| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-fp32-sse2-mul16.c | 158 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
160 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
161 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
303 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
305 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
306 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-fp32-sse41-mul16-add16.c | 145 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16() local
149 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16()
274 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16() local
277 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16()
|
| D | up8x9-minmax-fp32-avx-mul16-add16.c | 145 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16() local
149 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16()
274 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16() local
277 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16()
|
| D | up8x9-minmax-fp32-xop-mul16-add16.c | 150 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16() local
154 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16()
279 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16() local
282 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16()
|
| D | up16x9-minmax-fp32-sse2-mul16.c | 202 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
206 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
207 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
394 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
396 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
397 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
|
| D | up8x9-minmax-fp32-sse41-mul16.c | 151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
155 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
288 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
291 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-fp32-avx-mul16.c | 151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
155 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
288 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
291 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
|
| D | up8x9-minmax-fp32-sse2-mul16-add16.c | 149 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16() local
151 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16()
282 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16() local
284 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16()
|
| D | up16x9-minmax-fp32-xop-mul16-add16.c | 181 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16() local
189 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16()
351 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16() local
355 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16()
|
| D | up24x9-minmax-fp32-sse2-mul16.c | 246 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16() local
252 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
253 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
480 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16() local
482 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
483 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
|
| /external/XNNPACK/src/qu8-dwconv/gen/ |
| D | up8x9-minmax-fp32-sse41-mul16.c | 157 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
161 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
162 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
302 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
305 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
306 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-fp32-avx-mul16.c | 157 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
161 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
162 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
302 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
305 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
306 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
|
| D | up8x9-minmax-fp32-sse2-mul16.c | 160 …const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
162 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
163 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
305 …const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
307 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
308 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
|
| D | up16x9-minmax-fp32-avx-mul16.c | 199 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16() local
207 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16()
208 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16()
391 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16() local
395 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16()
396 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16()
|
| D | up16x9-minmax-fp32-sse41-mul16.c | 199 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16() local
207 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16()
208 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16()
391 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16() local
395 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16()
396 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16()
|
| D | up16x9-minmax-fp32-sse2-mul16.c | 204 …const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
208 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
209 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
395 …const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
397 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
398 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
|
| /external/XNNPACK/src/qc8-dwconv/gen/ |
| D | up8x9-minmax-fp32-sse2-mul16.c | 158 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
160 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
161 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
305 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16() local
307 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
308 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-fp32-xop-mul16-add16.c | 150 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16() local
154 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16()
281 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16() local
284 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16()
|
| D | up8x9-minmax-fp32-avx-mul16-add16.c | 145 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16() local
149 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16()
276 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16() local
279 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16_add16()
|
| D | up8x9-minmax-fp32-sse41-mul16-add16.c | 145 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16() local
149 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16()
276 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16() local
279 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16()
|
| D | up16x9-minmax-fp32-sse2-mul16.c | 202 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
206 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
207 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
398 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16() local
400 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
401 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16()
|
| D | up8x9-minmax-fp32-sse41-mul16.c | 151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
155 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
290 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16() local
293 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-fp32-sse2-mul16-add16.c | 149 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16() local
151 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16()
284 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16() local
286 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16()
|
| D | up8x9-minmax-fp32-avx-mul16.c | 151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
155 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
290 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16() local
293 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16()
|
| D | up24x9-minmax-fp32-sse2-mul16.c | 246 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16() local
252 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
253 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
486 … const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16() local
488 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
489 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_mul16()
|