/external/XNNPACK/src/f16-gemm/gen-inc/ |
D | 8x16inc-minmax-neonfp16arith-ld64.c | 104 …float16x8_t vacc7x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() local 136 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 162 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c0, vb89ABCDEFc0); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 183 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 209 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c1, vb89ABCDEFc1); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 230 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 256 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c2, vb89ABCDEFc2); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 277 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 303 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c3, vb89ABCDEFc3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 337 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/f16-gemm/gen/ |
D | 8x16-minmax-neonfp16arith-ld64.c | 102 float16x8_t vacc7x89ABCDEF = vacc0x89ABCDEF; in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() local 134 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 160 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c0, vb89ABCDEFc0); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 181 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 207 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c1, vb89ABCDEFc1); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 228 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 254 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c2, vb89ABCDEFc2); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 275 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 301 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c3, vb89ABCDEFc3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 335 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/f16-igemm/gen/ |
D | 8x16-minmax-neonfp16arith-ld64.c | 90 float16x8_t vacc7x89ABCDEF = vacc0x89ABCDEF; in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() local 166 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 192 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c0, vb89ABCDEFc0); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 213 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 239 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c1, vb89ABCDEFc1); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 260 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 286 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c2, vb89ABCDEFc2); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 307 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 333 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c3, vb89ABCDEFc3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 365 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 410 …const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x89AB), vacc7xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 419 int8x16_t vout7x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc7x01234567), vacc7x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 436 …const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x89AB), vqmovn_s32(vacc7x… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 445 …x16_t vout7x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc7x01234567), vqmovn_s16(vacc7x89ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 386 …const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x89AB), vacc7xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 395 int8x16_t vout7x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc7x01234567), vacc7x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 412 …const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x89AB), vqmovn_s32(vacc7x… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 421 …x16_t vout7x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc7x01234567), vqmovn_s16(vacc7x89ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
|