Home
last modified time | relevance | path

Searched refs:vs5 (Results 1 – 25 of 112) sorted by relevance

12345

/external/XNNPACK/src/f32-velu/gen/
Dvelu-scalar-rr2-lut16-p3-x6.c95 float vs5 = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx5] + ven5); in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6() local
124 vs5 = 0.0f; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
152 vt5 *= vs5; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
153 vs5 -= vone; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
172 const float ve5 = (vp5 + vs5) * valpha; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
Dvelu-scalar-rr2-p6-x6.c75 float vs5 = fp32_from_bits(fp32_to_bits(vn5) << 23); in xnn_f32_velu_ukernel__scalar_rr2_p6_x6() local
113 vs5 = 0.0f; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
162 vt5 *= vs5; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
163 vs5 -= vone; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
182 const float ve5 = (vp5 + vs5) * valpha; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
Dvelu-wasm-rr2-p6-x6.c75 float vs5 = fp32_from_bits(fp32_to_bits(vn5) << 23); in xnn_f32_velu_ukernel__wasm_rr2_p6_x6() local
138 vt5 *= vs5; in xnn_f32_velu_ukernel__wasm_rr2_p6_x6()
139 vs5 -= vone; in xnn_f32_velu_ukernel__wasm_rr2_p6_x6()
158 const float ve5 = (vp5 + vs5) * valpha; in xnn_f32_velu_ukernel__wasm_rr2_p6_x6()
Dvelu-wasm-rr2-lut16-p3-x6.c95 float vs5 = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx5] + ven5); in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6() local
128 vt5 *= vs5; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6()
129 vs5 -= vone; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6()
148 const float ve5 = (vp5 + vs5) * valpha; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6()
Dvelu-avx512f-rr1-lut16-p3-perm-x96.c85 __m512 vs5 = _mm512_castsi512_ps(_mm512_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96() local
113 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
120 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
140 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
Dvelu-avx512f-rr1-lut16-p3-perm-x112.c90 __m512 vs5 = _mm512_castsi512_ps(_mm512_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112() local
122 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
131 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
153 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
Dvelu-avx2-rr1-p6-x48.c72 __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x48() local
121 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x48()
133 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_p6_x48()
146 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x48()
Dvelu-avx2-rr1-lut16-p3-gather-x48.c98 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48() local
119 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48()
131 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48()
144 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x48()
Dvelu-avx2-rr1-lut8-p4-perm-x48.c90 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48() local
118 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48()
130 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48()
143 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48()
Dvelu-avx2-rr1-lut4-p4-perm-x48.c90 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48() local
118 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48()
130 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48()
143 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48()
Dvelu-avx512f-rr1-p6-x96.c74 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96() local
123 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
130 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
150 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
Dvelu-avx2-rr1-lut8-p4-perm-x56.c96 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56() local
128 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56()
142 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56()
157 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56()
Dvelu-avx2-rr1-lut4-p4-perm-x56.c96 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56() local
128 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56()
142 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56()
157 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56()
Dvelu-avx512f-rr1-p6-x112.c77 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112() local
133 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
142 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
164 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
Dvelu-avx2-rr1-p6-x56.c75 __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56() local
131 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
145 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
160 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56()
Dvelu-avx2-rr1-lut16-p3-gather-x56.c105 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x56() local
129 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x56()
143 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x56()
158 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x56()
Dvelu-avx2-rr1-p6-x64.c78 __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64() local
141 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
157 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
174 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64()
Dvelu-avx512f-rr1-lut16-p3-perm-x128.c95 __m512 vs5 = _mm512_castsi512_ps(_mm512_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128() local
131 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
142 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
166 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
Dvelu-avx512f-rr1-p6-x128.c80 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128() local
143 vt5 = _mm512_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
154 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
178 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
Dvelu-avx2-rr1-lut16-p3-gather-x64.c112 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64() local
139 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
155 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
172 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x64()
Dvelu-avx2-rr1-lut8-p4-perm-x64.c102 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64() local
138 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
154 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
171 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x64()
Dvelu-avx-rr2-lut4-p4-perm-x48.c120 __m256 vs5 = _mm256_mul_ps(vl5, ven5); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48() local
153 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48()
154 vs5 = _mm256_sub_ps(vs5, vone); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48()
173 const __m256 ve5 = _mm256_mul_ps(_mm256_add_ps(vp5, vs5), valpha); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48()
Dvelu-avx2-rr1-lut4-p4-perm-x64.c102 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64() local
138 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
154 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
171 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x64()
Dvelu-avx-rr2-p6-x48.c94 __m256 vs5 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs5_lo), vs5_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_p6_x48() local
148 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx_rr2_p6_x48()
149 vs5 = _mm256_sub_ps(vs5, vone); in xnn_f32_velu_ukernel__avx_rr2_p6_x48()
168 const __m256 ve5 = _mm256_mul_ps(_mm256_add_ps(vp5, vs5), valpha); in xnn_f32_velu_ukernel__avx_rr2_p6_x48()
Dvelu-avx2-rr1-p6-x72.c81 __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72() local
151 vt5 = _mm256_mul_ps(vt5, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()
169 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()
188 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x72()

12345