Home
last modified time | relevance | path

Searched refs:vfKLMN (Results 1 – 18 of 18) sorted by relevance

/external/XNNPACK/src/f32-sigmoid/gen/
Dpsimd-p5-div-x24.c165 psimd_f32 vfKLMN = psimd_div_f32(veKLMN, psimd_add_f32(veKLMN, vone)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24() local
174 vfKLMN = psimd_andnotmask_f32(vzKLMN > vdenorm_cutoff, vfKLMN); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
182 vfKLMN = psimd_signblend_f32(vxKLMN, vfKLMN, psimd_sub_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
189 psimd_store_f32(y + 20, vfKLMN); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
Dneonfma-rr1-p5-div-x24.c161 float32x4_t vfKLMN = vdivq_f32(veKLMN, vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() local
170vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
185 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
192 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
Dsse41-p5-div-x24.c172 __m128 vfKLMN = _mm_div_ps(veKLMN, vdKLMN); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24() local
181 vfKLMN = _mm_andnot_ps(_mm_cmplt_ps(vzKLMN, vdenorm_cutoff), vfKLMN); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
189 vfKLMN = _mm_blendv_ps(_mm_sub_ps(vone, vfKLMN), vfKLMN, vxKLMN); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
196 _mm_storeu_ps(y + 20, vfKLMN); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
Dneonfma-rr1-p5-nr1recps1fma-x24.c185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local
194vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
209 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
216 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
Dneonfma-rr1-p5-nr2recps-x24.c185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local
194vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
209 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
216 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
Dsse2-p5-div-x24.c172 __m128 vfKLMN = _mm_div_ps(veKLMN, vdKLMN); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24() local
181 vfKLMN = _mm_andnot_ps(_mm_cmplt_ps(vzKLMN, vdenorm_cutoff), vfKLMN); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
196 vfKLMN = _mm_or_ps(_mm_and_ps(vfKLMN, vmKLMN), _mm_andnot_ps(vmKLMN, _mm_sub_ps(vone, vfKLMN))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
203 _mm_storeu_ps(y + 20, vfKLMN); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
Dneonfma-rr1-p5-nr2fma-x24.c185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local
194vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
209 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
216 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
Dneon-rr2-p5-nr2recps-x24.c195 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
204vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
219 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
226 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
Dneonfma-rr1-lut64-p2-div-x24.c211 float32x4_t vfKLMN = vdivq_f32(vyKLMN, vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local
220vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
235 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
242 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
Dneonfma-rr1-lut2048-p1-div-x24.c203 float32x4_t vfKLMN = vdivq_f32(vyKLMN, vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local
212vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
227 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
234 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
Dneonfma-rr1-lut64-p2-nr1recps1fma-x24.c235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() local
244vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
259 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
266 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
Dneonfma-rr1-lut2048-p1-nr2fma-x24.c227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local
236vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
251 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
258 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
Dneonfma-rr1-lut64-p2-nr2fma-x24.c235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local
244vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
259 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
266 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
Dneonfma-rr1-lut2048-p1-nr2recps-x24.c227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() local
236vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
251 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
258 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
Dneonfma-rr1-lut64-p2-nr2recps-x24.c235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() local
244vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
259 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
266 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
Dneon-rr2-lut64-p2-nr2recps-x24.c245 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24() local
254vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24()
269 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24()
276 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24()
Dneon-rr2-lut2048-p1-nr2recps-x24.c237 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24() local
246vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
261 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
268 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
Dneonfma-rr1-lut2048-p1-nr1recps1fma-x24.c227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24() local
236vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
251 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
258 vst1q_f32(y, vfKLMN); y += 4; in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()