Home
last modified time | relevance | path

Searched refs:va2c0x0 (Results 1 – 25 of 56) sorted by relevance

123

/external/XNNPACK/src/qs8-igemm/gen/
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c133 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
138 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
148 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
158 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
168 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
178 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
188 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
198 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
208 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c127 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
132 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
142 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
152 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
162 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
172 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
182 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
192 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
202 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c127 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
132 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
142 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
152 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
162 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
172 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
182 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
192 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
202 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x8c4-minmax-rndnu-neon-mlal-dup.c107 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
112 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
122 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
132 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c113 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
118 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
128 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
138 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
148 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c107 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
112 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
122 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
132 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c153 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
160 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
173 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
186 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
199 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
212 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
225 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
238 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
251 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c145 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
152 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
165 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
178 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
191 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
204 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
217 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
230 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
243 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c145 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
152 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
165 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
178 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
191 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
204 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
217 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
230 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
243 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c129 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
136 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
149 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
162 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
175 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c121 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
128 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
141 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
154 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c121 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
128 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
141 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
154 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
/external/XNNPACK/src/qs8-gemm/gen/
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c112 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
117 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
127 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
137 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
147 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
157 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
167 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
177 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
187 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c112 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
117 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
127 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
137 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
147 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
157 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
167 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
177 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
187 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c118 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
123 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
133 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
143 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
153 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
163 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
173 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
183 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
193 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c92 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
97 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
107 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
117 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c98 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
103 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
113 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
123 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
133 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-dup.c92 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
97 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
107 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
117 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-dup.c128 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
135 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
148 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
161 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
174 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
187 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
200 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
213 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
226 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c136 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
143 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
156 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
169 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
182 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
195 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
208 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
221 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
234 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c128 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
135 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
148 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
161 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
174 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
187 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
200 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
213 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
226 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c104 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
111 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
124 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
137 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c104 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
111 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
124 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
137 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c112 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
119 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
132 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
145 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
158 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D3x16c2-minmax-rndnu-neon-mlal-dup.c98 const int8x8_t va2c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
103 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
113 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
123 int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
133 int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()

123