n\t" \
|
|
310
|
-
"mov.d "#f2", "#f0" \n\t"
|
|
311
|
-
|
|
312
|
-
#define MMI_PRED_H_16X16_TWO_LINE_DEC \
|
|
313
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t" \
|
|
314
|
-
COPY_16_TIMES(%[pPred], $f0, $f2, $f4, $f6, $f8) \
|
|
315
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t" \
|
|
316
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t" \
|
|
317
|
-
COPY_16_TIMES(%[pPred], $f0, $f2, $f4, $f6, $f8) \
|
|
318
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
319
|
-
|
|
320
|
-
void WelsDecoderI16x16LumaPredH_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
321
|
-
__asm__ volatile (
|
|
322
|
-
".set arch=loongson3a \n\t"
|
|
323
|
-
"dli $8, 56 \n\t"
|
|
324
|
-
"dmtc1 $8, $f4 \n\t"
|
|
325
|
-
"gsldxc1 $f6, 0x0(%[mmi_01bytes], $0) \n\t"
|
|
326
|
-
"xor $f8, $f8, $f8 \n\t"
|
|
327
|
-
|
|
328
|
-
COPY_16_TIMES(%[pPred], $f0, $f2, $f4, $f6, $f8)
|
|
329
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
330
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
331
|
-
COPY_16_TIMES(%[pPred], $f0, $f2, $f4, $f6, $f8)
|
|
332
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
333
|
-
|
|
334
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
335
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
336
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
337
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
338
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
339
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
340
|
-
MMI_PRED_H_16X16_TWO_LINE_DEC
|
|
341
|
-
: [pPred]"+&r"((unsigned char *)pPred)
|
|
342
|
-
: [kiStride]"r"((int)kiStride),
|
|
343
|
-
[mmi_01bytes]"r"((unsigned char *)mmi_01bytes)
|
|
344
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6", "$f8"
|
|
345
|
-
);
|
|
346
|
-
}
|
|
347
|
-
|
|
348
|
-
void WelsDecoderI16x16LumaPredV_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
349
|
-
__asm__ volatile(
|
|
350
|
-
".set arch=loongson3a \n\t"
|
|
351
|
-
PTR_SUBU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
352
|
-
"gslqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
353
|
-
|
|
354
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
355
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
356
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
357
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
358
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
359
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
360
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
361
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
362
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
363
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
364
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
365
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
366
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
367
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
368
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
369
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
370
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
371
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
372
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
373
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
374
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
375
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
376
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
377
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
378
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
