/external/XNNPACK/src/f32-gemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 53 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local 93 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 73 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local 139 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 83 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local 162 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 93 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local 185 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 1x16s4inc-minmax-fma3-broadcast.c | 55 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local 95 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4inc-minmax-fma3-broadcast.c | 75 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local 141 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4inc-minmax-fma3-broadcast.c | 85 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local 164 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4inc-minmax-fma3-broadcast.c | 95 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local 187 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast() local
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 1x16s4-minmax-fma3-broadcast.c | 66 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local 106 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast() local
|
D | 3x16s4-minmax-fma3-broadcast.c | 92 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local 158 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast() local
|
D | 4x16s4-minmax-fma3-broadcast.c | 105 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local 184 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast() local
|
D | 5x16s4-minmax-fma3-broadcast.c | 118 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); in xnn_f32_igemm_minmax_ukernel_5x16s4__fma3_broadcast() local 210 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8); 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 | 58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm() local 146 const int8x8_t vb89ABCDEFc0 = 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 | 57 … const int8x8_t vb89ABCDEFc0 = 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 139 … const int8x8_t vb89ABCDEFc0 = 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 | 58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 144 const int8x8_t vb89ABCDEFc0 = 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 | 58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 144 const int8x8_t vb89ABCDEFc0 = 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 | 70 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm() local 158 const int8x8_t vb89ABCDEFc0 = 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 | 70 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 156 const uint8x8_t vb89ABCDEFc0 = 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 | 70 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 156 const uint8x8_t vb89ABCDEFc0 = 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 | 59 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane() local 145 const int8x8_t vb89ABCDEFc0 = 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 | 58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm() local 146 const int8x8_t vb89ABCDEFc0 = 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 | 59 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 145 const uint8x8_t vb89ABCDEFc0 = 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 | 59 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane() local 145 const uint8x8_t vb89ABCDEFc0 = 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 | 68 … const int8x8_t vb89ABCDEFc0 = 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 150 … const int8x8_t vb89ABCDEFc0 = 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 | 69 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local 155 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane() local
|