| /external/libaom/libaom/av1/common/x86/ |
| D | highbd_warp_plane_sse4.c | 533 __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_highbd_warp_affine_sse4_1() local
540 res_lo = _mm_add_epi32(res_lo, res_add_const); in av1_highbd_warp_affine_sse4_1()
541 res_lo = _mm_sra_epi32(_mm_add_epi32(res_lo, reduce_bits_vert_const), in av1_highbd_warp_affine_sse4_1()
549 res_lo = _mm_add_epi32(_mm_mullo_epi32(p_32, wt0), in av1_highbd_warp_affine_sse4_1()
550 _mm_mullo_epi32(res_lo, wt1)); in av1_highbd_warp_affine_sse4_1()
551 res_lo = _mm_srai_epi32(res_lo, DIST_PRECISION_BITS); in av1_highbd_warp_affine_sse4_1()
553 res_lo = _mm_srai_epi32(_mm_add_epi32(p_32, res_lo), 1); in av1_highbd_warp_affine_sse4_1()
556 __m128i res32_lo = _mm_add_epi32(res_lo, res_sub_const); in av1_highbd_warp_affine_sse4_1()
564 res_lo = _mm_packus_epi32(res_lo, res_lo); in av1_highbd_warp_affine_sse4_1()
565 _mm_storel_epi64(p, res_lo); in av1_highbd_warp_affine_sse4_1()
[all …]
|
| D | convolve_sse2.c | 167 __m128i s[8], src6, res_lo, res_hi; in av1_convolve_y_sr_sse2() local
199 res_lo = convolve_lo_y(s, coeffs); // Filter low index pixels in av1_convolve_y_sr_sse2()
203 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_convolve_y_sr_sse2()
213 res_lo = convolve_lo_y(s + 1, coeffs); // Filter low index pixels in av1_convolve_y_sr_sse2()
217 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_convolve_y_sr_sse2()
276 const __m128i res_lo = convolve_lo_x(s, coeffs); in av1_convolve_x_sr_sse2() local
278 _mm_sra_epi32(_mm_add_epi32(res_lo, round_0_const), round_0_shift); in av1_convolve_x_sr_sse2()
319 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_convolve_x_sr_sse2() local
322 _mm_sra_epi32(_mm_add_epi32(res_lo, round_0_const), round_0_shift); in av1_convolve_x_sr_sse2()
|
| D | warp_plane_sse4.c | 544 __m128i *res_lo, __m128i *res_hi, in filter_src_pixels_vertical() argument
577 *res_lo = _mm_unpacklo_epi32(res_even, res_odd); in filter_src_pixels_vertical()
582 __m128i *res_lo, __m128i *res_hi, const __m128i *res_add_const, in store_vertical_filter_output() argument
587 __m128i res_lo_1 = *res_lo; in store_vertical_filter_output()
693 __m128i res_lo; in warp_vertical_filter() local
695 filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k); in warp_vertical_filter()
697 store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt, in warp_vertical_filter()
721 __m128i res_lo; in warp_vertical_filter_gamma0() local
723 filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k); in warp_vertical_filter_gamma0()
725 store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt, in warp_vertical_filter_gamma0()
[all …]
|
| D | jnt_convolve_sse2.c | 67 const __m128i res_lo = convolve_lo_x(s, coeffs); in av1_dist_wtd_convolve_x_sse2() local
69 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_dist_wtd_convolve_x_sse2()
119 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_dist_wtd_convolve_x_sse2() local
122 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_dist_wtd_convolve_x_sse2()
285 __m128i s[8], src6, res_lo, res_hi, res_lo_shift, res_hi_shift; in av1_dist_wtd_convolve_y_sse2() local
