| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 1x16s4-minmax-fma3-broadcast.c | 76 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
116 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4-minmax-fma3-broadcast.c | 114 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
180 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4-minmax-fma3-broadcast.c | 133 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
212 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 1x16s4inc-minmax-fma3-broadcast.c | 78 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
118 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4inc-minmax-fma3-broadcast.c | 116 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
182 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4inc-minmax-fma3-broadcast.c | 135 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
214 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 1x16s4-minmax-fma3-broadcast.c | 89 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
129 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
| D | 3x16s4-minmax-fma3-broadcast.c | 133 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
199 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
| D | 4x16s4-minmax-fma3-broadcast.c | 155 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
234 const __m256 vb01234567c3 = _mm256_load_ps(w + 48); 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 | 65 … const int8x8_t vb01234567c3 = 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
117 … const int8x8_t vb01234567c3 = 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 | 66 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
121 const int8x8_t vb01234567c3 = 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 | 66 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
122 const int8x8_t vb01234567c3 = 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 | 77 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm() local
133 const int8x8_t vb01234567c3 = 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 | 76 … const int8x8_t vb01234567c3 = 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
128 … const int8x8_t vb01234567c3 = 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 | 77 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
132 const int8x8_t vb01234567c3 = 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 | 67 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane_prfm() local
123 const int8x8_t vb01234567c3 = 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 | 67 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
122 const int8x8_t vb01234567c3 = 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 | 66 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
121 const int8x8_t vb01234567c3 = 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 | 66 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
122 const int8x8_t vb01234567c3 = 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 | 67 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
122 const uint8x8_t vb01234567c3 = 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 | 67 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
122 const uint8x8_t vb01234567c3 = 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 | 78 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane() local
133 const uint8x8_t vb01234567c3 = 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 | 78 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane() local
133 const uint8x8_t vb01234567c3 = 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 | 78 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane() local
133 const int8x8_t vb01234567c3 = 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 | 77 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
133 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm() local
|