Home
last modified time | relevance | path

Searched refs:float32x4_t (Results 1 – 25 of 800) sorted by relevance

12345678910>>...32

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-neonfma-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
170 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
171 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
172 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
173 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
177 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
178 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
[all …]
Dup8x25-minmax-neon-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
170 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
171 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
172 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
173 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
177 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
178 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
[all …]
Dup8x25-minmax-neonfma.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
170 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
171 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
172 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
173 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
177 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
178 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
[all …]
Dup8x25-minmax-neon.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
170 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
171 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
172 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
173 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
177 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
178 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
[all …]
Dup8x9-minmax-neonfma-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
87 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
90 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
91 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
92 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
93 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
97 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
98 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
[all …]
Dup8x9-minmax-neon-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
87 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
90 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
91 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
92 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
93 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
97 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
98 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
[all …]
Dup8x9-minmax-neon.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
87 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
90 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
91 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
92 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
93 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
97 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
98 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
[all …]
Dup8x9-minmax-neonfma.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
87 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
90 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
91 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
92 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
93 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
97 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
98 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
[all …]
Dup4x25-minmax-neonfma-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
169 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
170 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
173 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
174 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
177 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
178 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
[all …]
Dup4x25-minmax-neon-acc2.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
169 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
170 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
173 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
174 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
177 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
178 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
[all …]
Dup4x25-minmax-neon.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
169 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
170 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
173 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
174 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
177 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
178 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
181 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
[all …]
Dup4x25-minmax-neonfma.c32 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
33 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
169 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
170 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
173 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
174 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
177 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
178 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
181 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
[all …]
/external/XNNPACK/src/f32-vsqrt/gen/
Dneonfma-nr1rsqrts1fma1adj-x40.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
37 const float32x4_t vxSTUV = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
38 const float32x4_t vxWXYZ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
[all …]
Dneonfma-nr1rsqrts1fma1adj-x36.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
37 const float32x4_t vxSTUV = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
38 const float32x4_t vxWXYZ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
[all …]
Dneonfma-nr1rsqrts1fma1adj-x32.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
37 const float32x4_t vxSTUV = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
39 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
[all …]
Dneonfma-nr1rsqrts1fma1adj-x28.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
38 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
39 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
[all …]
Dneonfma-nr1rsqrts1fma1adj-x24.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
37 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
38 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
39 float32x4_t vrsqrtx89AB = vrsqrteq_f32(vx89AB); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
