/external/XNNPACK/src/f32-velu/gen/ |
D | velu-avx-rr2-lut16-p3-x8.c | 61 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local 65 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 69 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 127 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local 131 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 135 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
|
D | velu-avx-rr2-lut16-p3-x16.c | 176 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local 180 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 184 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 242 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local 246 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 250 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
|
D | velu-avx-rr2-lut16-p3-x24.c | 225 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local 229 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 233 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 291 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local 295 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 299 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
|
D | velu-avx-rr2-lut16-p3-x32.c | 274 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local 278 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 282 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 340 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local 344 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 348 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
|
D | velu-avx-rr2-lut16-p3-x40.c | 323 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local 327 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 331 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 389 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local 393 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 397 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
|
D | velu-avx-rr2-lut16-p3-x48.c | 372 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local 376 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 380 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 438 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local 442 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 446 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
|
/external/XNNPACK/src/math/ |
D | expm1minus-avx-rr2-lut16-p3.c | 79 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local 83 …mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh))); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 87 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
|
D | sigmoid-avx-rr2-lut64-p2-div.c | 83 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() local 87 …mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hh))); in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 91 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hh >> 32))), 1); in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
|
/external/XNNPACK/src/f32-velu/ |
D | avx-rr2-lut16-p3.c.in | 143 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); variable 147 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); 151 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1); 209 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1); variable 213 …h = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)); 217 …l_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
|