Home
last modified time | relevance | path

Searched refs:vzGHIJ (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/f32-sigmoid/gen/
Dpsimd-p5-div-x20.c61 const psimd_f32 vzGHIJ = psimd_abs_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20() local
73 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
96 psimd_f32 vtGHIJ = psimd_qfma_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
159 vfGHIJ = psimd_andnotmask_f32(vzGHIJ > vdenorm_cutoff, vfGHIJ); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
Dpsimd-p5-div-x24.c62 const psimd_f32 vzGHIJ = psimd_abs_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24() local
75 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
101 psimd_f32 vtGHIJ = psimd_qfma_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
173 vfGHIJ = psimd_andnotmask_f32(vzGHIJ > vdenorm_cutoff, vfGHIJ); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
Dsse41-p5-div-x20.c61 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20() local
73 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
96 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
165 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
Dsse2-p5-div-x20.c61 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20() local
73 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
96 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
165 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
Dsse41-p5-div-x24.c62 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24() local
75 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
101 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
180 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
Dsse2-p5-div-x24.c62 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24() local
75 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
101 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
180 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
Dneonfma-rr1-p5-div-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() local
70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
92 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
Dneonfma-rr1-p5-nr1recps1fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local
70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
92 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
Dneonfma-rr1-p5-nr2recps-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local
70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
92 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
Dneonfma-rr1-p5-div-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() local
72 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
97 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
Dneon-rr2-p5-nr2recps-x20.c60 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
72 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
95 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
Dneonfma-rr1-p5-nr2fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local
70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
92 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
Dneonfma-rr1-p5-nr1recps1fma-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local
72 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
97 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
Dneonfma-rr1-p5-nr2recps-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local
72 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
97 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
Dneonfma-rr1-lut2048-p1-div-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
Dneonfma-rr1-p5-nr2fma-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local
72 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
97 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
Dneonfma-rr1-lut64-p2-div-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
Dneon-rr2-p5-nr2recps-x24.c61 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
74 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
100 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
Dneonfma-rr1-lut64-p2-nr1recps1fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
Dneonfma-rr1-lut64-p2-nr2fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
Dneonfma-rr1-lut2048-p1-nr2recps-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
Dneonfma-rr1-lut64-p2-div-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
Dneonfma-rr1-lut2048-p1-nr2fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
Dneonfma-rr1-lut2048-p1-nr1recps1fma-x20.c58 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
Dneonfma-rr1-lut2048-p1-div-x24.c59 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()

12