[all …]
Dneonfma-nr2fma1adj-x40.c28 const float32x4_t vhalf = vmovq_n_f32(0.5f); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
37 const float32x4_t vxSTUV = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
38 const float32x4_t vxWXYZ = vld1q_f32(x); x += 4; in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
[all …]
/external/XNNPACK/src/f32-gavgpool/
D7p7x-minmax-neon-c4.c39 const float32x4_t vi0 = vld1q_f32(i0); i0 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
40 const float32x4_t vi1 = vld1q_f32(i1); i1 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
41 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
42 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
43 const float32x4_t vi4 = vld1q_f32(i4); i4 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
44 const float32x4_t vi5 = vld1q_f32(i5); i5 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
45 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
47 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
48 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
49 const float32x4_t vsum45 = vaddq_f32(vi4, vi5); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
[all …]
/external/XNNPACK/src/f32-ibilinear/gen/
Dneonfma-c8.c40 const float32x4_t valphah = vdupq_lane_f32(valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8()
41 const float32x4_t valphav = vdupq_lane_f32(valphahv, 1); in xnn_f32_ibilinear_ukernel__neonfma_c8()
46 const float32x4_t vtl0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
47 const float32x4_t vtr0123 = vld1q_f32(i1); i1 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
48 const float32x4_t vbl0123 = vld1q_f32(i2); i2 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
49 const float32x4_t vbr0123 = vld1q_f32(i3); i3 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
50 const float32x4_t vtl4567 = vld1q_f32(i0); i0 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
51 const float32x4_t vtr4567 = vld1q_f32(i1); i1 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
52 const float32x4_t vbl4567 = vld1q_f32(i2); i2 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
53 const float32x4_t vbr4567 = vld1q_f32(i3); i3 += 4; in xnn_f32_ibilinear_ukernel__neonfma_c8()
[all …]
/external/XNNPACK/src/f32-maxpool/
D9p8x-minmax-neon-c4.c28 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
29 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
78 const float32x4_t vi0 = vld1q_f32(i0); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
80 const float32x4_t vi1 = vld1q_f32(i1); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
82 const float32x4_t vi2 = vld1q_f32(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
84 const float32x4_t vi3 = vld1q_f32(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
86 const float32x4_t vi4 = vld1q_f32(i4); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
88 const float32x4_t vi5 = vld1q_f32(i5); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
90 const float32x4_t vi6 = vld1q_f32(i6); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
92 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
[all …]
/external/webrtc/common_audio/third_party/ooura/fft_size_128/
Dooura_fft_neon.cc27 const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign); in cft1st_128_neon()
31 float32x4_t a00v = vld1q_f32(&a[j + 0]); in cft1st_128_neon()
32 float32x4_t a04v = vld1q_f32(&a[j + 4]); in cft1st_128_neon()
33 float32x4_t a08v = vld1q_f32(&a[j + 8]); in cft1st_128_neon()
34 float32x4_t a12v = vld1q_f32(&a[j + 12]); in cft1st_128_neon()
35 float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v)); in cft1st_128_neon()
36 float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v)); in cft1st_128_neon()
37 float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v)); in cft1st_128_neon()
38 float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v)); in cft1st_128_neon()
39 const float32x4_t wk1rv = vld1q_f32(&rdft_wk1r[k2]); in cft1st_128_neon()
[all …]
/external/XNNPACK/src/f32-sigmoid/gen/
Dneonfma-rr1-p5-div-x24.c26 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
28 const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
29 const float32x4_t vc5 = vmovq_n_f32(-0x1.0F9F9Cp-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
30 const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
31 const float32x4_t vc3 = vmovq_n_f32(-0x1.555A80p-3f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
32 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
33 const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFF6p-1f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
34 const float32x4_t vone = vmovq_n_f32(1.0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
35 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
[all …]
/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5p2-minmax-neonfma-1x4-acc3.c34 const float32x4_t vmax = vld1q_dup_f32(&params->neon.max); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
35 const float32x4_t vmin = vld1q_dup_f32(&params->neon.min); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
37 const float32x4_t vw0123 = vld1q_f32(weights); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
38 const float32x4_t vw4567 = vld1q_f32(weights + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
39 const float32x4_t vw89AB = vld1q_f32(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
40 const float32x4_t vwCDEF = vld1q_f32(weights + 12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
41 const float32x4_t vwGHIJ = vld1q_f32(weights + 16); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
42 const float32x4_t vwKLMN = vld1q_f32(weights + 20); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
64 float32x4_t vi0x0123 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
65 float32x4_t vi1x0123 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
[all …]
D5x5p2-minmax-neon-1x4-acc3.c34 const float32x4_t vmax = vld1q_dup_f32(&params->neon.max); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
35 const float32x4_t vmin = vld1q_dup_f32(&params->neon.min); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
37 const float32x4_t vw0123 = vld1q_f32(weights); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
38 const float32x4_t vw4567 = vld1q_f32(weights + 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
39 const float32x4_t vw89AB = vld1q_f32(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
40 const float32x4_t vwCDEF = vld1q_f32(weights + 12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
41 const float32x4_t vwGHIJ = vld1q_f32(weights + 16); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
42 const float32x4_t vwKLMN = vld1q_f32(weights + 20); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
64 float32x4_t vi0x0123 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
65 float32x4_t vi1x0123 = vmovq_n_f32(0.0f); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
[all …]

12345678910>>...32