/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp8/common/arm/neon/ |
H A D | idct_dequant_full_2x_neon.asm | 28 vld1.16 {q4, q5}, [r0] ; r q 47 vmul.i16 q5, q5, q1 52 ; q4: l4r4 q5: l12r12 62 vqdmulh.s16 q7, q5, d0[2] 64 vqdmulh.s16 q9, q5, d0[0] 81 ; q5: 12 + 12 * cospi : c1/temp2 83 vqadd.s16 q5, q5, q9 87 vqsub.s16 q2, q6, q5 [all...] |
H A D | sixtappredict4x4_neon.asm | 64 vld1.u8 {q5}, [r0], r1 89 vmov q6, q5 101 vshr.u64 q5, q6, #32 115 vshr.u64 q5, q6, #24 130 vld1.u8 {q5}, [r0], r1 155 vmov q6, q5 169 vshr.u64 q5, q6, #32 189 vshr.u64 q5, q6, #24 210 vld1.s32 {q5, q6}, [r3] ;load second_pass filter 217 vabs.s32 q7, q5 [all...] |
H A D | vp8_subpixelvariance16x16s_neon.asm | 57 vext.8 q5, q4, q5, #1 63 vrhadd.u8 q2, q4, q5 67 vsubl.u8 q5, d1, d23 81 vpadal.s16 q8, q5 113 vmull.s32 q5, d0, d0 150 vld1.8 {q5}, [r2], r3 209 vmull.s32 q5, d0, d0 255 vext.8 q5, q4, q5, # [all...] |
H A D | sixtappredict8x8_neon.asm | 74 vld1.u8 {q5}, [r0], r1 135 vmull.u8 q5, d30, d3 142 vqadd.s16 q9, q5 155 vld1.u8 {q5}, [r0], r1 165 ;vld1.u8 {q5}, [r0], r1 231 vmull.u8 q5, d29, d3 237 vqadd.s16 q10, q5 250 vld1.s32 {q5, q6}, [r3] ;load second_pass filter 255 vabs.s32 q7, q5 274 vmull.u8 q5, d2 [all...] |
H A D | sixtappredict8x4_neon.asm | 71 vld1.u8 {q5}, [r0], r1 131 vmull.u8 q5, d30, d3 136 vqadd.s16 q9, q5 148 vld1.u8 {q5}, [r0], r1 218 vmull.u8 q5, d29, d3 224 vqadd.s16 q10, q5 239 vld1.s32 {q5, q6}, [r3] ;load second_pass filter 242 vabs.s32 q7, q5 256 vmull.u8 q5, d24, d0 261 vmlsl.u8 q5, d2 [all...] |
H A D | buildintrapredictorsmby_neon.asm | 245 vmull.u8 q5, d17, d0 248 vsub.s16 q5, q5, q7 265 vqadd.s16 q9, q0, q5 268 vqadd.s16 q11, q1, q5 271 vqadd.s16 q13, q2, q5 274 vqadd.s16 q15, q3, q5 528 vmull.u8 q5, d17, d0 531 vsub.s16 q5, q5, q [all...] |
H A D | sixtappredict16x16_neon.asm | 169 vmull.u8 q5, d29, d3 173 vqadd.s16 q10, q5 203 vld1.s32 {q5, q6}, [r3] ;load second_pass filter 206 vabs.s32 q7, q5 234 vmull.u8 q5, d20, d0 239 vmlsl.u8 q5, d21, d1 244 vmlsl.u8 q5, d24, d4 249 vmlal.u8 q5, d22, d2 254 vmlal.u8 q5, d25, d5 266 vqadd.s16 q9, q5 [all...] |
H A D | vp8_subpixelvariance8x8_neon.asm | 74 vld1.u8 {q5}, [r0], r1 117 vmull.u8 q5, d26, d0 126 vmlal.u8 q5, d27, d1 135 vqrshrn.u16 d26, q5, #7 173 vsubl.u8 q5, d23, d1 182 vpadal.s16 q8, q5 207 vmull.s32 q5, d0, d0
|
H A D | dequant_idct_neon.c | 24 int16x8_t q1, q2, q3, q4, q5, q6; local 39 q5 = vld1q_s16(dq); 54 vreinterpretq_u16_s16(q5)));
|
H A D | variance_neon.asm | 84 vmull.s32 q5, d0, d0 146 vmull.s32 q5, d0, d0 201 vmull.s32 q5, d0, d0 266 vmull.s32 q5, d0, d0
|
H A D | vp8_subpixelvariance16x16_neon.asm | 204 vmull.u8 q5, d26, d0 213 vmlal.u8 q5, d28, d1 224 vqrshrn.u16 d6, q5, #7 331 vmull.u8 q5, d26, d0 340 vmlal.u8 q5, d28, d1 349 vqrshrn.u16 d6, q5, #7 411 vmull.s32 q5, d0, d0
|
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/common/arm/neon/ |
H A D | vp9_idct32x32_add_neon.asm | 180 vrshr.s16 q5, q5, #6 185 vaddw.u8 q5, q5, d5 190 vqmovun.s16 d5, q5 214 vrshr.s16 q5, q5, #6 219 vaddw.u8 q5, q5, d5 224 vqmovun.s16 d5, q5 [all...] |
H A D | vp9_iht8x8_add_neon.asm | 135 vmull.s16 q5, d26, d2 143 vmlsl.s16 q5, d22, d3 151 vqrshrn.s32 d10, q5, #14 ; >> 14 241 vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5] 242 vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5] 273 vadd.s16 q10, q2, q5 ; output[2] = step1[2] + step1[5]; 276 vsub.s16 q13, q2, q5 ; output[5] = step1[2] - step1[5]; 310 vmull.s16 q5, d22, d30 318 vmlal.s16 q5, d24, d31 326 vadd.s32 q11, q1, q5 [all...] |
H A D | vp9_idct16x16_add_neon.asm | 92 vmull.s16 q5, d18, d1 100 vmlal.s16 q5, d30, d0 111 vqrshrn.s32 d14, q5, #14 ; >> 14 206 vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5]; 207 vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5]; 250 vadd.s16 q10, q2, q5 ; step2[2] = step1[2] + step1[5]; 253 vsub.s16 q13, q2, q5 ; step2[5] = step1[2] - step1[5]; 376 vmull.s16 q5, d25, d31 384 vmlal.s16 q5, d23, d30 395 vqrshrn.s32 d13, q5, #1 [all...] |
/hardware/samsung_slsi/exynos5/libswconverter/ |
H A D | csc_ARGB8888_to_YUV420SP_NEON.s | 39 @q5: temp3, output 79 vand.u16 q5,#0x00FF @G 85 vmls.u16 q8,q5,q12 @-(74 * G[k]) 90 vmls.u16 q7,q5,q14 @q0:U -(94 * G[k]) @128<<6+ 32 + v>>2 107 vmla.u16 q7,q5,q15 @q0 += 129 *G[k] 121 vshr.u16 q5,q5,#8 @G 125 vmla.u16 q0,q5,q15 @q0 += 129 *G[k] 152 vand.u16 q5,#0x00FF @G 158 vmla.u16 q7,q5,q1 [all...] |
H A D | csc_interleave_memcpy_neon.s | 86 vld1.8 {q5}, [r12]! 95 vst2.8 {q4, q5}, [r10]!
