| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x20-acc2.c | 66 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2() local
74 const psimd_f32 vsGHIJ = (psimd_f32) ((psimd_u32) vnGHIJ << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
81 vnGHIJ = psimd_sub_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
89 psimd_f32 vtGHIJ = psimd_qfma_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
95 vtGHIJ = psimd_qfma_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
|
| D | psimd-p5-x20.c | 65 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20() local
73 const psimd_f32 vsGHIJ = (psimd_f32) ((psimd_u32) vnGHIJ << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
80 vnGHIJ = psimd_sub_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
88 psimd_f32 vtGHIJ = psimd_qfma_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
94 vtGHIJ = psimd_qfma_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
|
| D | neonfma-p5-x20-acc2.c | 69 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2() local
77 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
84 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
92 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
98 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
|
| D | sse2-p5-x20.c | 65 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vxGHIJ, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20() local
73 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
80 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
88 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vxGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
94 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
|
| D | neonfma-p5-x20-acc5.c | 72 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5() local
80 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
87 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
95 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
101 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
|
| D | neon-p5-x20-acc5.c | 73 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5() local
81 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
88 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
96 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
102 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
|
| D | neonfma-p5-x20.c | 68 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20() local
76 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
83 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
91 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
97 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
|
| D | psimd-p5-x20-acc5.c | 69 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5() local
77 const psimd_f32 vsGHIJ = (psimd_f32) ((psimd_u32) vnGHIJ << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5()
84 vnGHIJ = psimd_sub_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5()
92 psimd_f32 vtGHIJ = psimd_qfma_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5()
98 vtGHIJ = psimd_qfma_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc5()
|
| D | sse2-p5-x20-acc2.c | 66 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vxGHIJ, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2() local
74 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
81 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
89 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vxGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
95 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
|
| D | neon-p5-x20.c | 69 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20() local
77 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
84 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
92 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
98 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
|
| D | neon-p5-x20-acc2.c | 70 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2() local
78 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
93 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
99 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
|
| D | sse2-p5-x20-acc5.c | 69 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vxGHIJ, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5() local
77 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
84 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
92 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vxGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
98 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
|
| D | neonfma-lut64-p2-x20.c | 69 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20() local
85 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
100 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
143 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
151 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
157 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
|
| D | neonfma-lut64-p2-x20-acc5.c | 73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5() local
89 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
104 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
147 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
155 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
161 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
|
| D | neonfma-lut64-p2-x20-acc2.c | 70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2() local
86 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
101 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
144 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
152 float32x4_t vtGHIJ = vfmaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
158 vtGHIJ = vfmaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
|
| D | neon-lut64-p2-x20-acc2.c | 71 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2() local
87 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
102 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
145 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
153 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
159 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
|
| D | neon-lut64-p2-x20.c | 70 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20() local
86 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
101 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
144 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
152 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
158 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
|
| D | neon-lut64-p2-x20-acc5.c | 74 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vxGHIJ, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5() local
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
105 …const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
148 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
156 float32x4_t vtGHIJ = vmlaq_f32(vxGHIJ, vnGHIJ, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
162 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | psimd-p5-div-x20.c | 73 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20() local
81 const psimd_f32 vsGHIJ = (psimd_f32) ((psimd_u32) vnGHIJ << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
88 vnGHIJ = psimd_sub_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
96 psimd_f32 vtGHIJ = psimd_qfma_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
102 vtGHIJ = psimd_qfma_f32(vtGHIJ, vnGHIJ, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x20()
|
| D | psimd-p5-div-x24.c | 75 psimd_f32 vnGHIJ = psimd_qfma_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24() local
84 const psimd_f32 vsGHIJ = (psimd_f32) ((psimd_u32) vnGHIJ << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
92 vnGHIJ = psimd_sub_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
101 psimd_f32 vtGHIJ = psimd_qfma_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
108 vtGHIJ = psimd_qfma_f32(vtGHIJ, vnGHIJ, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x24()
|
| D | sse41-p5-div-x20.c | 73 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20() local
81 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
88 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
96 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
102 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x20()
|
| D | sse2-p5-div-x20.c | 73 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20() local
81 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
88 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
96 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
102 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
|
| D | neonfma-rr1-p5-div-x20.c | 70 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() local
78 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
85 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
92 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
| D | sse41-p5-div-x24.c | 75 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24() local
84 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
92 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
101 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
108 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x24()
|
| D | neon-rr2-p5-nr2recps-x20.c | 72 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
80 … const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
87 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
95 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
101 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|