/external/libvpx/libvpx/vp9/encoder/mips/msa/ |
H A D | vp9_fdct8x8_msa.c | 18 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 20 LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7); 22 SLLI_4V(in4, in5, in6, in7, 2); 26 VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, 27 in5, in6, in7); 28 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, 29 in3, in4, in5, in6, in7); 30 VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, 31 in5, in6, in7); 34 VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in [all...] |
H A D | vp9_fdct_msa.h | 18 #define VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \ 35 ILVRL_H2_SH(in0, in7, vec1_m, vec0_m); \ 38 cnst2_m, cnst3_m, in7, in0, in4, in3); \ 52 BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5); \
|
H A D | vp9_fdct16x16_msa.c | 368 v8i16 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11; local 371 LD_SH8(temp, 16, in0, in1, in2, in3, in4, in5, in6, in7); 374 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 375 in4, in5, in6, in7); 381 FDCT_POSTPROC_2V_NEG_H(in6, in7); 386 BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, 396 in4, in5, in6, in7); 400 TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, tmp4, in4, 401 tmp5, in5, tmp6, in6, tmp7, in7); 403 ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, ou [all...] |
/external/libvpx/libvpx/vp9/common/mips/msa/ |
H A D | vp9_idct8x8_msa.c | 18 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 21 LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7); 23 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 24 in4, in5, in6, in7); 29 VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 30 in4, in5, in6, in7); 32 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, 33 in3, in4, in5, in6, in7); 34 VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 35 in4, in5, in6, in7); [all...] |
/external/libvpx/libvpx/vpx_dsp/mips/ |
H A D | idct8x8_msa.c | 15 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 18 LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7); 21 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 22 in4, in5, in6, in7); 24 VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 25 in4, in5, in6, in7); 27 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 28 in4, in5, in6, in7); 30 VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 31 in4, in5, in6, in7); 43 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
H A D | fwd_txfm_msa.c | 16 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 29 LD_SH16(input, src_stride, in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, 32 SLLI_4V(in4, in5, in6, in7, 2); 36 ADD4(in4, in11, in5, in10, in6, in9, in7, in8, tmp4, tmp5, tmp6, tmp7); 41 SUB4(in4, in11, in5, in10, in6, in9, in7, in8, in11, in10, in9, in8); 135 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 138 LD_SH8(input, 16, in0, in1, in2, in3, in4, in5, in6, in7); 140 TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 141 in4, in5, in6, in7); 145 ADD4(in4, 1, in5, 1, in6, 1, in7, 201 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 219 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 249 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
H A D | macros_msa.h | 421 #define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \ 424 ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride); \ 447 #define ST_H8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \ 450 ST_H4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride); \ 614 #define AVER_UB4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 618 AVER_UB2(RTYPE, in4, in5, in6, in7, out2, out3) \ 1130 #define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 1134 ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ 1187 #define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 1191 ILVR_B2(RTYPE, in4, in5, in6, in7, out [all...] |
H A D | fwd_dct32x32_msa.c | 16 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 23 LD_SH4(input + (28 * src_stride), src_stride, in4, in5, in6, in7); 27 SLLI_4V(in4, in5, in6, in7, 2); 30 BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, step0, step1, step2, 31 step3, in4, in5, in6, in7); 35 ST_SH4(in4, in5, in6, in7, temp_buff + (28 * 8), 8); 41 LD_SH4(input + (20 * src_stride), src_stride, in4, in5, in6, in7); 45 SLLI_4V(in4, in5, in6, in7, 2); 48 BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, step0, step1, step2, 49 step3, in4, in5, in6, in7); 59 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 255 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 288 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 392 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 596 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 695 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local 931 v8i16 in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
