/external/XNNPACK/src/f32-gemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 69 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local 109 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 101 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local 167 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 117 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local 196 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 133 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local 225 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 1x16s4inc-minmax-fma3-broadcast.c | 71 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local 111 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4inc-minmax-fma3-broadcast.c | 103 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local 169 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4inc-minmax-fma3-broadcast.c | 119 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local 198 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4inc-minmax-fma3-broadcast.c | 135 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local 227 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 82 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local 122 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 120 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local 186 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 139 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local 218 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 158 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_5x16s4__fma3_broadcast() local 250 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40); in xnn_f32_igemm_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16-minmax-rndnu-neon-mull-addw-dup.c | 77 … const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup() local 161 … const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup() local
|
D | 1x16-minmax-rndnu-neon-mlal-lane-prfm.c | 78 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm() local 168 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm() local
|
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 78 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 166 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-fp32-neonv8-mlal-lane.c | 79 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 167 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local
|
D | 1x16-minmax-fp32-neon-mlal-lane.c | 78 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 166 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16-minmax-fp32-neon-mlal-lane.c | 89 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 177 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 89 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 177 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-fp32-neonv8-mlal-lane.c | 90 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 178 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 1x16-minmax-fp32-neon-mlal-lane.c | 79 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 167 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 79 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 167 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 1x16-minmax-fp32-neonv8-mlal-lane.c | 91 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 179 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 1x16-minmax-fp32-neonv8-mlal-lane.c | 90 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 178 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 1x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 79 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm() local 169 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm() local
|