/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse41-lut64-p2-div-x12.c | 80 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local 108 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
|
D | sse41-lut64-p2-div-x16.c | 85 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local 120 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
|
D | sse2-lut64-p2-div-x12.c | 86 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local 120 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
|
D | sse41-lut64-p2-div-x20.c | 90 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local 132 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local
|
D | sse2-lut64-p2-div-x16.c | 91 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local 134 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local
|
D | sse41-lut64-p2-div-x24.c | 95 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() local 144 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() local
|
D | sse2-lut64-p2-div-x20.c | 96 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20() local 148 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20() local
|
D | neonfma-rr1-lut64-p2-div-x12.c | 69 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12() local
|
D | neonfma-rr1-lut64-p2-nr2recps-x12.c | 69 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() local
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c | 68 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxAB]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() local
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x12.c | 86 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local 114 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
|
D | velu-sse41-rr2-lut16-p3-x16.c | 91 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 126 …const __m128i vlAB = _mm_insert_epi32(vlA, *((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-x12.c | 92 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local 126 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
|
D | velu-sse2-rr2-lut16-p3-x16.c | 97 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 140 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local
|
D | velu-sse41-rr2-lut16-p3-x20.c | 96 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 138 …const __m128i vlAB = _mm_insert_epi32(vlA, *((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 | 101 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 150 …const __m128i vlAB = _mm_insert_epi32(vlA, *((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 | 102 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 154 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local
|
D | velu-neon-rr2-lut16-p3-x12.c | 82 …int32x2_t vlAB = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12() local
|
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neonfma-lut64-p2-x12-acc2.c | 96 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2() local
|
D | neon-lut64-p2-x12-acc3.c | 98 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3() local
|
D | neon-lut64-p2-x12.c | 96 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12() local
|
D | neonfma-lut64-p2-x12.c | 95 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12() local
|
D | neon-lut64-p2-x12-acc2.c | 97 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2() local
|
D | neonfma-lut64-p2-x12-acc3.c | 97 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3() local
|
D | neonfma-lut64-p2-x16.c | 102 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxAB]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16() local
|