Home
last modified time | relevance | path

Searched refs:vout0123x0123456789ABCDEF (Results 1 – 7 of 7) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D4x16c8-minmax-fp32-avx512skx.c165 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
166 vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
169 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
170 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
171 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
172 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
189 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
191 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
193 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
195 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/qc8-gemm/gen/
D4x16c8-minmax-fp32-avx512skx.c167 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
168 vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
171 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
172 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
173 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
174 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
191 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
193 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
195 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
197 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/qu8-igemm/gen/
D4x16c8-minmax-fp32-avx512skx.c183 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
184 vout0123x0123456789ABCDEF = _mm512_max_epu8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
187 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
188 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
189 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
190 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
204 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
206 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
208 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
210 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/qs8-igemm/gen/
D4x16c8-minmax-fp32-avx512skx.c182 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
183 vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
186 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
187 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
188 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
189 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
203 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
205 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
207 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
209 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/qu8-gemm/gen/
D4x16c8-minmax-fp32-avx512skx.c166 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
167 vout0123x0123456789ABCDEF = _mm512_max_epu8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
170 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
171 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
172 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
173 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
190 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
192 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
194 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
196 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/qc8-igemm/gen/
D4x16c8-minmax-fp32-avx512skx.c184 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
185 vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
188 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
189 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
190 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
191 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
205 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
207 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
209 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
211 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
/external/XNNPACK/src/amalgam/
Davx512skx.c1628 …__m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_e… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local
1629 vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1632 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1633 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1634 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1635 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1652 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1654 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1656 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
1658 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
[all …]