Home
last modified time | relevance | path

Searched refs:_mm_mullo_epi32 (Results 1 – 25 of 116) sorted by relevance

12345

/external/XNNPACK/src/qc8-dwconv/gen/
Dup24x25-minmax-fp32-sse41-mul32.c187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
191 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
192 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
210 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
211 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
[all …]
Dup24x25-minmax-fp32-avx-mul32.c187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
191 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
192 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
210 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
211 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
[all …]
Dup16x25-minmax-fp32-sse41-mul32.c181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
196 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
197 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
198 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
199 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
[all …]
Dup16x25-minmax-fp32-avx-mul32.c181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
196 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
197 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
198 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
199 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
[all …]
Dup8x25-minmax-fp32-sse41-mul32.c175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
[all …]
Dup8x25-minmax-fp32-avx-mul32.c175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
[all …]
Dup24x9-minmax-fp32-sse41-mul32.c107 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
111 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
112 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
[all …]
Dup24x9-minmax-fp32-avx-mul32.c107 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
111 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
112 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
[all …]
Dup16x9-minmax-fp32-sse41-mul32.c101 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
103 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
104 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
116 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
117 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
118 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
119 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
[all …]
Dup16x9-minmax-fp32-avx-mul32.c101 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
103 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
104 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
116 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
117 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
118 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
119 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
[all …]
/external/XNNPACK/src/qs8-dwconv/gen/
Dup24x25-minmax-fp32-sse41-mul32.c187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
191 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
192 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
210 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
211 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32()
[all …]
Dup24x25-minmax-fp32-avx-mul32.c187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
191 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
192 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
210 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
211 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul32()
[all …]
Dup16x25-minmax-fp32-sse41-mul32.c181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
196 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
197 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
198 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
199 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
[all …]
Dup16x25-minmax-fp32-avx-mul32.c181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
196 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
197 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
198 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
199 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
[all …]
Dup8x25-minmax-fp32-sse41-mul32.c175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
[all …]
Dup8x25-minmax-fp32-avx-mul32.c175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
[all …]
Dup24x9-minmax-fp32-avx-mul32.c107 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
111 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
112 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32()
[all …]
Dup24x9-minmax-fp32-sse41-mul32.c107 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
111 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
112 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse41_mul32()
[all …]
Dup16x9-minmax-fp32-avx-mul32.c101 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
103 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
104 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
116 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
117 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
118 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
119 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
[all …]
Dup16x9-minmax-fp32-sse41-mul32.c101 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
103 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
104 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
116 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
117 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
118 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
119 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32()
[all …]
/external/XNNPACK/src/qu8-dwconv/gen/
Dup16x25-minmax-fp32-sse41-mul32.c182 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
183 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
184 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
185 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
198 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
199 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
200 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul32()
[all …]
Dup16x25-minmax-fp32-avx-mul32.c182 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
183 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
184 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
185 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
198 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
199 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
200 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32()
[all …]
Dup8x25-minmax-fp32-avx-mul32.c176 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
177 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
194 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
195 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32()
[all …]
Dup8x25-minmax-fp32-sse41-mul32.c176 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
177 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
194 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
195 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32()
[all …]
Dup16x9-minmax-fp32-avx-mul32.c102 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
103 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
104 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
105 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
117 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
118 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
119 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
120 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
132 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
133 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32()
[all …]

12345