316 res_lo = convolve_lo_y(s, coeffs); // Filter low index pixels in av1_dist_wtd_convolve_y_sse2()
318 res_lo_shift = _mm_sll_epi32(res_lo, left_shift); in av1_dist_wtd_convolve_y_sse2()
346 res_lo = convolve_lo_y(s + 1, coeffs); // Filter low index pixels in av1_dist_wtd_convolve_y_sse2()
348 res_lo_shift = _mm_sll_epi32(res_lo, left_shift); in av1_dist_wtd_convolve_y_sse2()
580 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_dist_wtd_convolve_2d_sse2() local
[all …]
|
| D | jnt_convolve_avx2.c | 272 __m256i res_lo = convolve_lowbd_4tap(s, coeffs + 1); in av1_dist_wtd_convolve_y_avx2() local
274 res_lo = _mm256_add_epi16(res_lo, offset_const_1); in av1_dist_wtd_convolve_y_avx2()
276 const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero); in av1_dist_wtd_convolve_y_avx2()
282 const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero); in av1_dist_wtd_convolve_y_avx2()
451 __m256i res_lo = convolve_lowbd(s, coeffs); in av1_dist_wtd_convolve_y_avx2() local
453 res_lo = _mm256_add_epi16(res_lo, offset_const_1); in av1_dist_wtd_convolve_y_avx2()
455 const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero); in av1_dist_wtd_convolve_y_avx2()
461 const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero); in av1_dist_wtd_convolve_y_avx2()
|
| D | warp_plane_avx2.c | 627 __m256i *res_lo, in filter_src_pixels_vertical_avx2() argument
654 *res_lo = _mm256_unpacklo_epi32(res_even, res_odd); in filter_src_pixels_vertical_avx2()
659 const __m256i *res_lo, const __m256i *res_hi, const __m256i *res_add_const, in store_vertical_filter_output_avx2() argument
664 __m256i res_lo_1 = *res_lo; in store_vertical_filter_output_avx2()
810 __m256i res_lo, res_hi; in warp_vertical_filter_avx2() local
811 filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi, in warp_vertical_filter_avx2()
813 store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt, in warp_vertical_filter_avx2()
860 __m256i res_lo, res_hi; in warp_vertical_filter_gamma0_avx2() local
861 filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi, in warp_vertical_filter_gamma0_avx2()
863 store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt, in warp_vertical_filter_gamma0_avx2()
[all …]
|
| D | convolve_avx2.c | 94 const __m256i res_lo = convolve_lowbd_4tap(s, coeffs + 1); in av1_convolve_y_sr_avx2() local
98 _mm256_add_epi16(res_lo, right_shift_const), right_shift); in av1_convolve_y_sr_avx2()
206 const __m256i res_lo = convolve_lowbd(s, coeffs); in av1_convolve_y_sr_avx2() local
211 _mm256_add_epi16(res_lo, right_shift_const), right_shift); in av1_convolve_y_sr_avx2()
|
| D | convolve_2d_sse2.c | 188 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_convolve_2d_sr_sse2() local
192 _mm_sra_epi32(_mm_add_epi32(res_lo, sum_round), sum_shift); in av1_convolve_2d_sr_sse2()
402 const __m128i res_lo = _mm_sll_epi16(d16_lo, left_shift); in av1_dist_wtd_convolve_2d_copy_sse2() local
403 const __m128i res_unsigned_lo = _mm_add_epi16(res_lo, offset_const); in av1_dist_wtd_convolve_2d_copy_sse2()
|
| D | wiener_convolve_sse2.c | 183 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_wiener_convolve_add_src_sse2() local
