/external/XNNPACK/src/f32-prelu/gen/ |
D | wasmsimd-minmax-1x16.c | 58 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_1x16() local
|
D | neon-1x16.c | 55 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB); in xnn_f32_prelu_ukernel__neon_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 | wasmsimd-minmax-2x16.c | 69 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_2x16() local
|
D | wasmsimd-bitselect-2x16.c | 69 v128_t vacc0x89AB = wasm_f32x4_mul(vi0x89AB, vw89AB); in xnn_f32_prelu_ukernel__wasmsimd_bitselect_2x16() local
|
D | neon-2x16.c | 65 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB); in xnn_f32_prelu_ukernel__neon_2x16() local
|
D | wasmsimd-minmax-4x16.c | 91 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_4x16() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 130 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 153 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 146 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 169 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c4-minmax-neondot.c | 48 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x16c8-minmax-avx512skx.c | 57 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 219 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 242 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 2x16c8-minmax-avx512skx.c | 63 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 193 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 220 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 225 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 252 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-avx512skx.c | 69 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 144 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 167 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 160 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 183 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c4-minmax-neondot.c | 49 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local
|
D | 1x16c8-minmax-avx512skx.c | 58 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 233 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 256 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 2x16c8-minmax-avx512skx.c | 62 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 241 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 268 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 209 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 236 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-avx512skx.c | 66 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx() local
|