Home
last modified time | relevance | path

Searched refs:vf0123 (Results 1 – 25 of 158) sorted by relevance

1234567

/external/XNNPACK/src/f32-sigmoid/gen/
Dpsimd-p5-div-x8.c108 psimd_f32 vf0123 = psimd_div_f32(ve0123, psimd_add_f32(ve0123, vone)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() local
113 vf0123 = psimd_andnotmask_f32(vz0123 > vdenorm_cutoff, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
117 vf0123 = psimd_signblend_f32(vx0123, vf0123, psimd_sub_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
120 psimd_store_f32(y, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
Dneonfma-rr1-p5-div-x8.c104 float32x4_t vf0123 = vdivq_f32(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() local
109 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
116 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
119 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
Dsse41-p5-div-x8.c111 __m128 vf0123 = _mm_div_ps(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local
116 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
120 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
123 _mm_storeu_ps(y, vf0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
Dpsimd-p5-div-x12.c121 psimd_f32 vf0123 = psimd_div_f32(ve0123, psimd_add_f32(ve0123, vone)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
127 vf0123 = psimd_andnotmask_f32(vz0123 > vdenorm_cutoff, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
132 vf0123 = psimd_signblend_f32(vx0123, vf0123, psimd_sub_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
136 psimd_store_f32(y, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
Dneonfma-rr1-p5-nr2recps-x8.c116 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() local
121 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8()
128 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8()
131 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8()
Dsse2-p5-div-x8.c111 __m128 vf0123 = _mm_div_ps(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() local
116 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
123 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
126 _mm_storeu_ps(y, vf0123); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
Dneonfma-rr1-p5-nr1recps1fma-x8.c116 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() local
121 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8()
128 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8()
131 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8()
Dneonfma-rr1-p5-div-x12.c117 float32x4_t vf0123 = vdivq_f32(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() local
123 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
132 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
136 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
Dneonfma-rr1-p5-nr2fma-x8.c116 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8() local
121 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8()
128 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8()
131 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8()
Dneon-rr2-p5-nr2recps-x8.c122 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() local
127 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
134 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
137 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
Dneonfma-rr1-lut64-p2-div-x8.c130 float32x4_t vf0123 = vdivq_f32(vy0123, vd0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8() local
135 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
142 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
145 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
Dneonfma-rr1-lut2048-p1-div-x8.c126 float32x4_t vf0123 = vdivq_f32(vy0123, vd0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8() local
131 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
138 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
141 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
Dpsimd-p5-div-x16.c134 psimd_f32 vf0123 = psimd_div_f32(ve0123, psimd_add_f32(ve0123, vone)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16() local
141 vf0123 = psimd_andnotmask_f32(vz0123 > vdenorm_cutoff, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
147 vf0123 = psimd_signblend_f32(vx0123, vf0123, psimd_sub_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
152 psimd_store_f32(y, vf0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
Dsse41-p5-div-x12.c125 __m128 vf0123 = _mm_div_ps(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12() local
131 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
136 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
140 _mm_storeu_ps(y, vf0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
Dneonfma-rr1-lut2048-p1-nr1recps1fma-x8.c138 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8() local
143 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
150 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
153 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
Dneonfma-rr1-lut2048-p1-nr2recps-x8.c138 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() local
143 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
150 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
153 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
Dneonfma-rr1-p5-div-x16.c130 float32x4_t vf0123 = vdivq_f32(ve0123, vd0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() local
137 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
148 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
153 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
Dneonfma-rr1-lut64-p2-nr1recps1fma-x8.c142 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8() local
147 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
154 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
157 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
Dneonfma-rr1-p5-nr2fma-x12.c132 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() local
138 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12()
147 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12()
151 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12()
Dneonfma-rr1-p5-nr2recps-x12.c132 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() local
138 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12()
147 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12()
151 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12()
Dneon-rr2-p5-nr2recps-x12.c139 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local
145 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
154 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
158 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
Dneonfma-rr1-lut64-p2-nr2recps-x8.c142 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8() local
147 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
154 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
157 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
Dneonfma-rr1-lut2048-p1-nr2fma-x8.c138 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8() local
143 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8()
150 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8()
153 vst1q_f32(y, vf0123); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dpsimd-p5-x4.c83 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() local
87 vf0123 = psimd_andnotmask_f32(vx0123 < vdenorm_cutoff, vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
90 psimd_store_f32(output, vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
94 vacc0 = psimd_add_f32(vacc0, vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
Dsse2-p5-x4.c83 __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() local
87 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vx0123, vdenorm_cutoff), vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
90 _mm_storeu_ps(output, vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
94 vacc0 = _mm_add_ps(vacc0, vf0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()

1234567