/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x8.c | 68 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local 85 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
|
D | velu-sse2-rr2-lut16-p3-x8.c | 70 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local 91 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
|
D | velu-sse41-rr2-lut16-p3-x12.c | 73 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local 97 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
|
D | velu-sse41-rr2-lut16-p3-x16.c | 78 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 109 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
|
D | velu-sse2-rr2-lut16-p3-x12.c | 75 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local 105 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
|
D | velu-sse2-rr2-lut16-p3-x16.c | 80 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 119 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local
|
D | velu-sse41-rr2-lut16-p3-x20.c | 83 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 121 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local
|
D | velu-neon-rr2-lut16-p3-x8.c | 66 const int32x4_t vl0123 = vcombine_s32(vl01, vl23); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x8() local
|
D | velu-neonfma-rr1-lut16-p3-x8.c | 65 const int32x4_t vl0123 = vcombine_s32(vl01, vl23); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x8() local
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse41-lut64-p2-div-x8.c | 62 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local 79 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
|
D | sse2-lut64-p2-div-x8.c | 64 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local 85 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
|
D | sse41-lut64-p2-div-x12.c | 67 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local 91 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
|
D | sse41-lut64-p2-div-x16.c | 72 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local 103 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
|
D | sse2-lut64-p2-div-x12.c | 69 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local 99 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
|
D | sse41-lut64-p2-div-x20.c | 77 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local 115 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local
|
D | sse2-lut64-p2-div-x16.c | 74 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local 113 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local
|
D | neonfma-rr1-lut2048-p1-div-x8.c | 63 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8() local
|
D | neonfma-rr1-lut64-p2-nr2fma-x8.c | 64 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8() local
|
D | wasmsimd-lut64-p2-div-x8.c | 60 const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8() local
|
D | neonfma-rr1-lut2048-p1-nr2recps-x8.c | 63 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() local
|
D | neonfma-rr1-lut2048-p1-nr2fma-x8.c | 63 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neon-lut64-p2-x8.c | 91 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8() local
|
D | neonfma-lut64-p2-x8.c | 90 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8() local
|
D | neonfma-lut64-p2-x8-acc2.c | 91 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2() local
|
D | neon-lut64-p2-x8-acc2.c | 92 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2() local
|