| /external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neonfp16arith-3x4.c | 156 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4() local
169 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
172 vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
175 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
337 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4() local
350 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
353 vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
356 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
511 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4() local
519 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4()
[all …]
|
| D | 5x5p2-minmax-neonfp16arith-3x4-acc2.c | 156 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2() local
169 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
172 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
175 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
340 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2() local
353 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
356 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
359 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
517 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2() local
525 vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfp16arith-2x4-acc2.c | 135 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2() local
148 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
150 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
281 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2() local
294 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
296 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
421 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2() local
430 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
432 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2()
|
| D | 5x5p2-minmax-neonfp16arith-2x4.c | 135 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4() local
148 vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
150 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
279 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4() local
292 vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
294 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
417 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4() local
426 vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
428 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4()
|
| D | 5x5p2-minmax-neonfp16arith-2x4-acc3.c | 135 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3() local
148 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
150 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
283 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3() local
296 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
298 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
425 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3() local
434 vo1p1 = vfma_laneq_f16(vo1p1, vi2x2345, vw01234567, 6); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
436 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3()
|
| D | 5x5p2-minmax-neonfp16arith-1x4-acc4.c | 114 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local
125 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4()
224 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local
235 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4()
329 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4() local
337 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4()
|
| D | 5x5p2-minmax-neonfp16arith-1x4.c | 114 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local
125 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4()
221 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local
232 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4()
323 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4() local
331 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4()
|
| D | 5x5p2-minmax-neonfp16arith-1x4-acc2.c | 114 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local
125 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2()
222 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local
233 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2()
325 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2() local
333 vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2()
|
| D | 5x5p2-minmax-neonfp16arith-1x4-acc3.c | 114 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local
125 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3()
223 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local
234 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3()
327 const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3() local
335 vo0p2 = vfma_laneq_f16(vo0p2, vi2x2345, vw89ABCDEF, 3); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3()
|
| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 185 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local
198 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
201 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
204 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
368 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local
381 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
384 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
387 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
544 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2() local
552 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-3x4.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
171 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
174 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
339 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
352 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
355 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
358 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
513 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
521 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 185 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
198 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
201 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
204 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
365 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
378 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
381 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
384 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
538 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
546 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 185 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
198 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
201 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
204 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
368 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
381 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
384 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
387 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
544 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
552 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 185 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
198 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
201 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
204 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
365 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
378 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
381 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
384 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
538 const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
546 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4-acc2.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
171 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
174 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
342 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
355 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
358 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
361 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
519 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
527 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-3x4-acc2.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
171 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
174 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
342 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
355 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
358 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
361 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
519 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
527 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4.c | 158 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
171 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
174 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
339 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
352 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
355 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
358 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
513 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
521 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
[all …]
|
| D | 5x5p2-minmax-neon-2x4-acc2.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
150 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
283 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
296 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
298 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
423 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
432 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
434 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-2x4.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
150 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
152 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
281 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
294 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
296 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
419 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
428 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
430 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
|
| D | 5x5p2-minmax-neon-2x4.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
150 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
281 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
294 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
296 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
419 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
428 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
430 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
|
| D | 5x5p2-minmax-neonfma-2x4-acc2.c | 137 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
150 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
152 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
283 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
296 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
298 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
423 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
432 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
434 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
127 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
223 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
234 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
325 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
333 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
127 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
224 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
235 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
327 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
335 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
127 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
225 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
236 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
329 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
337 vo0p2 = vfmaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 116 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
127 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
225 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
236 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
329 const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
337 vo0p2 = vmlaq_lane_f32(vo0p2, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|