187 _mm_add_epi32(res_lo, round_const), conv_params->round_1); in av1_wiener_convolve_add_src_sse2()
|
| D | highbd_wiener_convolve_ssse3.c | 185 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_highbd_wiener_convolve_add_src_ssse3() local
189 _mm_add_epi32(res_lo, round_const), conv_params->round_1); in av1_highbd_wiener_convolve_add_src_ssse3()
|
| D | highbd_wiener_convolve_avx2.c | 225 const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd); in av1_highbd_wiener_convolve_add_src_avx2() local
229 _mm256_add_epi32(res_lo, round_const), conv_params->round_1); in av1_highbd_wiener_convolve_add_src_avx2()
|
| D | jnt_convolve_ssse3.c | 196 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_dist_wtd_convolve_2d_ssse3() local
200 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_dist_wtd_convolve_2d_ssse3()
|
| D | highbd_convolve_2d_sse4.c | 348 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); in av1_highbd_dist_wtd_convolve_2d_sse4_1() local
352 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
|
| /external/libaom/libaom/av1/common/arm/ |
| D | warp_plane_neon.c | 488 int32x4_t res_lo, res_hi; in av1_warp_affine_neon() local
627 vertical_filter_neon(v_src, &res_lo, &res_hi, sy, gamma); in av1_warp_affine_neon()
629 res_lo = vaddq_s32(res_lo, add_const_vert); in av1_warp_affine_neon()
637 res_lo = vrshlq_s32(res_lo, shift_vert); in av1_warp_affine_neon()
644 res_lo = vmulq_s32(res_lo, bwd); in av1_warp_affine_neon()
646 tmp32_lo = vaddq_s32(tmp32_lo, res_lo); in av1_warp_affine_neon()
649 tmp32_lo = vaddq_s32(tmp32_lo, res_lo); in av1_warp_affine_neon()
659 uint16x4_t res_u16_low = vqmovun_s32(res_lo); in av1_warp_affine_neon()
696 res_lo = vrshlq_s32(res_lo, shift_vert); in av1_warp_affine_neon()
699 result_final = vcombine_s16(vmovn_s32(res_lo), vmovn_s32(res_hi)); in av1_warp_affine_neon()
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | blend_mask_sse4.h | 52 __m128i res_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0, s1), in blend_a64_d16_mask_w8_sse41() local
56 res_lo = _mm_srai_epi32(_mm_sub_epi32(res_lo, *v_round_offset), shift); in blend_a64_d16_mask_w8_sse41()
58 const __m128i res_e = _mm_packs_epi32(res_lo, res_hi); in blend_a64_d16_mask_w8_sse41()
|
| D | convolve_sse2.h | 90 const __m128i res_lo = _mm_srai_epi32(wt_res_lo, DIST_PRECISION_BITS); in comp_avg() local
93 res = _mm_packs_epi32(res_lo, res_hi); in comp_avg()
|
| D | aom_subpixel_8t_intrin_sse2.c | 62 __m128i res_lo = _mm_unpacklo_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2); in aom_filter_block1d16_h4_sse2() local
64 srcRegFilt32b1_1 = _mm_packs_epi32(res_lo, res_hi); in aom_filter_block1d16_h4_sse2()
86 res_lo = _mm_unpacklo_epi32(srcRegFilt32b2_1, srcRegFilt32b2_2); in aom_filter_block1d16_h4_sse2()
88 srcRegFilt32b2_1 = _mm_packs_epi32(res_lo, res_hi); in aom_filter_block1d16_h4_sse2()
300 __m128i res_lo = _mm_unpacklo_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2); in aom_filter_block1d8_h4_sse2() local
302 srcRegFilt32b1_1 = _mm_packs_epi32(res_lo, res_hi); in aom_filter_block1d8_h4_sse2()
|
| D | convolve_avx2.h | 414 const __m256i res_lo = _mm256_srai_epi32(wt_res_lo, DIST_PRECISION_BITS); in comp_avg() local