379
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
380
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
381
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
382
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
383
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
384
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
385
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
386
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
387
|
-
: [kiStride] "r"((int)kiStride)
|
|
388
|
-
: "memory", "$f0", "$f2"
|
|
389
|
-
);
|
|
390
|
-
}
|
|
391
|
-
|
|
392
|
-
void WelsDecoderI16x16LumaPredDcTop_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
393
|
-
BACKUP_REG;
|
|
394
|
-
__asm__ volatile (
|
|
395
|
-
".set arch=loongson3a \n\t"
|
|
396
|
-
PTR_SUBU "$8, %[pPred], %[kiStride] \n\t"
|
|
397
|
-
"gslqc1 $f2, $f0, 0x0($8) \n\t"
|
|
398
|
-
"xor $f28, $f28, $f28 \n\t"
|
|
399
|
-
"pasubub $f0, $f0, $f28 \n\t"
|
|
400
|
-
"pasubub $f2, $f2, $f28 \n\t"
|
|
401
|
-
"biadd $f0, $f0 \n\t"
|
|
402
|
-
"biadd $f2, $f2 \n\t"
|
|
403
|
-
"paddh $f0, $f0, $f2 \n\t"
|
|
404
|
-
"dmfc1 $8, $f0 \n\t"
|
|
405
|
-
|
|
406
|
-
PTR_ADDIU "$8, $8, 0x8 \n\t"
|
|
407
|
-
"dsra $8, $8, 0x4 \n\t"
|
|
408
|
-
MMI_Copy16Times($f4, $f6, $f28, $8)
|
|
409
|
-
"mov.d $f0, $f4 \n\t"
|
|
410
|
-
"mov.d $f2, $f6 \n\t"
|
|
411
|
-
|
|
412
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
413
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
414
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
415
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
416
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
417
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
418
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
419
|
-
|
|
420
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
421
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
422
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
423
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
424
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
425
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
426
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
427
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
428
|
-
|
|
429
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
430
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
431
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
432
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
433
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
434
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
435
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
436
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
437
|
-
|
|
438
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
439
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
440
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
441
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
442
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
443
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
444
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
445
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
446
|
-
: [pPred]"+&r"((unsigned char *)pPred)
|
|
447
|
-
: [kiStride]"r"((int)kiStride)
|
|
448
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6"
|
|
449
|
-
);
|
|
450
|
-
RECOVER_REG;
|
|
451
|
-
}
|
|
452
|
-
|
|
453
|
-
void WelsDecoderI16x16LumaPredDcNA_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
454
|
-
__asm__ volatile(
|
