/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x16.c | 99 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 136 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
|
D | velu-sse2-rr2-lut16-p3-x16.c | 107 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 152 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local
|
D | velu-sse41-rr2-lut16-p3-x20.c | 104 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 148 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local
|
D | velu-sse41-rr2-lut16-p3-x24.c | 109 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 160 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local
|
D | velu-sse2-rr2-lut16-p3-x20.c | 112 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 166 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local
|
D | velu-sse2-rr2-lut16-p3-x24.c | 117 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local 180 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local
|
D | velu-neon-rr2-lut16-p3-x16.c | 97 const int32x4_t vlCDEF = vcombine_s32(vlCD, vlEF); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x16() local
|
D | velu-neonfma-rr1-lut16-p3-x16.c | 96 const int32x4_t vlCDEF = vcombine_s32(vlCD, vlEF); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16() local
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse41-lut64-p2-div-x16.c | 93 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local 130 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
|
D | sse2-lut64-p2-div-x16.c | 101 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local 146 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local
|
D | sse41-lut64-p2-div-x20.c | 98 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local 142 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local
|
D | sse41-lut64-p2-div-x24.c | 103 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() local 154 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() local
|
D | sse2-lut64-p2-div-x20.c | 106 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20() local 160 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20() local
|
D | sse2-lut64-p2-div-x24.c | 111 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24() local 174 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24() local
|
D | neonfma-rr1-lut64-p2-div-x16.c | 91 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neon-lut64-p2-x16.c | 118 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16() local
|
D | neonfma-lut64-p2-x16.c | 117 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16() local
|
D | neonfma-lut64-p2-x16-acc4.c | 120 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4() local
|
D | neon-lut64-p2-x16-acc4.c | 121 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4() local
|
D | neon-lut64-p2-x16-acc2.c | 119 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2() local
|
D | neonfma-lut64-p2-x16-acc2.c | 118 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2() local
|
D | neonfma-lut64-p2-x20.c | 126 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20() local
|
D | neonfma-lut64-p2-x20-acc2.c | 127 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2() local
|
D | neon-lut64-p2-x20.c | 127 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20() local
|
D | neon-lut64-p2-x20-acc2.c | 128 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2() local
|