| /external/libhevc/common/x86/ |
| ihevc_32x32_itrans_recon_sse42_intr.c | 295 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3);
318 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3);
358 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
374 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff1);
387 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2);
401 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2);
415 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
429 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff3);
446 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff4);
460 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff4)
[all...] |
| ihevc_itrans_recon_32x32_ssse3_intr.c | 304 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3);
327 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3);
368 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
385 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff1);
398 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2);
412 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2);
426 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
440 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff3);
457 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff4);
471 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff4)
[all...] |
| ihevc_16x16_itrans_recon_sse42_intr.c | 256 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
275 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1);
294 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
311 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3);
331 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
347 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1);
365 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
380 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3);
406 m_temp_reg_20 = _mm_madd_epi16(m_temp_reg_0, m_coeff1);
407 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2)
[all...] |
| ihevc_itrans_recon_16x16_ssse3_intr.c | 262 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
281 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1);
300 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
317 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3);
337 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1);
353 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1);
371 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3);
386 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3);
412 m_temp_reg_20 = _mm_madd_epi16(m_temp_reg_0, m_coeff1);
413 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2)
[all...] |
| ihevc_itrans_recon_sse42_intr.c | 887 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1);
888 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2);
905 m_temp_reg_16 = _mm_madd_epi16(m_temp_reg_4, m_coeff1);
906 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_4, m_coeff2);
941 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_60, m_coeff1);
967 m_temp_reg_32 = _mm_madd_epi16(m_temp_reg_60, m_coeff3);
[all...] |
| ihevc_itrans_recon_ssse3_intr.c | [all...] |
| ihevc_inter_pred_filters_ssse3_intr.c | [all...] |
| /external/chromium_org/third_party/libvpx/source/libvpx/vp9/encoder/x86/ |
| vp9_dct_sse2.c | 160 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
161 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
162 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
163 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
201 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
202 const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
203 const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
204 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
312 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
313 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); //
[all...] |
| vp9_dct32x32_sse2.c | 332 const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
333 const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
334 const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
335 const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
336 const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
337 const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
338 const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
339 const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
340 const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
341 const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16)
[all...] |
| /external/libvpx/libvpx/vp9/encoder/x86/ |
| vp9_dct_sse2.c | 129 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
130 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
131 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
132 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
170 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
171 const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
172 const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
173 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
281 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
282 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); //
[all...] |
| vp9_dct_avx2.c | 71 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
72 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
73 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
74 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
180 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
181 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2
182 u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1
183 u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3
216 v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2
217 v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s
[all...] |
| vp9_dct32x32_sse2.c | 329 const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
330 const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
331 const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
332 const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
333 const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
334 const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
335 const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
336 const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
337 const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
338 const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16)
[all...] |
| /hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/encoder/x86/ |
| vp9_dct_sse2.c | 129 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
130 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
131 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
132 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
170 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
171 const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
172 const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
173 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
281 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
282 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); //
[all...] |
| vp9_dct_avx2.c | 71 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
72 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
73 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
74 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
180 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