|
455
|
-
".set arch=loongson3a \n\t"
|
|
456
|
-
"gslqc1 $f2, $f0, 0x0(%[mmi_dc_0x80]) \n\t"
|
|
457
|
-
"mov.d $f4, $f0 \n\t"
|
|
458
|
-
"mov.d $f6, $f2 \n\t"
|
|
459
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
460
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
461
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
462
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
463
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
464
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
465
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
466
|
-
|
|
467
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
468
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
469
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
470
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
471
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
472
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
473
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
474
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
475
|
-
|
|
476
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
477
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
478
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
479
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
480
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
481
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
482
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
483
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
484
|
-
|
|
485
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
486
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
487
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
488
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
489
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
490
|
-
"gssqc1 $f2, $f0, 0x0(%[pPred]) \n\t"
|
|
491
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
492
|
-
"gssqc1 $f6, $f4, 0x0(%[pPred]) \n\t"
|
|
493
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
494
|
-
: [kiStride] "r"((int)kiStride), [mmi_dc_0x80] "r"(mmi_dc_0x80)
|
|
495
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6"
|
|
496
|
-
);
|
|
497
|
-
}
|
|
498
|
-
|
|
499
|
-
void WelsDecoderIChromaPredPlane_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
500
|
-
BACKUP_REG;
|
|
501
|
-
__asm__ volatile (
|
|
502
|
-
".set arch=loongson3a \n\t"
|
|
503
|
-
"move $10, %[pPred] \n\t"
|
|
504
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x1 \n\t"
|
|
505
|
-
PTR_SUBU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
506
|
-
|
|
507
|
-
"gsldlc1 $f0, 0x7(%[pPred]) \n\t"
|
|
508
|
-
"xor $f28, $f28, $f28 \n\t"
|
|
509
|
-
"gsldrc1 $f0, 0x0(%[pPred]) \n\t"
|
|
510
|
-
"gsldxc1 $f20, 0x0(%[mmi_plane_dec_c], $0) \n\t"
|
|
511
|
-
"punpcklbh $f0, $f0, $f28 \n\t"
|
|
512
|
-
"gsldlc1 $f4, 0xc(%[pPred]) \n\t"
|
|
513
|
-
"pmullh $f0, $f0, $f20 \n\t"
|
|
514
|
-
"gsldrc1 $f4, 0x5(%[pPred]) \n\t"
|
|
515
|
-
"gsldxc1 $f24, 0x0(%[mmi_plane_inc_c], $0) \n\t"
|
|
516
|
-
"punpcklbh $f4, $f4, $f28 \n\t"
|
|
517
|
-
"pmullh $f4, $f4, $f24 \n\t"
|
|
518
|
-
"psubh $f4, $f4, $f0 \n\t"
|
|
519
|
-
|
|
520
|
-
"xor $f6, $f6, $f6 \n\t"
|
|
521
|
-
"xor $f8, $f8, $f8 \n\t"
|
|
522
|
-
SUMH_HORIZON($f4, $f6, $f0, $f2, $f8)
|
|
523
|
-
"dmfc1 $8, $f4 \n\t"
|
|
524
|
-
"seh $8, $8 \n\t"
|
|
525
|
-
"mul $8, $8, 0x11 \n\t"
|
|
526
|
-
PTR_ADDIU "$8, $8, 0x10 \n\t"
|
|
527
|
-
"sra $8, $8, 0x5 \n\t"
|
|
528
|
-
MMI_Copy8Times($f4, $f6, $f8, $8)
|
|
529
|
-
|
|
530
|
-