417 res = _mm256_packs_epi32(res_lo, res_hi); in comp_avg()
|
| /external/libgav1/libgav1/src/dsp/arm/ |
| D | distance_weighted_blend_neon.cc | 109 const int16x8_t res_lo = in DistanceWeightedBlendLarge_NEON() local
118 vcombine_u8(vqmovun_s16(res_lo), vqmovun_s16(res_hi)); in DistanceWeightedBlendLarge_NEON()
|
| /external/mesa3d/src/compiler/nir/ |
| D | nir_lower_int64.c | 303 nir_ssa_def *res_lo = nir_iadd(b, x_lo, y_lo); in lower_iadd64() local
304 nir_ssa_def *carry = nir_b2i32(b, nir_ult(b, res_lo, x_lo)); in lower_iadd64()
307 return nir_pack_64_2x32_split(b, res_lo, res_hi); in lower_iadd64()
318 nir_ssa_def *res_lo = nir_isub(b, x_lo, y_lo); in lower_isub64() local
322 return nir_pack_64_2x32_split(b, res_lo, res_hi); in lower_isub64()
492 nir_ssa_def *res_lo = nir_ior(b, res_hi, nir_b2i32(b, is_non_zero)); in lower_isign64() local
494 return nir_pack_64_2x32_split(b, res_lo, res_hi); in lower_isign64()
775 nir_ssa_def *res_lo = nir_f2u32(b, nir_frem(b, x, div)); in lower_f2() local
776 nir_ssa_def *res = nir_pack_64_2x32_split(b, res_lo, res_hi); in lower_f2()
|
| /external/libgav1/libgav1/src/dsp/x86/ |
| D | distance_weighted_blend_sse4.cc | 139 const __m128i res_lo = in DistanceWeightedBlendLarge_SSE4_1() local
147 StoreUnaligned16(dst + x, _mm_packus_epi16(res_lo, res_hi)); in DistanceWeightedBlendLarge_SSE4_1()
|
| /external/u-boot/drivers/bios_emulator/x86emu/ |
| D | ops.c | 962 u32 res_lo,res_hi; in x86emuOp_imul_word_IMM() local
971 imul_long_direct(&res_lo,&res_hi,(s32)srcval,(s32)imm); in x86emuOp_imul_word_IMM()
972 if ((((res_lo & 0x80000000) == 0) && (res_hi == 0x00000000)) || in x86emuOp_imul_word_IMM()
973 (((res_lo & 0x80000000) != 0) && (res_hi == 0xFFFFFFFF))) { in x86emuOp_imul_word_IMM()
980 *destreg = (u32)res_lo; in x86emuOp_imul_word_IMM()
1007 u32 res_lo,res_hi; in x86emuOp_imul_word_IMM() local
1016 imul_long_direct(&res_lo,&res_hi,(s32)*srcreg,(s32)imm); in x86emuOp_imul_word_IMM()
1017 if ((((res_lo & 0x80000000) == 0) && (res_hi == 0x00000000)) || in x86emuOp_imul_word_IMM()
1018 (((res_lo & 0x80000000) != 0) && (res_hi == 0xFFFFFFFF))) { in x86emuOp_imul_word_IMM()
1025 *destreg = (u32)res_lo; in x86emuOp_imul_word_IMM()
[all …]
|
| D | ops2.c | 732 u32 res_lo,res_hi; in x86emuOp2_imul_R_RM() local
739 imul_long_direct(&res_lo,&res_hi,(s32)*destreg,(s32)srcval); in x86emuOp2_imul_R_RM()
747 *destreg = (u32)res_lo; in x86emuOp2_imul_R_RM()
771 u32 res_lo,res_hi; in x86emuOp2_imul_R_RM() local
777 imul_long_direct(&res_lo,&res_hi,(s32)*destreg,(s32)*srcreg); in x86emuOp2_imul_R_RM()
785 *destreg = (u32)res_lo; in x86emuOp2_imul_R_RM()
|
| D | prim_ops.c | 1913 void imul_long_direct(u32 *res_lo, u32* res_hi,u32 d, u32 s) in imul_long_direct() argument
1918 *res_lo = (u32)res; in imul_long_direct()
1936 *res_lo = (rlo_hi << 16) | (rlo_lo & 0xFFFF); in imul_long_direct()
1939 d = ~*res_lo; in imul_long_direct()
1941 *res_lo = ~*res_lo+1; in imul_long_direct()
|
| /external/u-boot/drivers/bios_emulator/include/x86emu/ |
| D | prim_ops.h | 119 void imul_long_direct(u32 *res_lo, u32* res_hi,u32 d, u32 s);
|