181 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2
182 u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1
183 u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3
216 v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2
217 v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s
[all...] |
| vp9_dct32x32_sse2.c | 329 const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
330 const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
331 const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
332 const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
333 const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
334 const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
335 const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
336 const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
337 const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
338 const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16)
[all...] |
| /external/chromium_org/third_party/libvpx/source/libvpx/vp9/common/x86/ |
| vp9_idct_intrin_ssse3.c | 72 v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
73 v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
74 v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
75 v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
76 v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
77 v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
78 v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
79 v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
80 v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
81 v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10)
[all...] |
| vp9_idct_intrin_sse2.c | 49 input0 = _mm_madd_epi16(input0, cst);
50 input1 = _mm_madd_epi16(input1, cst);
51 input2 = _mm_madd_epi16(input2, cst);
52 input3 = _mm_madd_epi16(input3, cst);
88 input0 = _mm_madd_epi16(input0, cst);
89 input1 = _mm_madd_epi16(input1, cst);
90 input2 = _mm_madd_epi16(input2, cst);
91 input3 = _mm_madd_epi16(input3, cst);
190 v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
191 v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16)
[all...] |
| /external/libvpx/libvpx/vp9/common/x86/ |
| vp9_idct_intrin_sse2.c | 54 input0 = _mm_madd_epi16(input0, cst);
55 input1 = _mm_madd_epi16(input1, cst);
56 input2 = _mm_madd_epi16(input2, cst);
57 input3 = _mm_madd_epi16(input3, cst);
93 input0 = _mm_madd_epi16(input0, cst);
94 input1 = _mm_madd_epi16(input1, cst);
95 input2 = _mm_madd_epi16(input2, cst);
96 input3 = _mm_madd_epi16(input3, cst);
195 v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
196 v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16)
[all...] |
| /hardware/intel/common/omx-components/videocodec/libvpx_internal/libvpx/vp9/common/x86/ |
| vp9_idct_intrin_sse2.c | 54 input0 = _mm_madd_epi16(input0, cst);
55 input1 = _mm_madd_epi16(input1, cst);
56 input2 = _mm_madd_epi16(input2, cst);
57 input3 = _mm_madd_epi16(input3, cst);
93 input0 = _mm_madd_epi16(input0, cst);
94 input1 = _mm_madd_epi16(input1, cst);
95 input2 = _mm_madd_epi16(input2, cst);
96 input3 = _mm_madd_epi16(input3, cst);
195 v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
196 v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16)
[all...] |
| /frameworks/rs/cpu_ref/ |
| rsCpuIntrinsics_x86.c | 113 o0 = _mm_madd_epi16(_mm_unpacklo_epi16(p0, p1), c0);
114 o1 = _mm_madd_epi16(_mm_unpacklo_epi16(p1, p2), c0);
116 o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p2, p4), c2));
117 o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p3, p5), c2));
119 o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p5, p6), c4));
120 o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p6, p7), c4));
122 o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p8, p9), c6));
123 o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p9, p10), c6));
125 o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p10, _mm_setzero_si128()), c8));
126 o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p11, _mm_setzero_si128()), c8))
[all...] |
| /external/neven/Embedded/common/src/b_BasicEm/ |
| MathSSE2.c | 63 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM1 );
68 m_XMM2 = _mm_madd_epi16( m_XMM2, m_XMM3 );
76 m_XMM4 = _mm_madd_epi16( m_XMM4, m_XMM5 );
79 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM8 );
172 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM2 );
178 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM3 );
276 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM2 );
282 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM3 );
|
| /external/jpeg/ |
| jidctintelsse.c | 132 tmp1 = _mm_madd_epi16(tmp1, * ( __m128i*)table1); /*x2*w13+x0*w12, x2*w9+x0*w8, x2*w5+x0*w4, x2*w1+x0*w0*/ \
133 tmp5 = _mm_madd_epi16(tmp5, * ( __m128i*)table2); \
140 tmp3 = _mm_madd_epi16(tmp3, * ( __m128i*)(table1+16)); /*x3*w29+x1*w28, x3*w25+x1*w24, x3*w21+x1*w20, x3*w17+x1*w16*/ \
141 tmp7 = _mm_madd_epi16(tmp7, * ( __m128i*)(table2+16) ); \
142 row0 = _mm_madd_epi16(row0, * ( __m128i*)(table1+24)); /*x7*w31+x5*w30, x7*w27+x5*w26, x7*w23+x5*w22, x7*w19+x5*w18*/ \
143 row2 = _mm_madd_epi16(row2, * ( __m128i*)(table2+24) ); \
144 tmp2 = _mm_madd_epi16(tmp2, * ( __m128i*)(table1+8) ); /*x6*w15+x4*w14, x6*w11+x4*w10, x6*w7+x4*w6, x6*w3+x4*w2*/ \
145 tmp6 = _mm_madd_epi16(tmp6, * ( __m128i*)(table2+8) ); \
|
| /external/qemu/distrib/jpeg-6b/ |
| jidctintelsse.c | 132 tmp1 = _mm_madd_epi16(tmp1, * ( __m128i*)table1); /*x2*w13+x0*w12, x2*w9+x0*w8, x2*w5+x0*w4, x2*w1+x0*w0*/ \
133 tmp5 = _mm_madd_epi16(tmp5, * ( __m128i*)table2); \
140 tmp3 = _mm_madd_epi16(tmp3, * ( __m128i*)(table1+16)); /*x3*w29+x1*w28, x3*w25+x1*w24, x3*w21+x1*w20, x3*w17+x1*w16*/ \
141 tmp7 = _mm_madd_epi16(tmp7, * ( __m128i*)(table2+16) ); \
142 row0 = _mm_madd_epi16(row0, * ( __m128i*)(table1+24)); /*x7*w31+x5*w30, x7*w27+x5*w26, x7*w23+x5*w22, x7*w19+x5*w18*/ \
143 row2 = _mm_madd_epi16(row2, * ( __m128i*)(table2+24) ); \
144 tmp2 = _mm_madd_epi16(tmp2, * ( __m128i*)(table1+8) ); /*x6*w15+x4*w14, x6*w11+x4*w10, x6*w7+x4*w6, x6*w3+x4*w2*/ \
145 tmp6 = _mm_madd_epi16(tmp6, * ( __m128i*)(table2+8) ); \
|
| /external/chromium_org/third_party/libwebp/dsp/ |
| enc_sse2.c | 394 const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ]
395 const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ]
396 const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
397 const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
434 const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
435 const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
532 const __m128i low_madd0 = _mm_madd_epi16(low0, low0);
533 const __m128i low_madd1 = _mm_madd_epi16(low1, low1);
534 const __m128i low_madd2 = _mm_madd_epi16(low2, low2);
535 const __m128i low_madd3 = _mm_madd_epi16(low3, low3)
[all...] |
| /external/webp/src/dsp/ |
| enc_sse2.c | 394 const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ]
395 const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ]
396 const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
397 const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
434 const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
435 const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
532 const __m128i low_madd0 = _mm_madd_epi16(low0, low0);
533 const __m128i low_madd1 = _mm_madd_epi16(low1, low1);
534 const __m128i low_madd2 = _mm_madd_epi16(low2, low2);
535 const __m128i low_madd3 = _mm_madd_epi16(low3, low3)
[all...] |