"lbu $9, 0x8(%[pPred]) \n\t"
|
|
531
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x3 \n\t"
|
|
532
|
-
LOAD_COLUMN_C($f0, $f8, $f12, $f16, %[pPred], %[kiStride], $11)
|
|
533
|
-
|
|
534
|
-
PTR_ADDIU "%[pPred], %[pPred], 0x3 \n\t"
|
|
535
|
-
"dsll $11, %[kiStride], 0x2 \n\t"
|
|
536
|
-
PTR_ADDU "$11, $11, %[pPred] \n\t"
|
|
537
|
-
"lbu $8, 0x0($11) \n\t"
|
|
538
|
-
PTR_ADDU "$9, $9, $8 \n\t"
|
|
539
|
-
"dsll $9, $9, 0x4 \n\t"
|
|
540
|
-
|
|
541
|
-
PTR_ADDIU "%[pPred], %[pPred], -0x3 \n\t"
|
|
542
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
543
|
-
LOAD_COLUMN_C($f28, $f8, $f12, $f16, %[pPred], %[kiStride], $11)
|
|
544
|
-
"xor $f16, $f16, $f16 \n\t"
|
|
545
|
-
"punpckhbh $f0, $f0, $f16 \n\t"
|
|
546
|
-
"pmullh $f0, $f0, $f20 \n\t"
|
|
547
|
-
"punpckhbh $f28, $f28, $f16 \n\t"
|
|
548
|
-
"pmullh $f28, $f28, $f24 \n\t"
|
|
549
|
-
"psubh $f28, $f28, $f0 \n\t"
|
|
550
|
-
|
|
551
|
-
"xor $f30, $f30, $f30 \n\t"
|
|
552
|
-
"xor $f8, $f8, $f8 \n\t"
|
|
553
|
-
SUMH_HORIZON($f28, $f30, $f0, $f2, $f8)
|
|
554
|
-
"dmfc1 $8, $f28 \n\t"
|
|
555
|
-
"seh $8, $8 \n\t"
|
|
556
|
-
|
|
557
|
-
"mul $8, $8, 0x11 \n\t"
|
|
558
|
-
PTR_ADDIU "$8, $8, 0x10 \n\t"
|
|
559
|
-
"sra $8, $8, 0x5 \n\t"
|
|
560
|
-
MMI_Copy8Times($f16, $f18, $f8, $8)
|
|
561
|
-
|
|
562
|
-
"move %[pPred], $10 \n\t"
|
|
563
|
-
PTR_ADDIU "$9, $9, 0x10 \n\t"
|
|
564
|
-
"mul $8, $8, -0x3 \n\t"
|
|
565
|
-
PTR_ADDU "$9, $9, $8 \n\t"
|
|
566
|
-
MMI_Copy8Times($f0, $f2, $f8, $9)
|
|
567
|
-
|
|
568
|
-
"xor $8, $8, $8 \n\t"
|
|
569
|
-
"gslqc1 $f22, $f20, 0x0(%[mmi_plane_mul_b_c]) \n\t"
|
|
570
|
-
|
|
571
|
-
"dli $11, 0x5 \n\t"
|
|
572
|
-
"dmtc1 $11, $f30 \n\t"
|
|
573
|
-
"1: \n\t"
|
|
574
|
-
"pmullh $f8, $f4, $f20 \n\t"
|
|
575
|
-
"pmullh $f10, $f6, $f22 \n\t"
|
|
576
|
-
"paddh $f8, $f8, $f0 \n\t"
|
|
577
|
-
"paddh $f10, $f10, $f2 \n\t"
|
|
578
|
-
"psrah $f8, $f8, $f30 \n\t"
|
|
579
|
-
"psrah $f10, $f10, $f30 \n\t"
|
|
580
|
-
"packushb $f8, $f8, $f10 \n\t"
|
|
581
|
-
"gssdxc1 $f8, 0x0(%[pPred], $0) \n\t"
|
|
582
|
-
"paddh $f0, $f0, $f16 \n\t"
|
|
583
|
-
"paddh $f2, $f2, $f18 \n\t"
|
|
584
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
585
|
-
PTR_ADDIU "$8, $8, 0x1 \n\t"
|
|
586
|
-
PTR_ADDIU "$11, $8, -0x8 \n\t"
|
|
587
|
-
"bnez $11, 1b \n\t"
|
|
588
|
-
"nop \n\t"
|
|
589
|
-
: [pPred]"+&r"((unsigned char *)pPred)
|
|
590
|
-
: [kiStride]"r"((int)kiStride), [mmi_plane_mul_b_c]"r"(mmi_plane_mul_b_c),
|
|
591
|
-
[mmi_plane_inc_c]"r"(mmi_plane_inc_c), [mmi_plane_dec_c]"r"(mmi_plane_dec_c)
|
|
592
|
-
: "memory", "$8", "$9", "$10", "$11", "$f0", "$f2", "$f4", "$f6", "$f8", "$f10",
|
|
593
|
-
"$f12", "$f14", "$f16", "$f18", "$f20", "$f22", "$f24", "$f26", "$f28", "$f30"
|
|
594
|
-
);
|
|
595
|
-
RECOVER_REG;
|
|
596
|
-
}
|
|
597
|
-
|
|
598
|
-
void WelsDecoderIChromaPredDc_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
599
|
-
__asm__ volatile(
|
|
600
|
-
".set arch=loongson3a \n\t"
|
|
601
|
-
"move $10, %[pPred] \n\t"
|
|
602
|
-
|
|
603
|
-
PTR_SUBU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
604
|
-
"gsldxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
605
|
-
|
|
606
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
607
|
-
"lbu $8, -0x1(%[pPred]) \n\t"
|
|
608
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
609
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
610
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
611
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
612
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
613
