/external/flac/libFLAC/ |
H A D | lpc_intrin_avx2.c | 64 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 71 q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); 85 mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); 97 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; local 104 q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); 116 mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); 130 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; local 137 q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); 147 mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); 159 __m256i q0, q1, q2, q3, q4, q5, q6, q local 190 __m256i q0, q1, q2, q3, q4, q5, q6, q7; local 215 __m256i q0, q1, q2, q3, q4, q5, q6; local 419 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 452 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; local 485 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; local 514 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8; local 545 __m256i q0, q1, q2, q3, q4, q5, q6, q7; local 570 __m256i q0, q1, q2, q3, q4, q5, q6; local 778 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 811 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; local 844 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; local 873 __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8; local 904 __m256i q0, q1, q2, q3, q4, q5, q6, q7; local 929 __m256i q0, q1, q2, q3, q4, q5, q6; local [all...] |
H A D | lpc_intrin_sse2.c | 70 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 77 q6 = _mm_cvtsi32_si128(0xffff & qlp_coeff[6]); q6 = _mm_shuffle_epi32(q6, _MM_SHUFFLE(0,0,0,0)); 91 mull = _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); 103 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; local 110 q6 = _mm_cvtsi32_si128(0xffff & qlp_coeff[6]); q6 = _mm_shuffle_epi32(q6, _MM_SHUFFLE(0,0,0,0)); 122 mull = _mm_madd_epi16(q6, _mm_loadu_si12 136 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; local 165 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8; local 196 __m128i q0, q1, q2, q3, q4, q5, q6, q7; local 221 __m128i q0, q1, q2, q3, q4, q5, q6; local [all...] |
H A D | lpc_intrin_sse41.c | 971 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 978 q6 = _mm_cvtsi32_si128(qlp_coeff[6]); q6 = _mm_shuffle_epi32(q6, _MM_SHUFFLE(0,0,0,0)); 992 mull = _mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); 1004 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; local 1011 q6 = _mm_cvtsi32_si128(qlp_coeff[6]); q6 = _mm_shuffle_epi32(q6, _MM_SHUFFLE(0,0,0,0)); 1023 mull = _mm_mullo_epi32(q6, _mm_loadu_si12 1037 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; local 1066 __m128i q0, q1, q2, q3, q4, q5, q6, q7, q8; local 1097 __m128i q0, q1, q2, q3, q4, q5, q6, q7; local 1122 __m128i q0, q1, q2, q3, q4, q5, q6; local [all...] |
/external/libvpx/libvpx/vp8/common/arm/neon/ |
H A D | dequant_idct_neon.c | 27 int16x8_t q1, q2, q3, q4, q5, q6; local 44 q6 = vld1q_s16(dq); 59 vmulq_u16(vreinterpretq_u16_s16(q4), vreinterpretq_u16_s16(q6)));
|
