Home
last modified time | relevance | path

Searched refs:va1c3x1 (Results 1 – 25 of 80) sorted by relevance

1234

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c210 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
216 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
223 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
230 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
237 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c198 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
204 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
211 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
218 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
225 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c198 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
204 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
211 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
218 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
225 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c202 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
208 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
215 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
222 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
229 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x8c2-minmax-rndnu-neon-mlal-ld4r.c144 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r() local
150 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r()
157 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r()
D2x8c2-minmax-fp32-neon-mlal-ld4r.c144 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r() local
150 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r()
157 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r()
D2x8c2-minmax-fp32-neonv8-mlal-ld2r.c149 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r() local
155 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r()
162 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r()
D2x8c2-minmax-fp32-neonv8-mlal-dup.c145 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup() local
151 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()
158 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()
D2x8c2-minmax-rndnu-neon-mlal-ld2r.c148 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r() local
154 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r()
161 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r()
D2x8c2-minmax-fp32-neonv8-mlal-ld4r.c145 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r() local
151 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
158 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c188 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
194 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
201 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
208 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
215 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c196 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
202 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
209 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
216 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
223 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c184 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
190 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
197 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
204 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
211 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c184 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
190 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
197 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
204 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
211 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x8c2-minmax-fp32-neon-mlal-ld1r.c142 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r() local
148 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r()
155 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r()
D2x8c2-minmax-rndnu-neon-mlal-ld4r.c130 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r() local
136 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r()
143 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r()
D2x8c2-minmax-rndnu-neon-mlal-ld1r.c142 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld1r() local
148 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld1r()
155 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld1r()
D2x8c2-minmax-fp32-neonv8-mlal-ld1r.c143 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r() local
149 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r()
156 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r()
/external/XNNPACK/src/qc8-gemm/gen/
D2x8c2-minmax-fp32-neonv8-mlal-ld4r.c131 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r() local
137 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
144 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
D2x8c2-minmax-fp32-neonv8-mlal-dup.c131 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup() local
137 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()
144 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()
D2x8c2-minmax-fp32-neon-mlal-ld1r.c142 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r() local
148 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r()
155 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r()
D2x8c2-minmax-fp32-neonv8-mlal-ld1r.c143 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r() local
149 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r()
156 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r()
D2x8c2-minmax-fp32-neon-mlal-ld4r.c130 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r() local
136 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r()
143 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c2-minmax-fp32-neonv8-mlal-ld4r.c145 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r() local
151 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
158 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r()
D2x8c2-minmax-fp32-neonv8-mlal-dup.c145 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup() local
151 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()
158 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup()

1234