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
614
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
615
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
616
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
617
|
-
"dmtc1 $8, $f2 \n\t"
|
|
618
|
-
|
|
619
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
620
|
-
"lbu $8, -0x1(%[pPred]) \n\t"
|
|
621
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
622
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
623
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
624
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
625
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
626
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
627
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
628
|
-
"lbu $9, -0x1(%[pPred]) \n\t"
|
|
629
|
-
PTR_ADDU "$8, $8, $9 \n\t"
|
|
630
|
-
"dmtc1 $8, $f4 \n\t"
|
|
631
|
-
|
|
632
|
-
"xor $f8, $f8, $f8 \n\t"
|
|
633
|
-
"punpcklwd $f6, $f0, $f8 \n\t"
|
|
634
|
-
"punpckhwd $f0, $f0, $f8 \n\t"
|
|
635
|
-
"pasubub $f0, $f0, $f8 \n\t"
|
|
636
|
-
"pasubub $f6, $f6, $f8 \n\t"
|
|
637
|
-
"biadd $f0, $f0 \n\t"
|
|
638
|
-
"biadd $f6, $f6 \n\t"
|
|
639
|
-
|
|
640
|
-
"paddd $f6, $f6, $f2 \n\t"
|
|
641
|
-
"paddd $f2, $f4, $f0 \n\t"
|
|
642
|
-
|
|
643
|
-
"dli $8, 0x2 \n\t"
|
|
644
|
-
"dmtc1 $8, $f8 \n\t"
|
|
645
|
-
"gsldxc1 $f12, 0x0(%[mmi_01bytes], $0) \n\t"
|
|
646
|
-
"dli $8, 0x3 \n\t"
|
|
647
|
-
"dmtc1 $8, $f10 \n\t"
|
|
648
|
-
|
|
649
|
-
"paddd $f0, $f0, $f8 \n\t"
|
|
650
|
-
"dsrl $f0, $f0, $f8 \n\t"
|
|
651
|
-
|
|
652
|
-
"paddd $f4, $f4, $f8 \n\t"
|
|
653
|
-
"dsrl $f4, $f4, $f8 \n\t"
|
|
654
|
-
|
|
655
|
-
"paddd $f6, $f6, $f8 \n\t"
|
|
656
|
-
"paddd $f6, $f6, $f8 \n\t"
|
|
657
|
-
"dsrl $f6, $f6, $f10 \n\t"
|
|
658
|
-
|
|
659
|
-
"paddd $f2, $f2, $f8 \n\t"
|
|
660
|
-
"paddd $f2, $f2, $f8 \n\t"
|
|
661
|
-
"dsrl $f2, $f2, $f10 \n\t"
|
|
662
|
-
|
|
663
|
-
"dli $8, 0x20 \n\t"
|
|
664
|
-
"dmtc1 $8, $f8 \n\t"
|
|
665
|
-
"pmuluw $f0, $f0, $f12 \n\t"
|
|
666
|
-
"pmuluw $f6, $f6, $f12 \n\t"
|
|
667
|
-
"dsll $f0, $f0, $f8 \n\t"
|
|
668
|
-
"xor $f0, $f0, $f6 \n\t"
|
|
669
|
-
|
|
670
|
-
"pmuluw $f4, $f4, $f12 \n\t"
|
|
671
|
-
"pmuluw $f2, $f2, $f12 \n\t"
|
|
672
|
-
"dsll $f2, $f2, $f8 \n\t"
|
|
673
|
-
"xor $f2, $f2, $f4 \n\t"
|
|
674
|
-
|
|
675
|
-
"gssdxc1 $f0, 0x0($10, $0) \n\t"
|
|
676
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
677
|
-
"gssdxc1 $f0, 0x0($10, $0) \n\t"
|
|
678
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
679
|
-
"gssdxc1 $f0, 0x0($10, $0) \n\t"
|
|
680
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
681
|
-
"gssdxc1 $f0, 0x0($10, $0) \n\t"
|
|
682
|
-
|
|
683
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
684
|
-
"gssdxc1 $f2, 0x0($10, $0) \n\t"
|
|
685
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
686
|
-
"gssdxc1 $f2, 0x0($10, $0) \n\t"
|
|
687
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
688
|
-
"gssdxc1 $f2, 0x0($10, $0) \n\t"
|
|
689
|
-
PTR_ADDU "$10, $10, %[kiStride] \n\t"
|
|
690
|
-
"gssdxc1 $f2, 0x0($10, $0) \n\t"
|
|
691
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
692
|
-
: [kiStride] "r"((int)kiStride),
|
|
693
|
-
[mmi_01bytes] "r"((unsigned char *)mmi_01bytes)
|
|
694
|
-
: "memory", "$8", "$9", "$10", "$f0", "$f2", "$f4", "$f6", "$f8", "$f10",
|
|
695
|
-
"$f12"
|
|
696
|
-
);
|
|
697
|
-
}
|
|
698
|
-
|
|
699
|
-
void WelsDecoderIChromaPredDcTop_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
700
|
-
BACKUP_REG;
|
|
701
|
-
__asm__ volatile (
|
|
