/aosp_15_r20/external/liblc3/src/ |
H A D | mdct_neon.h | 88 vst1_f32( (float *)(y + 0), vget_low_f32(y0) ); in neon_fft_5() 89 vst1_f32( (float *)(y + 1), vget_low_f32(y1) ); in neon_fft_5() 90 vst1_f32( (float *)(y + 2), vget_low_f32(y2) ); in neon_fft_5() 91 vst1_f32( (float *)(y + 3), vget_low_f32(y3) ); in neon_fft_5() 92 vst1_f32( (float *)(y + 4), vget_low_f32(y4) ); in neon_fft_5() 94 vst1_f32( (float *)(y + 5), vget_high_f32(y0) ); in neon_fft_5() 95 vst1_f32( (float *)(y + 6), vget_high_f32(y1) ); in neon_fft_5() 96 vst1_f32( (float *)(y + 7), vget_high_f32(y2) ); in neon_fft_5() 97 vst1_f32( (float *)(y + 8), vget_high_f32(y3) ); in neon_fft_5() 98 vst1_f32( (float *)(y + 9), vget_high_f32(y4) ); in neon_fft_5() [all …]
|
/aosp_15_r20/external/webrtc/common_audio/third_party/ooura/fft_size_128/ |
H A D | ooura_fft_neon.cc | 116 vst1_f32(&a[j + 0], vget_low_f32(xx0)); in cftmdl_128_neon() 117 vst1_f32(&a[j + 32], vget_high_f32(xx0)); in cftmdl_128_neon() 118 vst1_f32(&a[j + 16], vget_low_f32(xx1)); in cftmdl_128_neon() 119 vst1_f32(&a[j + 48], vget_high_f32(xx1_rev)); in cftmdl_128_neon() 123 vst1_f32(&a[j + 8], vget_low_f32(x1_x3_add)); in cftmdl_128_neon() 124 vst1_f32(&a[j + 24], vget_low_f32(x1_x3_sub)); in cftmdl_128_neon() 125 vst1_f32(&a[j + 40], vget_low_f32(yy4)); in cftmdl_128_neon() 126 vst1_f32(&a[j + 56], vget_high_f32(yy4_rev)); in cftmdl_128_neon() 170 vst1_f32(&a[j + 0], vget_low_f32(xx)); in cftmdl_128_neon() 171 vst1_f32(&a[j + 32], vget_high_f32(xx)); in cftmdl_128_neon() [all …]
|
/aosp_15_r20/external/XNNPACK/src/f32-conv-hwc2chw/ |
H A D | 3x3s2p1c3x4-neonfma-2x2.c | 341 vst1_f32(o1c0, vget_low_f32(vo1c01)); o1c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 342 vst1_f32(o1c1, vget_high_f32(vo1c01)); o1c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 343 vst1_f32(o1c2, vget_low_f32(vo1c23)); o1c2 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 344 vst1_f32(o1c3, vget_high_f32(vo1c23)); o1c3 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 346 vst1_f32(o0c0, vget_low_f32(vo0c01)); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 347 vst1_f32(o0c1, vget_high_f32(vo0c01)); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 348 vst1_f32(o0c2, vget_low_f32(vo0c23)); o0c2 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 349 vst1_f32(o0c3, vget_high_f32(vo0c23)); o0c3 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 601 vst1_f32(o1c0, vget_low_f32(vo1c01)); o1c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() 602 vst1_f32(o1c1, vget_high_f32(vo1c01)); o1c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() [all …]
|
H A D | 3x3s2p1c3x4-neon-2x2.c | 339 vst1_f32(o1c0, vget_low_f32(vo1c0123.val[0])); o1c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 340 vst1_f32(o1c1, vget_high_f32(vo1c0123.val[0])); o1c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 341 vst1_f32(o1c2, vget_low_f32(vo1c0123.val[1])); o1c2 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 342 vst1_f32(o1c3, vget_high_f32(vo1c0123.val[1])); o1c3 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 344 vst1_f32(o0c0, vget_low_f32(vo0c0123.val[0])); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 345 vst1_f32(o0c1, vget_high_f32(vo0c0123.val[0])); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 346 vst1_f32(o0c2, vget_low_f32(vo0c0123.val[1])); o0c2 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 347 vst1_f32(o0c3, vget_high_f32(vo0c0123.val[1])); o0c3 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 597 vst1_f32(o1c0, vget_low_f32(vo1c0123.val[0])); o1c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 598 vst1_f32(o1c1, vget_high_f32(vo1c0123.val[0])); o1c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() [all …]
|
/aosp_15_r20/external/XNNPACK/src/f32-ppmm/gen/ |
H A D | 8x8-minmax-neon.c | 205 vst1_f32(c7, vacc7x01); c7 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 206 vst1_f32(c6, vacc6x01); c6 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 207 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 208 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 209 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 210 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 211 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 212 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon()
|
H A D | 8x8-minmax-neonfma.c | 233 vst1_f32(c7, vacc7x01); c7 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 234 vst1_f32(c6, vacc6x01); c6 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 235 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 236 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 237 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 238 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 239 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 240 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()
|
/aosp_15_r20/external/XNNPACK/src/f32-gemm/gen/ |
H A D | 6x2-minmax-neon-lane-ld64.c | 141 vst1_f32(c0, vacc0x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() 143 vst1_f32(c1, vacc1x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() 145 vst1_f32(c2, vacc2x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() 147 vst1_f32(c3, vacc3x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() 149 vst1_f32(c4, vacc4x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() 151 vst1_f32(c5, vacc5x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
|
H A D | 6x2-minmax-neonfma-lane-ld64.c | 171 vst1_f32(c0, vacc0x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() 173 vst1_f32(c1, vacc1x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() 175 vst1_f32(c2, vacc2x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() 177 vst1_f32(c3, vacc3x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() 179 vst1_f32(c4, vacc4x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() 181 vst1_f32(c5, vacc5x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
|
H A D | 6x8-minmax-neon-lane-ld64.c | 230 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 231 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 232 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 233 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 234 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 235 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
|
H A D | 6x8-minmax-neonfma-lane-ld64.c | 230 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 231 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 232 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 233 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 234 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 235 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
