/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c4-minmax-neondot.c | 132 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local 137 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x16c2-minmax-neon-mull-padal-dup.c | 194 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mull_padal_dup() local 199 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mull_padal_dup() local
|
D | 4x16c4-minmax-neondot.c | 252 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local 266 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
|
D | 1x16-minmax-neon-mlal-lane.c | 260 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane() local 265 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-neon-mull-addw-dup.c | 257 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mull_addw_dup() local 262 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mull_addw_dup() local
|
D | 2x16-minmax-neon-mlal-lane.c | 350 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane() local 358 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 199 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 204 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 215 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 220 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 6x16c4-minmax-neondot.c | 332 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local 352 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
|
D | 2x16c2-minmax-neon-mull-padal-dup.c | 278 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local 286 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
|
D | 3x16-minmax-neon-mlal-lane.c | 440 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane() local 451 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane() local
|
D | 1x16c8-minmax-avx512skx.c | 123 …const __m128i vout0x0123456789ABCDEF = _mm_shuffle_epi8(vout0x084C2A6E195D3B7F, _mm_set_epi8(15, 7… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx() local
|
D | 1x16c2-minmax-neon-mlal-padal-dup.c | 282 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup() local 287 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c4-minmax-neondot.c | 122 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local 127 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x16-minmax-neon-mull-addw-dup.c | 243 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup() local 248 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup() local
|
D | 1x16c2-minmax-neon-mull-padal-dup.c | 180 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mull_padal_dup() local 185 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mull_padal_dup() local
|
D | 1x16-minmax-neon-mlal-lane.c | 247 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane() local 252 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane() local
|
D | 4x16c4-minmax-neondot.c | 236 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local 250 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
|
D | 2x16-minmax-neon-mlal-lane.c | 335 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane() local 343 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 201 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 206 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 185 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 190 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c8-minmax-avx512skx.c | 112 …const __m128i vout0x0123456789ABCDEF = _mm_shuffle_epi8(vout0x084C2A6E195D3B7F, _mm_set_epi8(15, 7… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx() local
|
D | 2x16c2-minmax-neon-mull-padal-dup.c | 262 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local 270 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
|
D | 3x16-minmax-neon-mlal-lane.c | 423 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane() local 434 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane() local
|
D | 2x16-minmax-neon-mull-addw-dup.c | 359 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup() local 367 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup() local
|