| /external/clang/test/CodeGen/ |
| D | arm-v8.1a-neon-intrinsics.c | 24 int32x2_t test_vqrdmlah_s32(int32x2_t a, int32x2_t b, int32x2_t c) { in test_vqrdmlah_s32()
66 int32x2_t test_vqrdmlah_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c) { in test_vqrdmlah_lane_s32()
90 int32x4_t test_vqrdmlahq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c) { in test_vqrdmlahq_lane_s32()
112 int32x2_t test_vqrdmlsh_s32(int32x2_t a, int32x2_t b, int32x2_t c) { in test_vqrdmlsh_s32()
154 int32x2_t test_vqrdmlsh_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c) { in test_vqrdmlsh_lane_s32()
178 int32x4_t test_vqrdmlshq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c) { in test_vqrdmlshq_lane_s32()
|
| D | aarch64-neon-2velem.c | 30 int32x2_t test_vmla_lane_s32(int32x2_t a, int32x2_t b, int32x2_t v) { in test_vmla_lane_s32()
39 int32x4_t test_vmlaq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t v) { in test_vmlaq_lane_s32()
66 int32x2_t test_vmla_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t v) { in test_vmla_laneq_s32()
102 int32x2_t test_vmls_lane_s32(int32x2_t a, int32x2_t b, int32x2_t v) { in test_vmls_lane_s32()
111 int32x4_t test_vmlsq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t v) { in test_vmlsq_lane_s32()
138 int32x2_t test_vmls_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t v) { in test_vmls_laneq_s32()
171 int32x2_t test_vmul_lane_s32(int32x2_t a, int32x2_t v) { in test_vmul_lane_s32()
179 int32x4_t test_vmulq_lane_s32(int32x4_t a, int32x2_t v) { in test_vmulq_lane_s32()
235 int32x2_t test_vmul_laneq_s32(int32x2_t a, int32x4_t v) { in test_vmul_laneq_s32()
518 int64x2_t test_vmlal_lane_s32(int64x2_t a, int32x2_t b, int32x2_t v) { in test_vmlal_lane_s32()
[all …]
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x16c8-minmax-rndnu-neon-mull.c | 335 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
336 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
337 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
338 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
339 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
340 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
342 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
343 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
344 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
345 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
[all …]
|
| D | 3x16c8-minmax-rndnu-neon-mull.c | 268 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
269 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
270 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
271 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
272 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
273 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
275 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
276 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
277 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
278 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
[all …]
|
| D | 4x16c16-minmax-rndnu-neon-mlal.c | 398 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
399 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
400 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
401 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
402 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
403 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
405 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
406 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
407 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
408 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
[all …]
|
| D | 4x8c8-minmax-rndnu-neon-mull.c | 207 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
208 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
209 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
210 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
211 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
212 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
214 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
215 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
216 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
217 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
[all …]
|
| D | 2x16c8-minmax-rndnu-neon-mull.c | 201 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
202 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
203 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
204 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
205 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
206 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
208 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
209 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
210 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
211 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
[all …]
|
| D | 3x16c16-minmax-rndnu-neon-mlal.c | 315 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
316 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
317 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
318 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
319 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
320 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
322 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
323 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
324 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
325 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
[all …]
|
| D | 3x8c8-minmax-rndnu-neon-mull.c | 170 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
171 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
172 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
173 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
174 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
175 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
177 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
178 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
179 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
180 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
[all …]
|
| D | 2x16c16-minmax-rndnu-neon-mlal.c | 232 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
233 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
234 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
235 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
236 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
237 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
239 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
240 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
241 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
242 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
[all …]
|
| D | 4x8c16-minmax-rndnu-neon-mlal.c | 238 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
239 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
240 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
241 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
242 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
243 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
245 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
246 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
247 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
248 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
[all …]
|
| D | 4x16c8-minmax-rndnu-neon-mlal.c | 574 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
575 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
576 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
577 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
578 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
579 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
581 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
582 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
583 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
584 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
[all …]
|
| D | 3x16c8-minmax-rndnu-neon-mlal.c | 457 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
458 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
459 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
460 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
461 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
462 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
464 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
465 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
466 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
467 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
[all …]
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x16c8-minmax-rndnu-neon-mull.c | 355 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
356 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
357 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
358 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
359 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
360 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
362 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
363 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
364 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
365 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
[all …]
|
| D | 3x16c8-minmax-rndnu-neon-mull.c | 286 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
287 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
288 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
289 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
290 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
291 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
293 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
294 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
295 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
296 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
[all …]
|
| D | 4x16c16-minmax-rndnu-neon-mlal.c | 418 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
419 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
420 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
421 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
422 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
423 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
425 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
426 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
427 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
428 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
[all …]
|
| D | 2x16c8-minmax-rndnu-neon-mull.c | 217 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
218 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
219 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
220 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
221 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
222 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
224 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
225 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
226 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
227 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
[all …]
|
| D | 4x8c8-minmax-rndnu-neon-mull.c | 227 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
228 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
229 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
230 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
231 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
232 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
234 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
235 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
236 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
237 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
[all …]
|
| D | 3x16c16-minmax-rndnu-neon-mlal.c | 333 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
334 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
335 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
336 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
337 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
338 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
340 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
341 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
342 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
343 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
[all …]
|
| D | 3x8c8-minmax-rndnu-neon-mull.c | 188 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
189 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
190 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
191 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
192 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
193 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
195 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
196 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
197 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
198 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
[all …]
|
| D | 4x8c16-minmax-rndnu-neon-mlal.c | 258 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
259 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
260 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
261 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
262 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
263 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
265 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
266 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
267 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
268 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
[all …]
|
| D | 2x16c16-minmax-rndnu-neon-mlal.c | 248 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
249 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
250 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
251 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
252 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
253 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
255 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
256 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
257 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
258 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
[all …]
|
| D | 4x16c8-minmax-rndnu-neon-mlal.c | 594 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
595 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
596 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
597 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
598 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
599 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
601 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
602 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
603 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
604 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
[all …]
|
| D | 3x16c8-minmax-rndnu-neon-mlal.c | 475 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
476 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
477 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
478 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
479 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
480 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
482 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
483 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
484 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
485 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
[all …]
|
| D | 3x8c16-minmax-rndnu-neon-mlal.c | 211 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
212 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
213 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
214 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
215 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
216 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
218 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
219 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
220 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
221 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
[all …]
|