/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x16.c | 96 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 133 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
|
D | velu-sse2-rr2-lut16-p3-x16.c | 103 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 148 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local
|
D | velu-sse41-rr2-lut16-p3-x20.c | 101 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 145 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local
|
D | velu-sse41-rr2-lut16-p3-x24.c | 106 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 157 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local
|
D | velu-sse2-rr2-lut16-p3-x20.c | 108 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 162 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local
|
D | velu-sse2-rr2-lut16-p3-x24.c | 113 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local 176 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local
|
D | velu-neonfma-rr1-lut16-p3-x16.c | 91 …int32x2_t vlCD = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16() local
|
D | velu-neon-rr2-lut16-p3-x16.c | 92 …int32x2_t vlCD = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x16() local
|
/external/XNNPACK/src/f32-vsigmoid/gen/ |
D | vsigmoid-sse41-rr2-lut64-p2-div-x16.c | 91 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x16() local 128 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x16() local
|
D | vsigmoid-sse41-rr2-lut64-p2-div-x20.c | 96 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20() local 140 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20() local
|
D | vsigmoid-sse2-rr2-lut64-p2-div-x16.c | 98 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x16() local 143 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x16() local
|
D | vsigmoid-sse41-rr2-lut64-p2-div-x24.c | 101 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x24() local 152 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x24() local
|
D | vsigmoid-sse2-rr2-lut64-p2-div-x20.c | 103 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x20() local 157 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x20() local
|
D | vsigmoid-sse2-rr2-lut64-p2-div-x24.c | 108 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24() local 171 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24() local
|
D | vsigmoid-neonfma-rr1-lut64-p2-div-x16.c | 77 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local
|
D | vsigmoid-neonfma-rr1-lut2048-p1-div-x16.c | 76 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxCD]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neonfma-rr1-lut64-p2-x16.c | 79 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x16() local
|
D | neon-rr2-lut64-p2-x16.c | 80 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x16() local
|
D | neon-rr2-lut64-p2-x16-acc2.c | 81 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x16_acc2() local
|
D | neon-rr2-lut64-p2-x16-acc4.c | 83 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x16_acc4() local
|
D | neonfma-rr1-lut64-p2-x16-acc2.c | 80 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x16_acc2() local
|
D | neonfma-rr1-lut64-p2-x16-acc4.c | 82 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x16_acc4() local
|
D | neon-rr2-lut64-p2-x20-acc2.c | 88 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20_acc2() local
|
D | neon-rr2-lut64-p2-x20.c | 87 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20() local
|
D | neonfma-rr1-lut64-p2-x20.c | 86 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxCD]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x20() local
|