| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 1x16s4-minmax-fma3-broadcast.c | 52 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
92 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4-minmax-fma3-broadcast.c | 72 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
138 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4-minmax-fma3-broadcast.c | 82 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
161 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 1x16s4inc-minmax-fma3-broadcast.c | 54 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
94 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4inc-minmax-fma3-broadcast.c | 74 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
140 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4inc-minmax-fma3-broadcast.c | 84 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
163 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 1x16s4-minmax-fma3-broadcast.c | 65 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
105 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4-minmax-fma3-broadcast.c | 91 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
157 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4-minmax-fma3-broadcast.c | 104 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
183 const __m256 vb01234567c0 = _mm256_load_ps(w + 0); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 1x8-minmax-rndnu-neon-mull-addw-dup.c | 50 … const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup() local
96 … const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup() local
|
| D | 1x8-minmax-rndnu-neon-mlal-lane.c | 51 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
100 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
|
| D | 1x8-minmax-rndnu-neon-mlal-lane-prfm.c | 51 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
101 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 1x8-minmax-rndnu-neon-mlal-lane-prfm.c | 62 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
112 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
|
| D | 1x8-minmax-rndnu-neon-mull-addw-dup.c | 61 … const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup() local
107 … const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup() local
|
| D | 1x8-minmax-rndnu-neon-mlal-lane.c | 62 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
111 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
|
| /external/XNNPACK/src/qc8-gemm/gen/ |
| D | 1x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 52 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane_prfm() local
102 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane_prfm() local
|
| D | 1x8-minmax-fp32-neonv8-mlal-lane.c | 52 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
101 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
|
| D | 1x8-minmax-fp32-neon-mlal-lane.c | 51 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
100 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
|
| D | 1x8-minmax-fp32-neon-mlal-lane-prfm.c | 51 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
101 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
|
| /external/XNNPACK/src/qu8-gemm/gen/ |
| D | 1x8-minmax-rndnu-neon-mlal-lane.c | 52 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
101 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
|
| D | 1x8-minmax-fp32-neon-mlal-lane.c | 52 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
101 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
|
| /external/XNNPACK/src/qu8-igemm/gen/ |
| D | 1x8-minmax-rndnu-neon-mlal-lane.c | 63 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
112 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
|
| D | 1x8-minmax-fp32-neon-mlal-lane.c | 63 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
112 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
|
| /external/XNNPACK/src/qc8-igemm/gen/ |
| D | 1x8-minmax-fp32-neonv8-mlal-lane.c | 63 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
112 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
|
| D | 1x8-minmax-fp32-neon-mlal-lane-prfm.c | 62 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
112 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
|