/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse41-lut64-p2-div-x12.c | 104 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local 108 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
|
D | sse41-lut64-p2-div-x16.c | 116 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local 120 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
|
D | sse2-lut64-p2-div-x12.c | 114 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local 119 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB))); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
|
D | sse41-lut64-p2-div-x20.c | 128 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() local 132 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
|
D | sse2-lut64-p2-div-x16.c | 128 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16() local 133 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB))); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16()
|
D | sse2-lut64-p2-div-x20.c | 142 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20() local 147 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB))); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20()
|
D | sse41-lut64-p2-div-x24.c | 140 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() local 144 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
|
D | sse2-lut64-p2-div-x24.c | 156 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24() local 161 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB))); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-sse41-rr2-lut16-p3-x12.c | 110 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local 114 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)), 1); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
|
D | velu-sse2-rr2-lut16-p3-x12.c | 120 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local 125 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB))); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
|
D | velu-sse41-rr2-lut16-p3-x16.c | 122 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local 126 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)), 1); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
|
D | velu-sse2-rr2-lut16-p3-x16.c | 134 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() local 139 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB))); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16()
|
D | velu-sse41-rr2-lut16-p3-x20.c | 134 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() local 138 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)), 1); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
|
D | velu-sse2-rr2-lut16-p3-x20.c | 148 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() local 153 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB))); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20()
|
D | velu-sse41-rr2-lut16-p3-x24.c | 146 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() local 150 …AB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)), 1); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
|
D | velu-sse2-rr2-lut16-p3-x24.c | 162 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() local 167 …m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB))); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24()
|