/external/XNNPACK/src/f16-gemm/gen-inc/ |
D | 8x16inc-minmax-neonfp16arith-ld64.c | 102 …float16x8_t vacc6x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() local 135 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc0, va6, 0); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 161 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c0, vb89ABCDEFc0); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 182 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc1, va6, 1); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 208 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c1, vb89ABCDEFc1); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 229 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc2, va6, 2); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 255 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c2, vb89ABCDEFc2); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 276 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc3, va6, 3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 302 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c3, vb89ABCDEFc3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() 336 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6, vb89ABCDEF); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/f16-gemm/gen/ |
D | 8x16-minmax-neonfp16arith-ld64.c | 100 float16x8_t vacc6x89ABCDEF = vacc0x89ABCDEF; in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() local 133 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc0, va6, 0); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 159 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c0, vb89ABCDEFc0); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 180 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc1, va6, 1); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 206 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c1, vb89ABCDEFc1); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 227 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc2, va6, 2); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 253 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c2, vb89ABCDEFc2); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 274 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc3, va6, 3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 300 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c3, vb89ABCDEFc3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() 334 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6, vb89ABCDEF); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/f16-igemm/gen/ |
D | 8x16-minmax-neonfp16arith-ld64.c | 88 float16x8_t vacc6x89ABCDEF = vacc0x89ABCDEF; in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() local 165 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc0, va6, 0); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 191 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c0, vb89ABCDEFc0); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 212 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc1, va6, 1); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 238 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c1, vb89ABCDEFc1); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 259 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc2, va6, 2); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 285 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c2, vb89ABCDEFc2); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 306 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc3, va6, 3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 332 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c3, vb89ABCDEFc3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() 364 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6, vb89ABCDEF); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() [all …]
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 408 …const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x89AB), vacc6xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 418 int8x16_t vout6x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc6x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 434 …const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x89AB), vqmovn_s32(vacc6x… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 444 …x16_t vout6x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc6x89ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 384 …const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x89AB), vacc6xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 394 int8x16_t vout6x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc6x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 410 …const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x89AB), vqmovn_s32(vacc6x… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 420 …x16_t vout6x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc6x89ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
|