| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx-rr2-lut16-p3-x8.c | 62 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local
66 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
71 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local
75 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
83 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
128 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local
132 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
137 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local
141 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
149 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
|
| D | velu-sse41-rr2-lut16-p3-x4.c | 58 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
60 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
63 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
65 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
101 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
103 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
106 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
108 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
|
| D | velu-sse2-rr2-lut16-p3-x4.c | 58 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
61 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
65 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
68 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
106 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
109 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
113 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
116 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
|
| D | velu-avx-rr2-lut16-p3-x16.c | 177 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local
181 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
186 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local
190 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
198 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
243 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local
247 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
252 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local
256 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
264 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
|
| D | velu-avx-rr2-lut16-p3-x24.c | 226 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
230 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
235 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
239 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
247 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
292 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
296 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
301 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
305 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
313 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
|
| D | velu-sse41-rr2-lut16-p3-x8.c | 148 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
150 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
153 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
155 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
191 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
193 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
196 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
198 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
|
| D | velu-avx-rr2-lut16-p3-x32.c | 275 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
279 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
284 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
288 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
296 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
341 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
345 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
350 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
354 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
362 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
|
| D | velu-sse2-rr2-lut16-p3-x8.c | 158 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
161 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
165 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
168 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
206 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
209 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
213 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
216 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
|
| D | velu-sse41-rr2-lut16-p3-x12.c | 182 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
184 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
187 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
189 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
225 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
227 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
230 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
232 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
|
| D | velu-avx-rr2-lut16-p3-x40.c | 324 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
328 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
333 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
337 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
345 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
390 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
394 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
399 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
403 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
411 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
|
| D | velu-sse2-rr2-lut16-p3-x12.c | 197 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
200 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
204 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
207 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
245 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
248 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
252 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
255 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
|
| D | velu-sse41-rr2-lut16-p3-x16.c | 216 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
218 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
221 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
223 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
259 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
261 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
264 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
266 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
|
| D | velu-avx-rr2-lut16-p3-x48.c | 373 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
377 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
382 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
386 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
394 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
439 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
443 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
448 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
452 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
460 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
|
| /external/XNNPACK/src/math/ |
| D | expm1minus-avx-rr2-lut16-p3.c | 80 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local
84 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
89 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local
93 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
98 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
|
| D | sigmoid-avx-rr2-lut64-p2-div.c | 84 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uin… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() local
88 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
93 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uin… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() local
97 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
102 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh); in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | sse41-lut64-p2-div-x4.c | 51 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
53 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
56 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
58 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
93 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
95 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
98 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
100 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
|
| D | sse2-lut64-p2-div-x4.c | 51 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
54 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
58 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
61 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
98 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
101 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
105 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
108 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
|
| D | sse41-lut64-p2-div-x8.c | 138 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
140 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
143 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
145 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
180 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
182 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
185 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
187 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
|
| D | sse2-lut64-p2-div-x8.c | 149 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
152 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
156 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
159 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
196 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
199 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
203 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
206 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
|
| D | sse41-lut64-p2-div-x12.c | 170 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
172 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
175 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
177 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
212 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
214 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
217 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
219 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
|
| D | sse41-lut64-p2-div-x16.c | 202 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
204 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
207 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
209 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
244 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
246 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
249 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
251 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
|
| D | sse2-lut64-p2-div-x12.c | 186 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
189 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
193 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
196 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
233 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
236 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
240 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
243 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
|
| /external/XNNPACK/src/f32-velu/ |
| D | avx-rr2-lut16-p3.c.in | 144 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
148 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
153 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
157 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
165 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
210 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
214 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
219 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
223 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
231 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
|
| D | sse-rr2-lut16-p3.c.in | 162 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… variable
165 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
168 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
175 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … variable
178 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
181 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
229 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_1… variable
232 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
235 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
242 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … variable
[all …]
|
| /external/XNNPACK/src/f32-sigmoid/ |
| D | sse-lut64-p2-div.c.in | 157 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… variable
160 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
163 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
170 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
173 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
176 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
222 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_6… variable
225 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
228 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
235 …const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
[all …]
|