| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | scalar-lut2048-p1-div-x4.c | 31 const float vminus_log2e = -0x1.715476p0f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4() local
51 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
52 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
53 float vn2 = vz2 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
54 float vn3 = vz3 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
143 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
|
| D | scalar-lut64-p2-div-x4.c | 31 const float vminus_log2e = -0x1.715476p0f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4() local
51 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
52 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
53 float vn2 = vz2 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
54 float vn3 = vz3 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
148 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
|
| D | scalar-p5-div-x4.c | 28 const float vminus_log2e = -0x1.715476p+0f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4() local
51 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
52 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
53 float vn2 = vz2 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
54 float vn3 = vz3 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
154 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
|
| D | scalar-lut2048-p1-div-x2.c | 31 const float vminus_log2e = -0x1.715476p0f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2() local
47 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
48 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
102 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
|
| D | scalar-p5-div-x2.c | 28 const float vminus_log2e = -0x1.715476p+0f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2() local
47 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
48 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
109 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
|
| D | scalar-lut64-p2-div-x2.c | 31 const float vminus_log2e = -0x1.715476p0f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2() local
47 float vn0 = vz0 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
48 float vn1 = vz1 * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
105 float vn = vz * vminus_log2e + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
|
| D | neonfma-rr1-p5-div-x24.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() local
52 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
53 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
54 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
55 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
56 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
57 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
169 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
195 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
| D | wasmsimd-p5-div-x24.c | 27 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24() local
54 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
55 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
56 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
57 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
58 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzGHIJ, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
59 v128_t vnKLMN = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzKLMN, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
173 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
201 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
|
| D | neonfma-rr1-p5-div-x16.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() local
48 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
49 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
50 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
51 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
133 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
159 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
|
| D | wasmsimd-p5-div-x16.c | 27 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16() local
50 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
51 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
52 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
53 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
137 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
165 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
|
| D | wasmsimd-p5-div-x20.c | 27 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20() local
52 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
53 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
54 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
55 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
56 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzGHIJ, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
155 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
183 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
|
| D | neonfma-rr1-p5-div-x20.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() local
50 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
51 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
52 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
53 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
54 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
151 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
177 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
| D | neonfma-rr1-p5-div-x12.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() local
46 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
47 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
48 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
115 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
141 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
|
| D | wasmsimd-p5-div-x12.c | 27 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12() local
48 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
49 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
50 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
119 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
147 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
|
| D | neonfma-rr1-p5-nr1recps1fma-x20.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local
50 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
51 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
52 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
53 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
54 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
169 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
199 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
| D | neonfma-rr1-p5-nr2recps-x20.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local
50 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
51 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
52 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
53 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
54 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
169 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
199 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
| D | neonfma-rr1-p5-nr2recps-x24.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local
52 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
53 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
54 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
55 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
56 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
57 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
190 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
220 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
| D | neonfma-rr1-p5-nr1recps1fma-x24.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local
52 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
53 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
54 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
55 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
56 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
57 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
190 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
220 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
| D | neonfma-rr1-p5-div-x8.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() local
44 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
45 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
97 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
123 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
|
| D | wasmsimd-p5-div-x8.c | 27 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8() local
46 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
47 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
101 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
129 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
|
| D | neonfma-rr1-p5-nr2fma-x24.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local
52 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
53 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
54 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
55 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
56 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
57 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
190 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
220 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
| D | neon-rr2-p5-nr2recps-x24.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
53 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
54 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
55 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
56 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
57 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
58 float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
198 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
229 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
|
| D | neon-rr2-p5-nr2recps-x20.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
51 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
52 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
53 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
54 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
55 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
176 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
207 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
| D | neonfma-rr1-p5-nr2fma-x20.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local
50 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
51 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
52 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
53 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
54 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
169 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
199 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
| D | neon-rr2-p5-nr2recps-x16.c | 27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local
49 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
50 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
51 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
52 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
154 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
185 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|