H A D | inv_txfm_msa.h | 18 #define VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \ 35 ILVRL_H2_SH(in0, in7, vec1_m, vec0_m); \ 38 cnst2_m, cnst3_m, in7, in0, in4, in3); \ 52 BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5); \ 214 #define VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 227 VP9_MADD(in1, in7, in3, in5, k0_m, k1_m, k2_m, k3_m, in1, in7, in3, in5); \ 228 SUB2(in1, in3, in7, in5, res0_m, res1_m); \ 238 tp7_m = in7 + in5; \ 247 #define VP9_IADST8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, out [all...] |
H A D | fwd_txfm_msa.h | 46 #define SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7) \ 51 SRLI_H4_SH(in4, in5, in6, in7, vec4_m, vec5_m, vec6_m, vec7_m, 15); \ 54 AVE_SH4_SH(vec4_m, in4, vec5_m, in5, vec6_m, in6, vec7_m, in7, in4, in5, \ 55 in6, in7); \ 58 #define VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \ 67 BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, s0_m, s1_m, s2_m, \ 118 #define FDCT8x16_EVEN(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 127 BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, s0_m, s1_m, s2_m, \
|
H A D | txfm_macros_msa.h | 40 #define DOT_ADD_SUB_SRARI_PCK(in0, in1, in2, in3, in4, in5, in6, in7, dst0, \ 48 DOTP_SH4_SW(in2, in3, in2, in3, in6, in6, in7, in7, tp5_m, tp6_m, tp7_m, \
|
H A D | deblock_msa.c | 17 #define VPX_TRANSPOSE8x16_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, out0, \ 25 ILVR_B4_SH(in1, in0, in3, in2, in5, in4, in7, in6, temp0, temp1, temp2, \ 31 ILVL_B4_SH(in1, in0, in3, in2, in5, in4, in7, in6, temp0, temp1, temp2, \ 74 #define TRANSPOSE12x16_B(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, \ 82 ILVR_B2_SH(in5, in4, in7, in6, temp0, temp1); \ 97 ILVL_B2_SH(in5, in4, in7, in6, temp0, temp1); \ 100 in7 = (v16u8)__msa_ilvl_d((v2i64)temp7, (v2i64)temp3); \ 113 #define VPX_TRANSPOSE12x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, \ 121 ILVR_B2_SH(in5, in4, in7, in6, temp0, temp1); \ 127 ILVL_B2_SH(in5, in4, in7, in [all...] |
/external/libvpx/libvpx/vp9/encoder/x86/ |
H A D | vp9_dct_ssse3.c | 51 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); local 66 in7 = _mm_slli_epi16(in7, 2); 75 in[7] = &in7; 85 const __m128i q0 = _mm_add_epi16(in0, in7); 92 const __m128i q7 = _mm_sub_epi16(in0, in7); 241 in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 264 const __m128i sign_in7 = _mm_srai_epi16(in7, 15); 272 in7 = _mm_sub_epi16(in7, sign_in [all...] |
H A D | vp9_dct_intrin_sse2.c | 115 __m128i in7 = _mm_add_epi16(in[0], in[1]); local 119 u[2] = _mm_unpacklo_epi16(in7, kZero); 213 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); local 228 in7 = _mm_slli_epi16(in7, 2); 237 in[7] = &in7; 247 const __m128i q0 = _mm_add_epi16(in0, in7); 254 const __m128i q7 = _mm_sub_epi16(in0, in7); 409 in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 432 const __m128i sign_in7 = _mm_srai_epi16(in7, 1 923 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
/external/webp/src/dsp/ |
H A D | msa_macro.h | 308 #define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, \ 311 ST_B4(RTYPE, in4, in5, in6, in7, pdst + 4 * stride, stride); \ 871 #define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, \ 874 ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ 897 #define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, \ 900 ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3); \ 921 #define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, \ 924 ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3); \ 982 #define PCKEV_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, \ 985 PCKEV_B2(RTYPE, in4, in5, in6, in7, out [all...] |
H A D | lossless_sse2.c | 471 __m128i in7 = _mm_loadu_si128(in + 7); local 473 VP8L32bToPlanar(&in4, &in5, &in6, &in7); 476 VP8PlanarTo24b(&in1, &in5, &in2, &in6, &in3, &in7); 482 _mm_storeu_si128(out + 5, in7);
|
/external/libvpx/libvpx/vp8/common/mips/msa/ |