702
|
-
".set arch=loongson3a \n\t"
|
|
703
|
-
"dli $8, 0x4e \n\t"
|
|
704
|
-
"dmtc1 $8, $f16 \n\t"
|
|
705
|
-
"dli $8, 0xb1 \n\t"
|
|
706
|
-
"dmtc1 $8, $f18 \n\t"
|
|
707
|
-
"dli $8, 0x2 \n\t"
|
|
708
|
-
"dmtc1 $8, $f20 \n\t"
|
|
709
|
-
PTR_SUBU "$8, %[pPred], %[kiStride] \n\t"
|
|
710
|
-
"gsldxc1 $f0, 0x0($8, $0) \n\t"
|
|
711
|
-
"xor $f28, $f28, $f28 \n\t"
|
|
712
|
-
"punpckhbh $f2, $f0, $f28 \n\t"
|
|
713
|
-
"punpcklbh $f0, $f0, $f28 \n\t"
|
|
714
|
-
"pshufh $f4, $f0, $f16 \n\t"
|
|
715
|
-
"pshufh $f6, $f2, $f16 \n\t"
|
|
716
|
-
"paddh $f0, $f0, $f4 \n\t"
|
|
717
|
-
"paddh $f2, $f2, $f6 \n\t"
|
|
718
|
-
|
|
719
|
-
"pshufh $f8, $f0, $f18 \n\t"
|
|
720
|
-
"pshufh $f14, $f2, $f18 \n\t"
|
|
721
|
-
"paddh $f2, $f2, $f14 \n\t"
|
|
722
|
-
"paddh $f0, $f0, $f8 \n\t"
|
|
723
|
-
|
|
724
|
-
"gslqc1 $f26, $f24, 0x0(%[mmi_wd_0x02]) \n\t"
|
|
725
|
-
"paddh $f0, $f0, $f24 \n\t"
|
|
726
|
-
"paddh $f2, $f2, $f26 \n\t"
|
|
727
|
-
"psrah $f0, $f0, $f20 \n\t"
|
|
728
|
-
"psrah $f2, $f2, $f20 \n\t"
|
|
729
|
-
"packushb $f0, $f0, $f2 \n\t"
|
|
730
|
-
|
|
731
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
732
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
733
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
734
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
735
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
736
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
737
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
738
|
-
|
|
739
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
740
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
741
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
742
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
743
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
744
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
745
|
-
PTR_ADDU "%[pPred], %[pPred], %[kiStride] \n\t"
|
|
746
|
-
"gssdxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
747
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
748
|
-
: [kiStride] "r"((int)kiStride), [mmi_wd_0x02] "r"((short *)mmi_wd_0x02)
|
|
749
|
-
: "memory", "$8", "$f0", "$f2", "$f4", "$f6"
|
|
750
|
-
);
|
|
751
|
-
RECOVER_REG;
|
|
752
|
-
}
|
|
753
|
-
|
|
754
|
-
void WelsDecoderI4x4LumaPredH_mmi(uint8_t *pPred, const int32_t kiStride) {
|
|
755
|
-
__asm__ volatile(
|
|
756
|
-
".set arch=loongson3a \n\t"
|
|
757
|
-
"gsldxc1 $f8, 0x0(%[mmi_01bytes], $0) \n\t"
|
|
758
|
-
"lbu $8, -0x1(%[pPred]) \n\t"
|
|
759
|
-
"dmtc1 $8, $f0 \n\t"
|
|
760
|
-
"pmuluw $f0, $f0, $f8 \n\t"
|
|
761
|
-
|
|
762
|
-
PTR_ADDU "$9, %[pPred], %[kiStride] \n\t"
|
|
763
|
-
"lbu $8, -0x1($9) \n\t"
|
|
764
|
-
"dmtc1 $8, $f2 \n\t"
|
|
765
|
-
"pmuluw $f2, $f2, $f8 \n\t"
|
|
766
|
-
|
|
767
|
-
PTR_ADDU "$10, $9, %[kiStride] \n\t"
|
|
768
|
-
"lbu $8, -0x1($10) \n\t"
|
|
769
|
-
"dmtc1 $8, $f4 \n\t"
|
|
770
|
-
"pmuluw $f4, $f4, $f8 \n\t"
|
|
771
|
-
|
|
772
|
-
PTR_ADDU "$11, $10, %[kiStride] \n\t"
|
|
773
|
-
"lbu $8, -0x1($11) \n\t"
|
|
774
|
-
"dmtc1 $8, $f6 \n\t"
|
|
775
|
-
"pmuluw $f6, $f6, $f8 \n\t"
|
|
776
|
-
|
|
777
|
-
"gsswxc1 $f0, 0x0(%[pPred], $0) \n\t"
|
|
778
|
-
"gsswxc1 $f2, 0x0($9, $0) \n\t"
|
|
779
|
-
"gsswxc1 $f4, 0x0($10, $0) \n\t"
|
|
780
|
-
"gsswxc1 $f6, 0x0($11, $0) \n\t"
|
|
781
|
-
: [pPred] "+&r"((unsigned char *)pPred)
|
|
782
|
-
: [kiStride] "r"((int)kiStride),
|
|
783
|
-
[mmi_01bytes] "r"((unsigned char *)mmi_01bytes)
|
|
784
|
-
: "memory", "$8", "$9", "$10", "$11", "$f0", "$f2", "$f4", "$f6", "$f8"
|
|
785
|
-
);
|
|
786
|
-
}
|