H A D | idct_dequant_full_2x_neon.c | 21 int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; local 86 q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2); 100 q2 = vqsubq_s16(q6, q5); 105 q6 = vqsubq_s16(q11, q2); 108 q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6)); 135 q6 = vqsubq_s16(q3, q8); 140 q6 = vrshrq_n_s16(q6, 3); 143 q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6)); 154 q6 [all...] |
H A D | mbloopfilter_neon.c | 20 uint8x16_t q6, // p0 40 q13u8 = vabdq_u8(q5, q6); 50 q12u8 = vabdq_u8(q6, q7); 68 q6 = veorq_u8(q6, q0u8); 79 vget_low_s8(vreinterpretq_s8_u8(q6))); 81 vget_high_s8(vreinterpretq_s8_u8(q6))); 109 q6s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q13s8); 158 uint8x16_t q5, q6, q7, q8, q9, q10; local 172 q6 14 vp8_mbloop_filter_neon(uint8x16_t qblimit, uint8x16_t qlimit, uint8x16_t qthresh, uint8x16_t q3, uint8x16_t q4, uint8x16_t q5, uint8x16_t q6, uint8x16_t q7, uint8x16_t q8, uint8x16_t q9, uint8x16_t q10, uint8x16_t *q4r, uint8x16_t *q5r, uint8x16_t *q6r, uint8x16_t *q7r, uint8x16_t *q8r, uint8x16_t *q9r) argument 206 uint8x16_t q5, q6, q7, q8, q9, q10; local 293 uint8x16_t q5, q6, q7, q8, q9, q10; local 457 uint8x16_t q5, q6, q7, q8, q9, q10; local [all...] |
H A D | vp8_loopfilter_neon.c | 21 uint8x16_t q6, // p0 38 q13u8 = vabdq_u8(q5, q6); 48 q9 = vabdq_u8(q6, q7); 65 q6 = veorq_u8(q6, q10); 74 vget_low_s8(vreinterpretq_s8_u8(q6))); 76 vget_high_s8(vreinterpretq_s8_u8(q6))); 108 q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8); 130 uint8x16_t q5, q6, q7, q8, q9, q10; local 143 q6 15 vp8_loop_filter_neon(uint8x16_t qblimit, uint8x16_t qlimit, uint8x16_t qthresh, uint8x16_t q3, uint8x16_t q4, uint8x16_t q5, uint8x16_t q6, uint8x16_t q7, uint8x16_t q8, uint8x16_t q9, uint8x16_t q10, uint8x16_t *q5r, uint8x16_t *q6r, uint8x16_t *q7r, uint8x16_t *q8r) argument 173 uint8x16_t q5, q6, q7, q8, q9, q10; local 321 uint8x16_t q5, q6, q7, q8, q9, q10; local 433 uint8x16_t q5, q6, q7, q8, q9, q10; local [all...] |
/external/guava/guava/src/com/google/common/cache/ |
H A D | Striped64.java | 100 volatile long q0, q1, q2, q3, q4, q5, q6; field in class:Striped64.Cell
|
/external/webrtc/webrtc/modules/audio_coding/codecs/g722/ |
H A D | g722_encode.c | 208 static const int q6[32] = local 347 wd1 = (q6[i]*s->band[0].det) >> 12;
|
/external/capstone/suite/MC/ARM/ |
H A D | neon-bitwise-encoding.s.cs | 92 0x5a,0xc1,0x0c,0xf2 = vand q6, q6, q5 external variable declarations 93 0x5a,0xc1,0x0c,0xf2 = vand q6, q6, q5 external variable declarations 97 0x5a,0xc1,0x0c,0xf3 = veor q6, q6, q5 external variable declarations 98 0x5a,0xc1,0x0c,0xf3 = veor q6, q6, q5 external variable declarations 102 0x5a,0xc1,0x0c,0xf3 = veor q6, q6, q external variable declarations 103 0x5a,0xc1,0x0c,0xf3 = veor q6, q6, q5 external variable declarations [all...] |
H A D | neon-mul-encoding.s.cs | 58 0x65,0xc9,0xac,0xf3 = vmul.f32 q6, q6, d5[1] external variable declarations 69 0x42,0xc8,0xa6,0xf3 = vmul.i32 q6, q3, d2[0] 72 0x65,0x69,0xac,0xf3 = vmul.f32 q3, q6, d5[1] external variable declarations
|
H A D | neon-add-encoding.s.cs | 43 0x4e,0xc0,0x2c,0xf3 = vhadd.u32 q6, q6, q7 external variable declarations 110 0x4a,0xc8,0x0c,0xf2 = vadd.i8 q6, q6, q5 external variable declarations 114 0x05,0xc1,0x8c,0xf2 = vaddw.s8 q6, q6, d5 external variable declarations 117 0x05,0xc1,0x8c,0xf3 = vaddw.u8 q6, q6, d5 external variable declarations