H A D | vp8_macros_msa.h | 360 #define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \ 363 ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride); \ 953 #define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 957 ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ 1008 #define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 1012 ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3); \ 1035 #define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 1039 ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3); \ 1067 #define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 1071 ILVR_D2(RTYPE, in4, in5, in6, in7, out [all...] |
/external/libvpx/libvpx/vpx_dsp/x86/ |
H A D | fwd_txfm_impl_sse2.h | 284 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); local 293 in7 = _mm_slli_epi16(in7, 2); 303 const __m128i q0 = ADD_EPI16(in0, in7); 310 const __m128i q7 = SUB_EPI16(in0, in7); 516 in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 539 const __m128i sign_in7 = _mm_srai_epi16(in7, 15); 547 in7 = _mm_sub_epi16(in7, sign_in7); 555 in7 [all...] |
H A D | inv_txfm_ssse3.c | 31 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local 45 in7 = load_input_data(input + 8 * 7); 50 TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, 51 in4, in5, in6, in7); 57 const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); 58 const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); 186 in7 = _mm_sub_epi16(stp1_0, stp2_7); 198 in7 = _mm_adds_epi16(in7, final_rounding); 207 in7 235 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
H A D | inv_txfm_sse2.c | 116 __m128i u[8], v[8], in7; local 119 in7 = _mm_srli_si128(in[1], 8); 120 in7 = _mm_add_epi16(in7, in[0]); 121 in7 = _mm_sub_epi16(in7, in[1]); 125 u[2] = _mm_unpacklo_epi16(in7, kZero); 178 #define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3, \ 183 const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \ 184 const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \ 262 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local 351 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local 386 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local 608 __m128i in0, in1, in2, in3, in4, in5, in6, in7; local [all...] |
H A D | inv_txfm_sse2.h | 49 #define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ 57 const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \ 59 const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
|
/external/libjpeg-turbo/simd/ |
H A D | jquanti-altivec.c | 57 __vector unsigned char in0, in1, in2, in3, in4, in5, in6, in7; local 80 out7 = (__vector short)VEC_UNPACKHU(in7);
|
/external/boringssl/src/crypto/fipsmodule/aes/asm/ |
H A D | aesp8-ppc.pl | 672 my ($in0, $in1, $in2, $in3, $in4, $in5, $in6, $in7 )=map("v$_",(0..3,10..13)); 789 lvx_u $in7,$x70,$inp 795 le?vperm $in7,$in7,$in7,$inpperm 798 vxor $out7,$in7,$rndkey0 858 # loop inX-in7 are loaded 927 vmr $ivec,$in7 929 lvx_u $in7,$x70,$inp 939 le?vperm $in7, [all...] |
/external/libvpx/libvpx/vpx_dsp/arm/ |
H A D | idct32x32_135_add_neon.c | 24 int16x8_t *const in7) { 39 *in7 = load_tran_low_to_s16q(input); 46 int16x4_t *const in7) { 61 *in7 = load_tran_low_to_s16d(input); 20 load_8x8_s16(const tran_low_t *input, int16x8_t *const in0, int16x8_t *const in1, int16x8_t *const in2, int16x8_t *const in3, int16x8_t *const in4, int16x8_t *const in5, int16x8_t *const in6, int16x8_t *const in7) argument 42 load_4x8_s16(const tran_low_t *input, int16x4_t *const in0, int16x4_t *const in1, int16x4_t *const in2, int16x4_t *const in3, int16x4_t *const in4, int16x4_t *const in5, int16x4_t *const in6, int16x4_t *const in7) argument
|
H A D | highbd_idct32x32_135_add_neon.c | 22 int32x4x2_t *const in5, int32x4x2_t *const in6, int32x4x2_t *const in7) { 44 in7->val[0] = vld1q_s32(input); 45 in7->val[1] = vld1q_s32(input + 4); 53 int32x4_t *const in7) { 68 *in7 = vld1q_s32(input); 19 load_8x8_s32_dual( const tran_low_t *input, int32x4x2_t *const in0, int32x4x2_t *const in1, int32x4x2_t *const in2, int32x4x2_t *const in3, int32x4x2_t *const in4, int32x4x2_t *const in5, int32x4x2_t *const in6, int32x4x2_t *const in7) argument 48 load_4x8_s32_dual(const tran_low_t *input, int32x4_t *const in0, int32x4_t *const in1, int32x4_t *const in2, int32x4_t *const in3, int32x4_t *const in4, int32x4_t *const in5, int32x4_t *const in6, int32x4_t *const in7) argument
|