一个较大的模型在infer时出现 _mm256_storeu_ps(d + PACK_UNIT * 1, t1); about mnn HOT 2 CLOSED

        TRANSPOSE_8x8;

        _mm256_storeu_ps(d + PACK_UNIT * 0, t0);
        _mm256_storeu_ps(d + PACK_UNIT * 1, t1);
        _mm256_storeu_ps(d + PACK_UNIT * 2, t2);
        _mm256_storeu_ps(d + PACK_UNIT * 3, t3);
        _mm256_storeu_ps(d + PACK_UNIT * 4, t4);
        _mm256_storeu_ps(d + PACK_UNIT * 5, t5);
        _mm256_storeu_ps(d + PACK_UNIT * 6, t6);
        _mm256_storeu_ps(d + PACK_UNIT * 7, t7);
    }
}
auto areaRemain  = areaC4 * PACK_UNIT;
auto depthRemain = depthC4 * PACK_UNIT;
// Down
int remain = depth - depthRemain;
if (remain > 0) {
    float* dstPlane       = depthC4 * dstAreaOffset * PACK_UNIT + dst;
    const float* srcPlane = src + depthC4 * srcAreaOffset * PACK_UNIT;
    {
        for (int x = 0; x < areaC4; ++x) {
            auto s  = srcPlane + PACK_UNIT * x;
            auto d  = dstPlane + PACK_UNIT * PACK_UNIT * x;
            auto r0 = _mm256_loadu_ps(s + 0 * srcAreaOffset);
            auto r1 = _mm256_setzero_ps();
            auto r2 = _mm256_setzero_ps();
            auto r3 = _mm256_setzero_ps();
            auto r4 = _mm256_setzero_ps();
            auto r5 = _mm256_setzero_ps();
            auto r6 = _mm256_setzero_ps();
            auto r7 = _mm256_setzero_ps();
            switch (remain) {
                case 7:
                    r6 = _mm256_loadu_ps(s + 6 * srcAreaOffset);
                case 6:
                    r5 = _mm256_loadu_ps(s + 5 * srcAreaOffset);
                case 5:
                    r4 = _mm256_loadu_ps(s + 4 * srcAreaOffset);
                case 4:
                    r3 = _mm256_loadu_ps(s + 3 * srcAreaOffset);
                case 3:
                    r2 = _mm256_loadu_ps(s + 2 * srcAreaOffset);
                case 2:
                    r1 = _mm256_loadu_ps(s + 1 * srcAreaOffset);
                default:
                    break;
            }

            TRANSPOSE_8x8;

            _mm256_storeu_ps(d + PACK_UNIT * 7, t7);
            _mm256_storeu_ps(d + PACK_UNIT * 6, t6);
            _mm256_storeu_ps(d + PACK_UNIT * 5, t5);
            _mm256_storeu_ps(d + PACK_UNIT * 4, t4);
            _mm256_storeu_ps(d + PACK_UNIT * 3, t3);
            _mm256_storeu_ps(d + PACK_UNIT * 2, t2);
            _mm256_storeu_ps(d + PACK_UNIT * 1, t1);
            _mm256_storeu_ps(d + PACK_UNIT * 0, t0);
        }
    }
    for (int x = areaRemain; x < area; ++x) {
        for (int y = 0; y < remain; y++) {
            dstPlane[PACK_UNIT * x + y] = srcPlane[y * srcAreaOffset + x];
        }
        for (int y = remain; y < PACK_UNIT; y++) {
            dstPlane[PACK_UNIT * x + y] = 0;
        }
    }
}
// Right
for (int z = 0; z < depthC4; ++z) {
    float* dstPlane       = z * dstAreaOffset * PACK_UNIT + dst;
    const float* srcPlane = src + z * srcAreaOffset * PACK_UNIT;
    for (int x = areaRemain; x < area; ++x) {
        float s0 = srcPlane[x];
        float s1 = srcPlane[x + srcAreaOffset];
        float s2 = srcPlane[x + srcAreaOffset * 2];
        float s3 = srcPlane[x + srcAreaOffset * 3];
        float s4 = srcPlane[x + srcAreaOffset * 4];
        float s5 = srcPlane[x + srcAreaOffset * 5];
        float s6 = srcPlane[x + srcAreaOffset * 6];
        float s7 = srcPlane[x + srcAreaOffset * 7];
        _mm256_storeu_ps(dstPlane + PACK_UNIT * x, _mm256_set_ps(s7, s6, s5, s4, s3, s2, s1, s0));
    }
}