Home
last modified time | relevance | path

Searched refs:vidx2_lo (Results 1 – 4 of 4) sorted by relevance

/external/XNNPACK/src/f32-velu/gen/
Dvelu-avx-rr2-lut16-p3-x24.c119 const __m128i vidx2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx2)), 2); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
122 const uint64_t vidx2_ll = (uint64_t) _mm_cvtsi128_si64(vidx2_lo); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
123 const uint64_t vidx2_lh = (uint64_t) _mm_extract_epi64(vidx2_lo, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
135 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
136 …st void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 2))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
139 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 1))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
140 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 3))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
Dvelu-avx-rr2-lut16-p3-x32.c122 const __m128i vidx2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx2)), 2); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
125 const uint64_t vidx2_ll = (uint64_t) _mm_cvtsi128_si64(vidx2_lo); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
126 const uint64_t vidx2_lh = (uint64_t) _mm_extract_epi64(vidx2_lo, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
138 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
139 …st void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 2))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
142 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 1))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
143 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 3))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
Dvelu-avx-rr2-lut16-p3-x40.c125 const __m128i vidx2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx2)), 2); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
128 const uint64_t vidx2_ll = (uint64_t) _mm_cvtsi128_si64(vidx2_lo); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
129 const uint64_t vidx2_lh = (uint64_t) _mm_extract_epi64(vidx2_lo, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
141 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
142 …st void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 2))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
145 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 1))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
146 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 3))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
Dvelu-avx-rr2-lut16-p3-x48.c128 const __m128i vidx2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx2)), 2); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
131 const uint64_t vidx2_ll = (uint64_t) _mm_cvtsi128_si64(vidx2_lo); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
132 const uint64_t vidx2_lh = (uint64_t) _mm_extract_epi64(vidx2_lo, 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
144 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
145 …st void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 2))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
148 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 1))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
149 …int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 3))), 1); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()