/external/XNNPACK/src/f16-vmulcaddc/gen/ |
D | c16-minmax-neonfp16arith-2x.c | 53 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 84 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 104 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 = (const __fp16*) ((uintptr_t) i0 + c); in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local
|
D | c8-minmax-neonfp16arith-2x.c | 52 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local 72 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 = (const __fp16*) ((uintptr_t) i0 + c); in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c4-minmax-neondot.c | 99 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot() local 103 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot() local
|
D | 1x16c4-minmax-neondot.c | 119 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local 124 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x8-minmax-neon-mull-addw-dup.c | 164 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup() local 168 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup() local
|
D | 1x8c2-minmax-neon-mull-padal-dup.c | 127 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mull_padal_dup() local 131 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mull_padal_dup() local
|
D | 1x8-minmax-neon-mlal-lane.c | 168 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane() local 172 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane() local
|
D | 4x8c4-minmax-neondot.c | 165 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local 173 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local
|
D | 1x8c8-minmax-neon-mull-padal.c | 125 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 128 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-sse41-c8-acc2.c | 60 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 106 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 171 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 239 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
|
D | 7p7x-minmax-neon-c8-acc2.c | 52 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 89 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 154 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 234 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
|
D | 7p7x-minmax-wasmsimd-c8-acc2.c | 59 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 104 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 169 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 239 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
|
D | 7p7x-minmax-sse2-c8-acc2.c | 67 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 121 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 194 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 278 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 67 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 121 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 194 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 275 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
|
D | 7p7x-minmax-sse41-c24-acc2.c | 74 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 129 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 192 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 247 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 329 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 451 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
|
D | 7p7x-minmax-sse41-c16-acc2.c | 67 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 130 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 212 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 306 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 73 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 127 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 189 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 243 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 325 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 455 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
|
D | 7p7x-minmax-neon-c24-acc2.c | 66 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 112 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 166 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 219 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 301 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 462 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c4-minmax-neondot.c | 109 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot() local 113 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot() local
|
D | 1x16c4-minmax-neondot.c | 129 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local 134 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x8c2-minmax-neon-mull-padal-dup.c | 141 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mull_padal_dup() local 145 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mull_padal_dup() local
|
D | 1x8-minmax-neon-mlal-lane.c | 181 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane() local 185 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane() local
|
D | 1x8-minmax-neon-mull-addw-dup.c | 178 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mull_addw_dup() local 182 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mull_addw_dup() local
|
D | 4x8c4-minmax-neondot.c | 181 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local 189 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local
|
/external/XNNPACK/src/f16-prelu/gen/ |
D | neonfp16arith-2x16.c | 83 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local 102 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local
|