| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx-rr2-lut16-p3-x8.c | 64 …__m128i vl_hl = _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
68 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
73 …__m128i vl_hl = _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
77 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
84 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
130 …__m128i vl_hl = _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
134 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
139 …__m128i vl_hl = _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
143 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
150 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
|
| D | velu-sse41-rr2-lut16-p3-x4.c | 59 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
61 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
64 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
66 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
102 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
104 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
107 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() local
109 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 59 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
63 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
66 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
70 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
107 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
111 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
114 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() local
118 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4()
|
| D | velu-avx-rr2-lut16-p3-x16.c | 179 …__m128i vl_hl = _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
183 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
188 …__m128i vl_hl = _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
192 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
199 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
245 …__m128i vl_hl = _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
249 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
254 …__m128i vl_hl = _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
258 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
265 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
|
| D | velu-avx-rr2-lut16-p3-x24.c | 228 …__m128i vl_hl = _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
232 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
237 …__m128i vl_hl = _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
241 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
248 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
294 …__m128i vl_hl = _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
298 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
303 …__m128i vl_hl = _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
307 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
314 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
|
| D | velu-sse41-rr2-lut16-p3-x8.c | 149 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
151 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
154 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
156 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
192 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
194 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8()
197 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() local
199 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 277 …__m128i vl_hl = _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
281 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
286 …__m128i vl_hl = _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
290 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
297 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
343 …__m128i vl_hl = _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
347 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
352 …__m128i vl_hl = _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
356 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
363 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
|
| D | velu-sse2-rr2-lut16-p3-x8.c | 159 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
163 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
166 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
170 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
207 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
211 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
214 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() local
218 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8()
|
| D | velu-sse41-rr2-lut16-p3-x12.c | 183 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
185 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
188 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
190 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
226 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
228 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12()
231 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() local
233 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 326 …__m128i vl_hl = _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
330 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
335 …__m128i vl_hl = _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
339 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
346 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
392 …__m128i vl_hl = _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
396 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
401 …__m128i vl_hl = _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
405 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
412 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
|
| D | velu-sse2-rr2-lut16-p3-x12.c | 198 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
202 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
205 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
209 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
246 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
250 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
253 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() local
257 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12()
|
| D | velu-sse41-rr2-lut16-p3-x16.c | 217 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
219 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
222 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
224 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
260 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
262 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16()
265 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() local
267 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 375 …__m128i vl_hl = _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
379 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
384 …__m128i vl_hl = _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
388 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
395 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
441 …__m128i vl_hl = _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
445 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
450 …__m128i vl_hl = _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
454 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
461 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
|
| /external/XNNPACK/src/math/ |
| D | expm1minus-avx-rr2-lut16-p3.c | 82 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local
86 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
91 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local
95 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
99 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
|
| D | sigmoid-avx-rr2-lut64-p2-div.c | 86 …__m128i vl_hl = _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
90 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
95 …__m128i vl_hl = _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
99 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
103 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh); in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | sse41-lut64-p2-div-x4.c | 52 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
54 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
57 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
59 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
94 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
96 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
99 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() local
101 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 52 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
56 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
59 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
63 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
99 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
103 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
106 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4() local
110 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4()
|
| D | sse41-lut64-p2-div-x8.c | 139 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
141 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
144 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
146 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
181 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
183 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8()
186 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() local
188 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 150 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
154 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
157 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
161 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
197 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
201 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
204 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8() local
208 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8()
|
| D | sse41-lut64-p2-div-x12.c | 171 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
173 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
176 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
178 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
213 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
215 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12()
218 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() local
220 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 203 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
205 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
208 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
210 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
245 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
247 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
250 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() local
252 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((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 | 187 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
191 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
194 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
198 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
234 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
238 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
241 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12() local
245 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12()
|
| /external/XNNPACK/src/f32-velu/ |
| D | avx-rr2-lut16-p3.c.in | 146 …__m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
150 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
155 …__m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
159 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
166 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
212 …__m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
216 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
221 …__m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… variable
225 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3…
232 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
|
| D | sse-rr2-lut16-p3.c.in | 163 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … variable
170 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
173 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
176 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … variable
183 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
186 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
230 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … variable
237 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
240 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
243 …const __m128i vl_hl = _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 | 158 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
165 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
168 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
171 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
178 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
181 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
223 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
230 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov…
233 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
236 …const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 … variable
[all …]
|