/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x20.c | 109 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 155 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((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 | 114 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 167 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((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 | 119 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 175 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local
|
D | velu-sse2-rr2-lut16-p3-x24.c | 124 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local 189 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local
|
D | velu-neonfma-rr1-lut16-p3-x20.c | 104 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20() local
|
D | velu-neon-rr2-lut16-p3-x20.c | 105 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20() local
|
/external/XNNPACK/src/f32-vsigmoid/gen/ |
D | vsigmoid-sse41-rr2-lut64-p2-div-x20.c | 104 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20() local 150 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20() local
|
D | vsigmoid-sse41-rr2-lut64-p2-div-x24.c | 109 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x24() local 162 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((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 | 114 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x20() local 170 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x20() local
|
D | vsigmoid-sse2-rr2-lut64-p2-div-x24.c | 119 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24() local 184 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24() local
|
D | vsigmoid-neonfma-rr1-lut64-p2-div-x20.c | 87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local
|
D | vsigmoid-neonfma-rr1-lut2048-p1-div-x20.c | 86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c | 87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local
|
D | vsigmoid-neon-rr2-lut64-p2-nr2recps-x20.c | 88 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20() local
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c | 86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x20.c | 86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x20.c | 87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() local
|
D | vsigmoid-neon-rr2-lut2048-p1-nr2recps-x20.c | 87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neon-rr2-lut64-p2-x20.c | 90 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20() local
|
D | neon-rr2-lut64-p2-x20-acc2.c | 91 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20_acc2() local
|
D | neonfma-rr1-lut64-p2-x20.c | 89 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x20() local
|
D | neonfma-rr1-lut64-p2-x20-acc2.c | 90 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x20_acc2() local
|
D | neon-rr2-lut64-p2-x20-acc5.c | 94 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20_acc5() local
|
D | neonfma-rr1-lut64-p2-x20-acc5.c | 93 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x20_acc5() local
|