Home
last modified time | relevance | path

Searched refs:vt89AB (Results 1 – 25 of 126) sorted by relevance

123456

/external/XNNPACK/src/f32-sigmoid/gen/
Dpsimd-p5-div-x12.c84 psimd_f32 vt89AB = psimd_qfma_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
88 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
94 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vt89AB, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
98 vp89AB = psimd_qfma_f32(vc3, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
102 vp89AB = psimd_qfma_f32(vc2, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
106 vp89AB = psimd_qfma_f32(vc1, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
114 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
118 const psimd_f32 ve89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
Dpsimd-p5-div-x16.c89 psimd_f32 vt89AB = psimd_qfma_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16() local
94 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
101 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vt89AB, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
106 vp89AB = psimd_qfma_f32(vc3, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
111 vp89AB = psimd_qfma_f32(vc2, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
116 vp89AB = psimd_qfma_f32(vc1, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
125 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
130 const psimd_f32 ve89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
Dsse41-p5-div-x12.c84 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12() local
88 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
93 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
97 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
101 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
105 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
113 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
117 __m128 ve89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
Dneon-rr2-p5-nr2recps-x12.c83 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local
87 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
92 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
96 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
100 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
104 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
112 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
116 float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dpsimd-p5-x12.c76 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local
80 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
85 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
89 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
93 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
97 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
105 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
109 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
Dpsimd-p5-x12-acc2.c77 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local
81 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
86 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
90 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
94 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
98 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
106 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
110 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
Dpsimd-p5-x12-acc3.c78 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local
82 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
87 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
91 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
95 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
99 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
107 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
111 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
Dsse2-p5-x12.c76 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local
80 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
85 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
89 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
93 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
97 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
105 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
109 __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
Dneon-p5-x12.c80 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
84 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
89 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
93 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
97 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
101 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
109 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
113 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
Dneonfma-p5-x12.c79 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
83 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
88 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
92 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
96 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
100 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
108 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
112 float32x4_t vf89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
Dneonfma-p5-x12-acc2.c80 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
84 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
89 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
93 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
97 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
101 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
109 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
113 float32x4_t vf89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
Dneon-p5-x12-acc2.c81 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local
85 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
90 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
94 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
98 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
102 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
110 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
114 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
Dpsimd-p5-x16-acc2.c82 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local
87 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
93 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
98 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
103 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
108 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
117 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
122 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
Dsse2-p5-x12-acc3.c78 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3() local
82 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
87 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
91 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
95 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
99 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
107 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
111 __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
Dpsimd-p5-x16.c81 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16() local
86 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
92 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
97 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
102 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
107 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
116 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
121 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
Dpsimd-p5-x16-acc4.c84 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4() local
89 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
95 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
100 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
105 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
110 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
119 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
124 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
Dneonfma-p5-x12-acc3.c81 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local
85 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
90 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
94 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
98 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
102 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
110 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
114 float32x4_t vf89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
Dneon-p5-x12-acc3.c82 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3() local
86 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
91 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
95 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
99 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
103 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
111 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
115 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
Dsse2-p5-x12-acc2.c77 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2() local
81 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
86 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
90 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
94 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
98 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
106 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
110 __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
Dneonfma-p5-x16-acc4.c87 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local
92 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
98 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
103 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
108 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
113 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
122 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
127 float32x4_t vf89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
Dpsimd-p5-x20-acc2.c87 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2() local
93 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
100 psimd_f32 vp89AB = psimd_qfma_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
106 vp89AB = psimd_qfma_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
112 vp89AB = psimd_qfma_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
118 vp89AB = psimd_qfma_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
128 vt89AB = psimd_mul_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
134 psimd_f32 vf89AB = psimd_qfma_f32(vs89AB, vt89AB, vp89AB); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
Dneon-p5-x16-acc2.c86 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local
91 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
97 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
102 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
107 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
112 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
121 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
126 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
Dsse2-p5-x16.c81 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16() local
86 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
92 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
97 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
102 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
107 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
116 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
121 __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
Dneon-p5-x16.c85 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local
90 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
96 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
101 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
106 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
111 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
120 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
125 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
Dneonfma-p5-x16.c84 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local
89 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
95 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
100 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
105 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
110 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
119 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
124 float32x4_t vf89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()

123456