Home
last modified time | relevance | path

Searched refs:va0c3x1 (Results 1 – 25 of 128) sorted by relevance

123456

/external/XNNPACK/src/qs8-gemm/gen/
D1x16c2-minmax-rndnu-neon-mlal-ld2r.c130 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
134 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
138 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
142 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
146 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
D1x16c2-minmax-rndnu-neon-mlal-dup.c128 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
132 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
136 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
140 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
144 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
D1x16c2-minmax-rndnu-neon-mlal-ld1r.c134 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
138 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
142 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
146 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
150 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
D1x16c2-minmax-rndnu-neon-mlal-ld4r.c128 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
132 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
136 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
140 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
144 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
D1x8c2-minmax-fp32-neonv8-mlal-ld4r.c95 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
99 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
103 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
D1x8c2-minmax-fp32-neonv8-mlal-ld1r.c101 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
105 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
109 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
D1x8c2-minmax-fp32-neon-mlal-ld4r.c94 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
98 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
102 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
D1x8c2-minmax-rndnu-neon-mlal-dup.c94 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup() local
98 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
102 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
D1x8c2-minmax-rndnu-neon-mlal-ld2r.c96 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
100 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
104 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
D1x8c2-minmax-fp32-neon-mlal-dup.c94 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
98 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
102 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
D1x8c2-minmax-fp32-neonv8-mlal-dup.c95 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
99 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
103 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
/external/XNNPACK/src/qs8-igemm/gen/
D1x16c2-minmax-rndnu-neon-mlal-ld2r.c142 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
146 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
150 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
154 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
158 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r()
D1x16c2-minmax-rndnu-neon-mlal-ld1r.c146 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
150 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
154 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
158 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
162 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r()
D1x16c2-minmax-rndnu-neon-mlal-dup.c140 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
144 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
148 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
152 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
156 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup()
D1x16c2-minmax-rndnu-neon-mlal-ld4r.c140 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
144 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
148 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
152 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
156 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r()
D1x8c2-minmax-fp32-neon-mlal-ld2r.c108 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
112 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
116 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
D1x8c2-minmax-fp32-neonv8-mlal-ld1r.c113 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
117 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
121 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
D1x8c2-minmax-fp32-neonv8-mlal-dup.c107 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
111 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
115 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
D1x8c2-minmax-fp32-neonv8-mlal-ld4r.c107 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
111 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
115 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
D1x8c2-minmax-fp32-neon-mlal-dup.c106 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
110 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
114 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
D1x8c2-minmax-fp32-neonv8-mlal-ld2r.c109 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
113 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
117 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c2-minmax-fp32-neonv8-mlal-ld4r.c107 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
111 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
115 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
D1x8c2-minmax-fp32-neonv8-mlal-ld1r.c113 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
117 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
121 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c2-minmax-fp32-neon-mlal-ld4r.c94 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
98 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
102 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
D1x8c2-minmax-fp32-neonv8-mlal-ld1r.c101 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
105 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
109 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()

123456