| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx512f-nr1fma1adj-x128.c | 66 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
67 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
68 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
69 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
70 const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
71 const __m512 vresidual5 = _mm512_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
72 const __m512 vresidual6 = _mm512_fnmadd_ps(vsqrtx6, vhalfrsqrtx6, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
73 const __m512 vresidual7 = _mm512_fnmadd_ps(vsqrtx7, vhalfrsqrtx7, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
92 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
93 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
[all …]
|
| D | avx512f-nr1fma1adj-x112.c | 62 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
63 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
64 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
65 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
66 const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
67 const __m512 vresidual5 = _mm512_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
68 const __m512 vresidual6 = _mm512_fnmadd_ps(vsqrtx6, vhalfrsqrtx6, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
85 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
86 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
87 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
[all …]
|
| D | avx512f-nr1fma1adj-x96.c | 58 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
59 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
60 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
61 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
62 const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
63 const __m512 vresidual5 = _mm512_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
78 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
79 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
80 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
81 const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
[all …]
|
| D | avx512f-nr1fma1adj-x80.c | 54 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
55 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
56 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
57 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
58 const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
71 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
72 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
73 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
74 const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
75 const __m512 vadjustment4 = _mm512_fnmadd_ps(vsqrtx4, vsqrtx4, vx4); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
[all …]
|
| D | avx512f-nr1fma1adj-x64.c | 50 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
51 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
52 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
53 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
64 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
65 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
66 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
67 const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
87 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
90 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
[all …]
|
| D | avx512f-nr1fma1adj-x48.c | 46 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
47 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
48 const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
57 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
58 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
59 const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
77 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
80 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
97 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
100 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
|
| D | avx512f-nr1fma1adj-x32.c | 42 const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
43 const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
50 const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
51 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
67 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
70 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
87 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
90 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
|
| D | avx512f-nr1fma1adj-x16.c | 36 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16()
39 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16()
56 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16()
59 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16()
|
| /external/XNNPACK/src/f32-vsqrt/ |
| D | avx512f-nr1fma1adj.c.in | 44 … const __m512 vresidual${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vhalfrsqrtx${ABC[N]}, vhalf);
51 …const __m512 vadjustment${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vsqrtx${ABC[N]}, vx${ABC[N]}…
68 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
71 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
88 const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
91 const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c | 159 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
160 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
161 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
162 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
163 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
164 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
165 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
166 vr7 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr7, vd7, vone), vr7, vr7); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
218 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
253 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x128.c | 156 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
157 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
158 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
159 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
160 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
161 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
162 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
163 vr7 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr7, vd7, vone), vr7, vr7); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
214 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
248 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x96.c | 130 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
131 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
132 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
133 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
134 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
135 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
180 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
214 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c | 146 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
147 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
148 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
149 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
150 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
151 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
152 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
201 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
236 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c | 152 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
153 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
154 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
155 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
156 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
157 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
158 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
207 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
242 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x112.c | 143 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
144 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
145 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
146 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
147 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
148 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
149 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
197 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
231 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c | 165 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
166 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
167 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
168 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
169 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
170 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
171 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
172 vr7 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr7, vd7, vone), vr7, vr7); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
224 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
259 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x64.c | 104 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
105 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
106 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
107 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
146 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
180 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x80.c | 117 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
118 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
119 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
120 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
121 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
163 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
197 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c | 126 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
127 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
128 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
129 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
130 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
173 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
208 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c | 120 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
121 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
122 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
123 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
124 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
167 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
202 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c | 133 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
134 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
135 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
136 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
137 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
138 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
184 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
219 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c | 139 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
140 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
141 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
142 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
143 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
144 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
190 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
225 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c | 107 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
108 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
109 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
110 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
150 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
185 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c | 113 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
114 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
115 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
116 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
156 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
191 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
|
| /external/XNNPACK/src/math/ |
| D | sqrt-avx512f-nr2fma.c | 36 __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_math_f32_sqrt__avx512f_nr2fma()
40 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf); in xnn_math_f32_sqrt__avx512f_nr2fma()
|