| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 122 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
343 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
558 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c | 122 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
343 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
558 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
|
| D | 5x5p2-minmax-neon-4x4-acc2.c | 119 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
341 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
557 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
|
| D | 5x5p2-minmax-neonfma-4x4-acc2.c | 119 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
341 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
557 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 148 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
369 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
584 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 148 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
369 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
584 v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
|
| D | 5x5p2-minmax-sse-4x4-acc2.c | 136 __m128 vo3p1 = _mm_mul_ps(vi4x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
354 __m128 vo3p1 = _mm_mul_ps(vi4x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
584 __m128 vo3p1 = _mm_mul_ps(vi4x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|