|
H A D | csc_tiled_to_linear_uv_neon.s | 90 vld1.8 {q4, q5}, [r8]! 99 vst1.8 {q5}, [r7], r2
|
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/common/mips/dspr2/ |
H A D | vp9_loopfilter_macros_dspr2.h | 374 "preceu.ph.qbl %[q5_l], %[q5] \n\t" \ 383 [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6), [q7] "r" (q7) \ 414 "preceu.ph.qbr %[q5_r], %[q5] \n\t" \ 423 [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6), [q7] "r" (q7) \ 455 "precr.qb.ph %[q5], %[q5_l], %[q5_r] \n\t" \ 461 [q5] "=&r" (q5),[q6] "=&r" (q6) \
|
H A D | vp9_loopfilter_filters_dspr2.h | 537 const uint32_t q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7; local 545 /* addition of p6,p5,p4,p3,p2,p1,p0,q0,q1,q2,q3,q4,q5,q6 558 "addu.ph %[add_p6toq6], %[add_p6toq6], %[q5] \n\t" 566 [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6), 581 "subu.ph %[res_op6], %[res_op6], %[q5] \n\t" 595 "subu.ph %[res_op5], %[res_op5], %[q5] \n\t" 607 "subu.ph %[res_op4], %[res_op4], %[q5] \n\t" 617 "subu.ph %[res_op3], %[res_op3], %[q5] \n\t" 627 "subu.ph %[res_op2], %[res_op2], %[q5] \ [all...] |
H A D | vp9_mblpf_vert_loopfilter_dspr2.c | 34 uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7; local 90 "lw %[q5], +4(%[s3]) \n\t" 95 [q5] "=&r" (q5), [q4] "=&r" (q4) 213 /* transpose q4, q5, q6, q7 218 q5 q5_0 q5_1 q5_2 q5_3 225 q5 q4_1 q5_1 q26_1 q7_1 231 "precrq.qb.ph %[prim3], %[q5], %[q4] \n\t" 232 "precr.qb.ph %[prim4], %[q5], %[q4] \n\t" 240 "precrq.ph.w %[q5], [all...] |
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp8/encoder/arm/neon/ |
H A D | fastquantizeb_neon.asm | 38 vabs.s16 q5, q1 50 vadd.s16 q5, q7 55 vqdmulh.s16 q5, q9 64 veor.s16 q5, q3 75 vshr.s16 q5, #1 80 vsub.s16 q5, q3 90 vst1.s16 {q4, q5}, [r7] ; store: qcoeff = x1 99 vmul.s16 q3, q7, q5 113 vtst.16 q15, q5, q8 ; non-zero element is set to all 1
|
H A D | vp8_memcpy_neon.asm | 36 vld1.8 {q4, q5}, [r1]! 39 vst1.8 {q4, q5}, [r0]!
|
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/common/ |
H A D | vp9_loopfilter_filters.c | 246 q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7; local 260 q0 + q1 + q2 + q3 + q4 + q5, 4); 262 q0 + q1 + q2 + q3 + q4 + q5 + q6, 4); 264 q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); 266 q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4); 268 q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4); 270 q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4); 272 q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4); 274 q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4); 276 q0 + q1 + q2 + q3 + q4 + q5 [all...] |
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vpx_scale/arm/neon/ |
H A D | vp8_vpxyv12_copysrcframe_func_neon.asm | 54 vld1.8 {q4, q5}, [r10]! 64 vst1.8 {q4, q5}, [r11]! 168 vld1.8 {q4, q5}, [r10]! 174 vst1.8 {q4, q5}, [r11]!
|
/hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/common/x86/ |
H A D | vp9_loopfilter_intrin_avx2.c | 403 __m128i q5, q6, q7; local 538 q5 = _mm_loadu_si128((__m128i *) (s + 5 * p)); 541 _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5))); 589 q256_5 = _mm256_cvtepu8_epi16(q5); 924 q5 = _mm_andnot_si128(flat2, q5); 926 q5 = _mm_or_si128(flat2_q5, q5); 927 _mm_storeu_si128((__m128i *) (s + 5 * p), q5); [all...] |