| /external/XNNPACK/src/qu8-requantization/ |
| D | precise-ssse3.c | 73 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
74 const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
75 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
76 const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
77 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
78 const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
79 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
80 const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__ssse3()
|
| D | precise-sse2.c | 78 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
79 const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
80 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
81 const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
82 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
83 const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
84 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
85 const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift); in xnn_qu8_requantize_precise__sse2()
|
| D | precise-sse4.c | 74 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshiftlo); in xnn_qu8_requantize_precise__sse4()
76 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshiftlo); in xnn_qu8_requantize_precise__sse4()
78 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshiftlo); in xnn_qu8_requantize_precise__sse4()
80 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshiftlo); in xnn_qu8_requantize_precise__sse4()
|
| /external/XNNPACK/src/qs8-requantization/ |
| D | precise-ssse3.c | 73 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
74 const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
75 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
76 const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
77 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
78 const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
79 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
80 const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__ssse3()
|
| D | precise-sse2.c | 78 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
79 const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
80 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
81 const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
82 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
83 const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
84 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
85 const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift); in xnn_qs8_requantize_precise__sse2()
|
| D | precise-sse4.c | 74 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshiftlo); in xnn_qs8_requantize_precise__sse4()
76 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshiftlo); in xnn_qs8_requantize_precise__sse4()
78 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshiftlo); in xnn_qs8_requantize_precise__sse4()
80 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshiftlo); in xnn_qs8_requantize_precise__sse4()
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-sse41-c24-acc2.c | 147 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
148 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
149 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
150 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
151 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
152 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
153 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
154 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
155 const __m128i vabsoutGI = _mm_srl_epi64(_mm_add_epi64(vabsprodGI, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
156 const __m128i vabsoutHJ = _mm_srl_epi64(_mm_add_epi64(vabsprodHJ, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
[all …]
|
| D | 7x-minmax-sse41-c16-acc2.c | 124 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
125 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
126 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
127 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
128 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
129 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
130 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
131 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
201 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
202 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c24-acc2.c | 171 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
172 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
173 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
174 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
175 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
176 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
177 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
178 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
179 const __m128i vabsoutGI = _mm_srl_epi64(_mm_add_epi64(vabsprodGI, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
180 const __m128i vabsoutHJ = _mm_srl_epi64(_mm_add_epi64(vabsprodHJ, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
[all …]
|
| D | 7x-minmax-sse2-c24-acc2.c | 178 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
179 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
180 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
181 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
182 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
183 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
184 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
185 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
186 const __m128i vabsoutGI = _mm_srl_epi64(_mm_add_epi64(vabsprodGI, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
187 const __m128i vabsoutHJ = _mm_srl_epi64(_mm_add_epi64(vabsprodHJ, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
[all …]
|
| D | 7x-minmax-sse41-c8-acc2.c | 101 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
102 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
103 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
104 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
168 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
169 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
170 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
171 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 140 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
141 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
142 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
143 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
144 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
145 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
146 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
147 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
234 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
235 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
[all …]
|
| D | 7x-minmax-sse2-c16-acc2.c | 145 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
146 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
147 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
148 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
149 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
150 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
151 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
152 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
242 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
243 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c8-acc2.c | 109 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
110 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
111 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
112 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
189 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
190 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
191 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
192 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 253 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
254 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
255 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
256 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
257 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
258 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
259 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
260 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
331 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
332 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
[all …]
|
| D | 7x-minmax-sse2-c8-acc2.c | 112 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
113 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
114 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
115 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
195 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
196 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
197 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
198 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 386 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
387 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
388 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
389 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
390 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
391 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
392 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
393 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
394 const __m128i vabsoutGI = _mm_srl_epi64(_mm_add_epi64(vabsprodGI, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
395 const __m128i vabsoutHJ = _mm_srl_epi64(_mm_add_epi64(vabsprodHJ, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 196 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
197 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
198 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
199 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
264 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
265 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
266 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
267 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 301 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
302 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
303 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
304 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
305 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
306 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
307 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
308 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
396 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
397 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
[all …]
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 223 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
224 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
225 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
226 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
307 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
308 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
309 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
310 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 220 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
221 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
222 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
223 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
301 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
302 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
303 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
304 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 306 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
307 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
308 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
309 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
310 const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
311 const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
312 const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
313 const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
404 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
405 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
[all …]
|
| /external/XNNPACK/src/qu8-gavgpool/ |
| D | 7x-minmax-sse2-c8.c | 104 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
105 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
106 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
107 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
173 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
174 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
175 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
176 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
|
| D | 7p7x-minmax-sse2-c8.c | 198 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
199 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
200 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
201 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
269 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
270 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
271 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
272 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
|
| /external/XNNPACK/src/qu8-avgpool/ |
| D | 9x-minmax-sse2-c8.c | 160 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
161 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
162 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
163 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
235 …const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
236 …const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
237 …const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
238 …const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shif… in xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8()
|