/external/XNNPACK/src/f16-vmulcaddc/gen/ |
D | c16-minmax-neonfp16arith-2x.c | 55 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 85 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local 105 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 = (const __fp16*) ((uintptr_t) i1 + c); in xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x() local
|
D | c8-minmax-neonfp16arith-2x.c | 53 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local 73 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 = (const __fp16*) ((uintptr_t) i1 + c); in xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x() local
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-sse41-c8-acc2.c | 61 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 107 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 172 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 240 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
|
D | 7p7x-minmax-neon-c8-acc2.c | 53 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 90 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 155 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 235 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
|
D | 7p7x-minmax-wasmsimd-c8-acc2.c | 60 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 105 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 170 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 240 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
|
D | 7p7x-minmax-sse2-c8-acc2.c | 68 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 122 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 279 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 68 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 122 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 276 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
|
D | 7p7x-minmax-sse41-c24-acc2.c | 77 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 130 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 248 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 332 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 452 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
|
D | 7p7x-minmax-sse41-c16-acc2.c | 69 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 132 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 214 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 307 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 76 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 128 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 192 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 244 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 328 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 456 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
|
D | 7p7x-minmax-neon-c24-acc2.c | 69 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 113 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 169 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 220 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 304 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 463 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
|
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 68 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 130 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 212 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 310 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
|
D | 7p7x-minmax-neon-c16-acc2.c | 61 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local 115 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local 199 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local 316 int16x8_t vacc1x01234567 = vaddl_s8(vi2x01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local
|
D | 7x-minmax-sse41-c8-acc2.c | 78 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 145 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 98 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 161 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 248 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 311 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 417 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 560 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
|
/external/XNNPACK/src/f16-prelu/gen/ |
D | neonfp16arith-2x16.c | 85 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local 104 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x8c4-minmax-neondot.c | 182 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local 190 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local
|
D | 2x8-minmax-neon-mlal-lane.c | 232 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane() local 237 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane() local
|
D | 6x8c4-minmax-neondot.c | 230 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() local 241 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() local
|
D | 2x8c2-minmax-neon-mull-padal-dup.c | 188 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mull_padal_dup() local 193 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mull_padal_dup() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x8c4-minmax-neondot.c | 166 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local 174 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local
|
D | 2x8-minmax-neon-mlal-lane.c | 217 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane() local 222 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane() local
|
D | 2x8c2-minmax-neon-mull-padal-dup.c | 172 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mull_padal_dup() local 177 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mull_padal_dup() local
|
D | 2x8-minmax-neon-mull-addw-dup.c | 226 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup() local 231 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup() local
|
D | 3x8-minmax-neon-mlal-lane.c | 265 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane() local 272 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane() local
|