|
H A D | neon-sub-encoding.s.cs | 21 0x4c,0xad,0x2a,0xf2 = vsub.f32 q5, q5, q6 76 0x4e,0xc2,0x2c,0xf3 = vhsub.u32 q6, q6, q7 external variable declarations 77 0x05,0xc3,0x8c,0xf2 = vsubw.s8 q6, q6, d5 external variable declarations 80 0x05,0xc3,0x8c,0xf3 = vsubw.u8 q6, q6, d5 external variable declarations
|
H A D | neon-shift-encoding.s.cs | 56 0x5c,0x61,0xa1,0xf2 = vsra.s32 q3, q6, #31 external variable declarations 64 0x5c,0xc1,0xa1,0xf2 = vsra.s32 q6, q6, #31 external variable declarations 72 0x5c,0x61,0xa1,0xf3 = vsra.u32 q3, q6, #31 external variable declarations 80 0x5c,0xc1,0xa1,0xf3 = vsra.u32 q6, q6, #31 external variable declarations 88 0x5c,0x64,0xa1,0xf3 = vsri.32 q3, q6, #31 external variable declarations 96 0x5c,0xc4,0xa1,0xf3 = vsri.32 q6, q6, #31 external variable declarations 104 0x5c,0x65,0xbf,0xf3 = vsli.32 q3, q6, #3 external variable declarations 112 0x5c,0xc5,0xbf,0xf3 = vsli.32 q6, q6, #31 external variable declarations 204 0xcc,0xc5,0x16,0xf3 = vrshl.u16 q6, q6, q11 external variable declarations 220 0x5c,0xc0,0x90,0xf3 = vshr.u16 q6, q6, #16 external variable declarations 236 0x5c,0xc2,0x90,0xf3 = vrshr.u16 q6, q6, #16 external variable declarations [all...] |
/external/libjpeg-turbo/simd/ |
H A D | jsimd_arm_neon.S | 111 JLONG q1, q2, q3, q4, q5, q6, q7; \ 126 q6 = MULTIPLY(q5, FIX_1_175875602_MINUS_1_961570560) + \ define 132 q4 = q6; \ 134 q6 += MULTIPLY(row5, -FIX_2_562915447) + \ 141 q1 = q1 + q6; \ 149 q1 = q1 - q6; \ 152 q1 = q1 - q6; \ 153 q6 = MULTIPLY(row2, FIX_0_541196100_PLUS_0_765366865) + \ define 161 q2 = q1 + q6; \ 162 q1 = q1 - q6; \ [all...] |
/external/libvpx/libvpx/vpx_dsp/ |
H A D | loopfilter.c | 244 q5 = *oq5, q6 = *oq6, q7 = *oq7; local 262 q1 + q2 + q3 + q4 + q5 + q6, 265 q2 + q3 + q4 + q5 + q6 + q7, 268 q3 + q4 + q5 + q6 + q7 * 2, 271 p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 274 p3 + p2 + p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4); 276 p2 + p1 + p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4); 278 p1 + p0 + q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4); 280 p0 + q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4); 598 const uint16_t q6 local [all...] |
/external/libvpx/libvpx/vpx_dsp/mips/ |
H A D | loopfilter_filters_dspr2.h | 514 const uint32_t q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7; local 522 /* addition of p6,p5,p4,p3,p2,p1,p0,q0,q1,q2,q3,q4,q5,q6 536 "addu.ph %[add_p6toq6], %[add_p6toq6], %[q6] \n\t" 542 [q3] "r"(q3), [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), 557 "subu.ph %[res_op6], %[res_op6], %[q6] \n\t" 571 "subu.ph %[res_op5], %[res_op5], %[q6] \n\t" 583 "subu.ph %[res_op4], %[res_op4], %[q6] \n\t" 593 "subu.ph %[res_op3], %[res_op3], %[q6] \n\t" 603 "subu.ph %[res_op2], %[res_op2], %[q6] \ [all...] |
H A D | loopfilter_mb_horiz_dspr2.c | 32 uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; local 98 "lw %[q6], (%[sq6]) \n\t" 102 [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4) 109 flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2); 160 "sw %[q6], (%[sq6]) \n\t" 169 : [q6] "r"(q6), [q5] "r"(q5), [q4] "r"(q4), [q3] "r"(q3),
