| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neon-lut64-p2-x8.c | 61 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8() local
74 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
80 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
102 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
107 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
110 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
|
| D | neonfma-lut64-p2-x8-acc2.c | 61 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2() local
74 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
80 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
102 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
107 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
110 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
|
| D | neon-lut64-p2-x8-acc2.c | 62 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2() local
75 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
81 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
103 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
108 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
111 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
|
| D | neonfma-lut64-p2-x8.c | 60 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8() local
73 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
79 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
101 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
106 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
109 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
|
| D | sse2-p5-x8-acc2.c | 57 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local
62 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
66 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
71 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
74 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
| D | neonfma-p5-x8-acc2.c | 60 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
65 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
74 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
77 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
| D | wasmsimd-p5-x8.c | 56 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8() local
61 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
65 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
70 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
73 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
|
| D | neon-p5-x8.c | 60 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
65 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
74 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
77 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
| D | sse2-p5-x8.c | 56 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local
61 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
65 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
70 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
73 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
| D | neon-p5-x8-acc2.c | 61 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
66 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
70 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
75 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
78 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
| D | neonfma-p5-x8.c | 59 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local
64 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
68 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
73 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
76 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
| D | wasmsimd-p5-x8-acc2.c | 57 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2() local
62 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2()
66 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2()
71 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2()
74 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2()
|
| D | wasmsimd-p5-x12.c | 58 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12() local
64 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12()
69 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12()
75 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12()
79 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12()
|
| D | neonfma-p5-x12.c | 61 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
67 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
72 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
78 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
82 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
| D | neon-p5-x12-acc2.c | 63 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local
69 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
74 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
80 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
84 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
| D | neonfma-p5-x12-acc2.c | 62 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
68 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
73 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
79 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
83 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | wasmsimd-lut64-p2-div-x8.c | 46 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8() local
49 const v128_t ve4567 = wasm_i32x4_shl(vn4567, 17); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
52 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
73 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
76 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
79 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
|
| D | wasmsimd-p5-div-x8.c | 47 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8() local
50 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
53 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
56 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
59 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
|
| D | sse41-p5-div-x8.c | 47 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local
50 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
53 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
56 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
59 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
|
| D | neon-rr2-lut64-p2-nr2recps-x8.c | 45 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8() local
48 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
52 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
74 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
77 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
80 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
|
| D | neon-rr2-lut2048-p1-nr2recps-x8.c | 45 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() local
48 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 12); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
51 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
73 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
76 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
79 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
|
| D | sse2-p5-div-x8.c | 47 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() local
50 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
53 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
56 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
59 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-neon-rr2-p6-x8.c | 53 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e); in xnn_f32_velu_ukernel__neon_rr2_p6_x8() local
57 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
58 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
61 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
64 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
|
| D | velu-wasmsimd-arm-rr2-lut16-p3-x8.c | 54 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8() local
58 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8()
59 const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8()
78 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8()
82 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8()
85 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x8()
|
| D | velu-sse41-rr2-p6-x8.c | 54 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8() local
57 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
60 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
63 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
66 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
|