H A D | 6x8-minmax-neon-dup-ld64.c | 242 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() 243 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() 244 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() 245 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() 246 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() 247 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
|
H A D | 6x8-minmax-neonfma-dup-ld64.c | 242 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() 243 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() 244 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() 245 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() 246 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() 247 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
H A D | 4x2-minmax-neon-lane-ld64.c | 113 vst1_f32(c0, vacc0x01); in xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64() 115 vst1_f32(c1, vacc1x01); in xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64() 117 vst1_f32(c2, vacc2x01); in xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64() 119 vst1_f32(c3, vacc3x01); in xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64()
|
H A D | 5x8-minmax-neon-lane-ld64.c | 203 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() 204 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() 205 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() 206 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() 207 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
|
/aosp_15_r20/external/XNNPACK/src/f32-igemm/gen/ |
H A D | 6x2-minmax-neon-lane-ld64.c | 171 vst1_f32(c5, vacc5x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() 173 vst1_f32(c4, vacc4x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() 175 vst1_f32(c3, vacc3x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() 177 vst1_f32(c2, vacc2x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() 179 vst1_f32(c1, vacc1x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() 181 vst1_f32(c0, vacc0x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
|
H A D | 6x2-minmax-neonfma-lane-ld64.c | 201 vst1_f32(c5, vacc5x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() 203 vst1_f32(c4, vacc4x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() 205 vst1_f32(c3, vacc3x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() 207 vst1_f32(c2, vacc2x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() 209 vst1_f32(c1, vacc1x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() 211 vst1_f32(c0, vacc0x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
|
H A D | 6x8-minmax-neonfma-lane-ld64.c | 255 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 256 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 257 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 258 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 259 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 260 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
H A D | 6x8-minmax-neon-lane-ld64.c | 255 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 256 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 257 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 258 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 259 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 260 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
|
H A D | 6x8-minmax-neon-dup-ld64.c | 267 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() 268 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() 269 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() 270 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() 271 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() 272 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
|
H A D | 6x8-minmax-neonfma-dup-ld64.c | 267 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() 268 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() 269 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() 270 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() 271 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() 272 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
/aosp_15_r20/external/XNNPACK/src/f32-gemm/gen-inc/ |
H A D | 6x8inc-minmax-neonfma-lane-ld64.c | 232 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 233 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 234 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 235 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 236 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 237 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
|
H A D | 6x8inc-minmax-neon-lane-ld64.c | 232 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 233 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 234 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 235 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 236 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 237 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
|
H A D | 6x8inc-minmax-neonfma-dup-ld64.c | 244 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() 245 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() 246 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() 247 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() 248 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() 249 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
|
H A D | 6x8inc-minmax-neon-dup-ld64.c | 244 vst1_f32(c5, vacc5x01); c5 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() 245 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() 246 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() 247 vst1_f32(c2, vacc2x01); c2 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() 248 vst1_f32(c1, vacc1x01); c1 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() 249 vst1_f32(c0, vacc0x01); c0 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/sgemm/ |
H A D | 6x8-neon.c | 278 vst1_f32(c0, vget_low_f32(vacc0x0123)); in pytorch_sgemm_ukernel_6x8__neon() 280 vst1_f32(c1, vget_low_f32(vacc1x0123)); in pytorch_sgemm_ukernel_6x8__neon() 282 vst1_f32(c2, vget_low_f32(vacc2x0123)); in pytorch_sgemm_ukernel_6x8__neon() 284 vst1_f32(c3, vget_low_f32(vacc3x0123)); in pytorch_sgemm_ukernel_6x8__neon() 286 vst1_f32(c4, vget_low_f32(vacc4x0123)); in pytorch_sgemm_ukernel_6x8__neon() 288 vst1_f32(c5, vget_low_f32(vacc5x0123)); in pytorch_sgemm_ukernel_6x8__neon()
|