|
H A D | loopfilter_mb_vert_dspr2.c | 30 uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; local 82 "lw %[q6], +4(%[s2]) \n\t" 87 [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4) 198 /* transpose q4, q5, q6, q7 202 q6 q6_0 q6_1 q6_2 q6_3 209 q6 q4_2 q5_2 q26_2 q7_2 214 "precrq.qb.ph %[prim1], %[q7], %[q6] \n\t" 215 "precr.qb.ph %[prim2], %[q7], %[q6] \n\t" 219 "precrq.qb.ph %[q6], [all...] |
/external/clang/test/SemaCXX/ |
H A D | overload-call.cpp | 207 float* q6 = quals_rank2(pp); local
|
/external/libvpx/libvpx/vp9/encoder/x86/ |
H A D | vp9_dct_ssse3.c | 91 const __m128i q6 = _mm_sub_epi16(in1, in6); local 146 const __m128i d0 = _mm_sub_epi16(q6, q5); 147 const __m128i d1 = _mm_add_epi16(q6, q5);
|
/external/libvpx/libvpx/vpx_dsp/arm/ |
H A D | highbd_loopfilter_neon.c | 197 const uint16x8_t q6, const uint16x8_t q7, uint16x8_t *op6, uint16x8_t *op5, 219 *op0 = apply_15_tap_filter_kernel(flat2, p7, p1, p0, q6, *op0, &sum); 226 *oq6 = apply_15_tap_filter_kernel(flat2, p1, q5, q6, q7, q6, &sum); 320 const uint16x8_t q4, const uint16x8_t q5, const uint16x8_t q6, 338 q4, q5, q6, q7, op6, op5, op4, op3, op2, op1, op0, 543 const uint16x8_t q6, const uint32_t flat_status, 554 vst1q_u16(s + 6 * p, q6); 670 q3, q4, q5, q6, q7, op6, op5, op4, op3, op2, op1, op0, oq0, oq1, oq2, oq3, local 675 &q3, &q4, &q5, &q6, 191 apply_15_tap_filter( const uint16x8_t flat2, const uint16x8_t p7, const uint16x8_t p6, const uint16x8_t p5, const uint16x8_t p4, const uint16x8_t p3, const uint16x8_t p2, const uint16x8_t p1, const uint16x8_t p0, const uint16x8_t q0, const uint16x8_t q1, const uint16x8_t q2, const uint16x8_t q3, const uint16x8_t q4, const uint16x8_t q5, const uint16x8_t q6, const uint16x8_t q7, uint16x8_t *op6, uint16x8_t *op5, uint16x8_t *op4, uint16x8_t *op3, uint16x8_t *op2, uint16x8_t *op1, uint16x8_t *op0, uint16x8_t *oq0, uint16x8_t *oq1, uint16x8_t *oq2, uint16x8_t *oq3, uint16x8_t *oq4, uint16x8_t *oq5, uint16x8_t *oq6) argument 313 filter16( const uint16x8_t mask, const uint16x8_t flat, const uint32_t flat_status, const uint16x8_t flat2, const uint32_t flat2_status, const uint16x8_t hev, const uint16x8_t p7, const uint16x8_t p6, const uint16x8_t p5, const uint16x8_t p4, const uint16x8_t p3, const uint16x8_t p2, const uint16x8_t p1, const uint16x8_t p0, const uint16x8_t q0, const uint16x8_t q1, const uint16x8_t q2, const uint16x8_t q3, const uint16x8_t q4, const uint16x8_t q5, const uint16x8_t q6, const uint16x8_t q7, uint16x8_t *op6, uint16x8_t *op5, uint16x8_t *op4, uint16x8_t *op3, uint16x8_t *op2, uint16x8_t *op1, uint16x8_t *op0, uint16x8_t *oq0, uint16x8_t *oq1, uint16x8_t *oq2, uint16x8_t *oq3, uint16x8_t *oq4, uint16x8_t *oq5, uint16x8_t *oq6, const int bd) argument 536 store_8x14(uint16_t *s, const int p, const uint16x8_t p6, const uint16x8_t p5, const uint16x8_t p4, const uint16x8_t p3, const uint16x8_t p2, const uint16x8_t p1, const uint16x8_t p0, const uint16x8_t q0, const uint16x8_t q1, const uint16x8_t q2, const uint16x8_t q3, const uint16x8_t q4, const uint16x8_t q5, const uint16x8_t q6, const uint32_t flat_status, const uint32_t flat2_status) argument 693 q3, q4, q5, q6, q7, op6, op5, op4, op3, op2, op1, op0, oq0, oq1, oq2, oq3, local [all...] |
