/external/XNNPACK/src/f32-gemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 61 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local 101 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 87 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local 153 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 100 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local 179 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 113 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local 205 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 1x16s4inc-minmax-fma3-broadcast.c | 63 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local 103 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4inc-minmax-fma3-broadcast.c | 89 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local 155 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4inc-minmax-fma3-broadcast.c | 102 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local 181 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4inc-minmax-fma3-broadcast.c | 115 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local 207 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 74 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local 114 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 106 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local 172 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 122 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local 201 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 138 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_5x16s4__fma3_broadcast() local 230 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24); in xnn_f32_igemm_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16-minmax-rndnu-neon-mlal-lane-prfm.c | 68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm() local 157 const int8x8_t vb89ABCDEFc1 = 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-mull-addw-dup.c | 67 … const int8x8_t vb89ABCDEFc1 = 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 150 … const int8x8_t vb89ABCDEFc1 = 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.c | 68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 155 const int8x8_t vb89ABCDEFc1 = 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-neon-mlal-lane.c | 68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 155 const int8x8_t vb89ABCDEFc1 = 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/qc8-igemm/gen/ |
D | 1x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 80 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm() local 169 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm() local
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 80 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 167 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-fp32-neon-mlal-lane.c | 80 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 167 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 1x16-minmax-fp32-neonv8-mlal-lane.c | 69 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 156 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local
|
D | 1x16-minmax-fp32-neon-mlal-lane-prfm.c | 68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm() local 157 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm() local
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 69 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 156 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|
D | 1x16-minmax-fp32-neon-mlal-lane.c | 69 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 156 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16-minmax-rndnu-neon-mull-addw-dup.c | 78 … const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup() local 161 … const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup() local
|
D | 1x16-minmax-rndnu-neon-mlal-lane.c | 79 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 166 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|