| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 3x4c2s4-xw-minmax-fp32-xop.c | 116 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__xop() local
121 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__xop()
126 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__xop()
130 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__xop()
|
| D | 3x4c2s4-minmax-fp32-avx-ld64.c | 115 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64() local
120 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
125 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
129 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
|
| D | 3x4c2s4-minmax-fp32-xop-ld64.c | 120 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64() local
125 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
130 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
134 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
|
| D | 3x4c2s4-minmax-fp32-sse41-ld64.c | 115 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64() local
120 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
125 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
129 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
|
| D | 3x4c2s4-xw-minmax-fp32-avx.c | 111 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__avx() local
116 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__avx()
121 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__avx()
125 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__avx()
|
| D | 3x4c2s4-minmax-fp32-xop-ld128.c | 118 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128() local
123 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
128 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
132 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
|
| D | 3x4c2s4-xw-minmax-fp32-sse41.c | 111 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse41() local
116 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse41()
121 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse41()
125 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse41()
|
| D | 3x4c2s4-minmax-fp32-sse41-ld128.c | 113 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128() local
118 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
123 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
127 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
|
| D | 3x4c2s4-minmax-fp32-avx-ld128.c | 113 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128() local
118 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
123 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
127 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
|
| D | 3x4c2s4-xw-minmax-fp32-sse2.c | 111 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse2() local
116 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse2()
121 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse2()
125 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2s4__sse2()
|
| D | 3x4c8-xw-minmax-fp32-sse41.c | 121 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41() local
126 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41()
131 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41()
135 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 3x4c2s4-minmax-fp32-xop-ld64.c | 136 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64() local
141 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
146 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
150 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
|
| D | 3x4c2s4-minmax-fp32-xop-ld128.c | 134 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128() local
139 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
144 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
148 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
|
| D | 3x4c2s4-minmax-fp32-sse2-ld128.c | 131 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld128() local
136 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld128()
141 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld128()
145 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld128()
|
| /external/XNNPACK/src/qc8-igemm/gen/ |
| D | 3x4c2s4-minmax-fp32-xop-ld128.c | 134 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128() local
140 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
145 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
149 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
|
| D | 3x4c2s4-minmax-fp32-xop-ld64.c | 136 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64() local
142 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
147 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
151 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
|
| /external/XNNPACK/src/qc8-gemm/gen/ |
| D | 3x4c2s4-minmax-fp32-sse41-ld64.c | 115 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64() local
121 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
126 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
130 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld64()
|
| D | 3x4c2s4-minmax-fp32-avx-ld128.c | 113 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128() local
119 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
124 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
128 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128()
|
| D | 3x4c2s4-minmax-fp32-xop-ld128.c | 118 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128() local
124 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
129 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
133 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
|
| D | 3x4c2s4-minmax-fp32-xop-ld64.c | 120 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64() local
126 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
131 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
135 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
|
| D | 3x4c2s4-minmax-fp32-sse41-ld128.c | 113 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128() local
119 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
124 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
128 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__sse41_ld128()
|
| D | 3x4c2s4-minmax-fp32-avx-ld64.c | 115 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64() local
121 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
126 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
130 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__avx_ld64()
|
| D | 3x4c8-minmax-fp32-sse41-ld128.c | 121 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128() local
127 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128()
132 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128()
136 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128()
|
| /external/XNNPACK/src/qu8-gemm/gen/ |
| D | 3x4c2s4-minmax-fp32-xop-ld64.c | 121 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64() local
126 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
131 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
135 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld64()
|
| D | 3x4c2s4-minmax-fp32-xop-ld128.c | 120 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128() local
125 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
130 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
134 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); in xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128()
|