H A D | loopfilter_neon.c | 313 const uint8x8_t q6, const uint8x8_t q7, uint8x8_t *op6, uint8x8_t *op5, 335 *op0 = apply_15_tap_filter_8_kernel(flat2, p7, p1, p0, q6, *op0, &sum); 342 *oq6 = apply_15_tap_filter_8_kernel(flat2, p1, q5, q6, q7, q6, &sum); 351 const uint8x16_t q6, const uint8x16_t q7, uint8x16_t *op6, uint8x16_t *op5, 389 apply_15_tap_filter_16_kernel(flat2, p7, p1, p0, q6, *op0, &sum0, &sum1); 399 *oq6 = apply_15_tap_filter_16_kernel(flat2, p1, q5, q6, q7, q6, &sum0, &sum1); 486 const uint8x##w##_t q4, const uint8x##w##_t q5, const uint8x##w##_t q6, \ 505 q2, q3, q4, q5, q6, q 307 apply_15_tap_filter_8( const uint8x8_t flat2, const uint8x8_t p7, const uint8x8_t p6, const uint8x8_t p5, const uint8x8_t p4, const uint8x8_t p3, const uint8x8_t p2, const uint8x8_t p1, const uint8x8_t p0, const uint8x8_t q0, const uint8x8_t q1, const uint8x8_t q2, const uint8x8_t q3, const uint8x8_t q4, const uint8x8_t q5, const uint8x8_t q6, const uint8x8_t q7, uint8x8_t *op6, uint8x8_t *op5, uint8x8_t *op4, uint8x8_t *op3, uint8x8_t *op2, uint8x8_t *op1, uint8x8_t *op0, uint8x8_t *oq0, uint8x8_t *oq1, uint8x8_t *oq2, uint8x8_t *oq3, uint8x8_t *oq4, uint8x8_t *oq5, uint8x8_t *oq6) argument 345 apply_15_tap_filter_16( const uint8x16_t flat2, const uint8x16_t p7, const uint8x16_t p6, const uint8x16_t p5, const uint8x16_t p4, const uint8x16_t p3, const uint8x16_t p2, const uint8x16_t p1, const uint8x16_t p0, const uint8x16_t q0, const uint8x16_t q1, const uint8x16_t q2, const uint8x16_t q3, const uint8x16_t q4, const uint8x16_t q5, const uint8x16_t q6, const uint8x16_t q7, uint8x16_t *op6, uint8x16_t *op5, uint8x16_t *op4, uint8x16_t *op3, uint8x16_t *op2, uint8x16_t *op1, uint8x16_t *op0, uint8x16_t *oq0, uint8x16_t *oq1, uint8x16_t *oq2, uint8x16_t *oq3, uint8x16_t *oq4, uint8x16_t *oq5, uint8x16_t *oq6) argument 980 uint8x8_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7, op6, local 997 uint8x16_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7, local 1020 uint8x8_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7, op6, local 1052 uint8x16_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7, local [all...] |
/external/libvpx/libvpx/vpx_dsp/x86/ |
H A D | highbd_loopfilter_sse2.c | 57 __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0; local 213 q6 = _mm_load_si128((__m128i *)(s + 6 * p)); 230 _mm_or_si128(_mm_subs_epu16(q6, q0), _mm_subs_epu16(q0, q6))); 255 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5), _mm_add_epi16(q4, q3)); 281 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6); 347 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6)), 4); 387 q6 = _mm_andnot_si128(flat2, q6); 388 // q6 remain [all...] |
H A D | loopfilter_avx2.c | 379 __m128i q5, q6, q7; local 550 q6 = _mm256_castsi256_si128(q256_6); 553 _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6))); 908 q6 = _mm_andnot_si128(flat2, q6); 910 q6 = _mm_or_si128(flat2_q6, q6); 911 _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
|