/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 129 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local 142 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 164 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 177 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 164 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 177 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 165 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 178 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c16-minmax-rndnu-neon-mlal.c | 161 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 190 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 146 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 175 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c8-minmax-fp32-avx512skx.c | 54 …__m512i vacc0xCDEF = _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 | 115 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local 128 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 151 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 164 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 151 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 164 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 152 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 165 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c8-minmax-fp32-avx512skx.c | 53 …__m512i vacc0xCDEF = _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 | 132 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 161 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c16-minmax-rndnu-neon-mlal.c | 147 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 176 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c4-minmax-rndnu-neondot.c | 50 int32x4_t vacc0xCDEF = 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 | 204 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 217 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 163 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local 180 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
/external/XNNPACK/src/f32-prelu/gen/ |
D | wasmsimd-minmax-1x16.c | 61 v128_t vacc0xCDEF = wasm_i32x4_max(vi0xCDEF, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_1x16() local
|
D | wasmsimd-bitselect-1x16.c | 60 v128_t vacc0xCDEF = wasm_f32x4_mul(vi0xCDEF, vwCDEF); in xnn_f32_prelu_ukernel__wasmsimd_bitselect_1x16() local
|
D | neon-1x16.c | 58 float32x4_t vacc0xCDEF = vmulq_f32(vi0xCDEF, vwCDEF); in xnn_f32_prelu_ukernel__neon_1x16() local
|
D | wasmsimd-minmax-2x16.c | 72 v128_t vacc0xCDEF = wasm_i32x4_max(vi0xCDEF, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_2x16() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 1x16c8-minmax-fp32-avx512skx.c | 52 …__m512i vacc0xCDEF = _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 | 53 …__m512i vacc0xCDEF = _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 | 53 …__m512i vacc0xCDEF = _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 | 54 …__m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w … in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx() local
|