/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 128 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local 139 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 163 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 174 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 163 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 174 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 164 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 175 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c16-minmax-rndnu-neon-mlal.c | 160 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 183 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 145 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 168 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c8-minmax-fp32-avx512skx.c | 53 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qs8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 114 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local 125 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 150 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 161 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 150 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 161 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 151 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 162 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c8-minmax-fp32-avx512skx.c | 52 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qs8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 131 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 154 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c16-minmax-rndnu-neon-mlal.c | 146 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 169 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c4-minmax-rndnu-neondot.c | 49 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot() local
|
D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 203 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 214 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 162 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local 177 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
/external/XNNPACK/src/f32-prelu/gen/ |
D | wasmsimd-minmax-1x16.c | 59 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_1x16() local
|
D | wasmsimd-bitselect-1x16.c | 58 v128_t vacc0x89AB = wasm_f32x4_mul(vi0x89AB, vw89AB); in xnn_f32_prelu_ukernel__wasmsimd_bitselect_1x16() local
|
D | neon-1x16.c | 56 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB); in xnn_f32_prelu_ukernel__neon_1x16() local
|
D | wasmsimd-minmax-2x16.c | 70 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_2x16() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 1x16c8-minmax-fp32-avx512skx.c | 51 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qc8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 1x16c8-minmax-fp32-avx512skx.c | 52 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 1x16c8-minmax-fp32-avx512skx.c | 52 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qc8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 1x16c8-minmax-fp32-avx512skx.c | 53 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|