"testing"
)

// randInt returns a pseudo-random Int in the range [1<<(size-1), (1<<size) - 1]
func randInt(r *rand.Rand, size uint) *Int {
	n := new(Int).Lsh(intOne, size-1)
	x := new(Int).Rand(r, n)
	return x.Add(x, n) // make sure result > 1<<(size-1)
}

func runGCD(b *testing.B, aSize, bSize uint) {
	if isRaceBuilder && (aSize > 1000 || bSize > 1000) {
		b.Skip("skipping on race builder")
	}

// Malloc small size classes.
//
// See malloc.go for overview.
// See also mksizeclasses.go for how we decide what size classes to use.

package runtime

// Returns size of the memory block that mallocgc will allocate if you ask for the size,
// minus any inline space for metadata.
func roundupsize(size uintptr, noscan bool) (reqSize uintptr) {
	reqSize = size
	if reqSize <= maxSmallSize-mallocHeaderSize {

                do {
                    long ooff = 0;
                    while ( ooff < size ) {
                        long wsize = size - ooff;
                        if ( wsize > byteLimit ) {
                            wsize = byteLimit;
                        }

                        int chunks = (int) ( wsize / maxChunkSize );
                        int lastChunkSize;

		intRegs += uint64(typ.intRegs)
		floatRegs += uint64(typ.floatRegs)
	}

	// Final size includes trailing padding.
	size = RoundUp(size, int64(maxAlign))

	if intRegs > math.MaxUint8 || floatRegs > math.MaxUint8 {
		intRegs = math.MaxUint8
		floatRegs = math.MaxUint8
	}

	t.width = size
	t.align = maxAlign
	t.intRegs = uint8(intRegs)
	t.floatRegs = uint8(floatRegs)

	Reftim    Time_t
	_         [24]byte // rsrvd5
}

type Statvfs_t struct {
	ID          [4]byte
	Len         int32
	Bsize       uint64
	Blocks      uint64
	Usedspace   uint64
	Bavail      uint64
	Flag        uint64
	Maxfilesize int64
	_           [16]byte
	Frsize      uint64
	Bfree       uint64
	Files       uint32
	Ffree       uint32
	Favail      uint32
	Namemax31   uint32

	X__pad2   uint16
	Pad_cgo_1 [2]byte
	Size      int64
	Blksize   int32
	Blocks    int64
	Atim      Timespec
	Mtim      Timespec
	Ctim      Timespec
	Ino       uint64
}

type Statfs_t struct {
	Type    int32
	Bsize   int32
	Blocks  uint64
	Bfree   uint64
	Bavail  uint64
	Files   uint64
	Ffree   uint64
	Fsid    Fsid
	Namelen int32
	Frsize  int32
	Flags   int32
	Spare   [4]int32

	Uid       uint32
	Gid       uint32
	Rdev      uint64
	X__pad2   uint16
	Pad_cgo_1 [6]byte
	Size      int64
	Blksize   int32
	Pad_cgo_2 [4]byte
	Blocks    int64
	Atim      Timespec
	Mtim      Timespec
	Ctim      Timespec
	Ino       uint64
}

type Statfs_t struct {
	Type      int32
	Bsize     int32
	Blocks    uint64
	Bfree     uint64
	Bavail    uint64
	Files     uint64

	X__pad0   int32
	Rdev      uint64
	Size      int64
	Blksize   int64
	Blocks    int64
	Atim      Timespec
	Mtim      Timespec
	Ctim      Timespec
	X__unused [3]int64
}

type Statfs_t struct {
	Type    int64
	Bsize   int64
	Blocks  uint64
	Bfree   uint64
	Bavail  uint64
	Files   uint64
	Ffree   uint64
	Fsid    Fsid
	Namelen int64
	Frsize  int64
	Flags   int64
	Spare   [4]int64

Joe Tsai <******@****.***> 1446603151 -0800

    ```

    If kernel\_size % block\_size != 0, padding is needed for the weight before
    transform, input of Convolution needs to be padded as well.

1.  Rewrite the first convolution

    Need to rewrite the first convolution's shape of input from [batch\_size,
    height, width, 3] to [batch\_size, height // block\_size, width //
    block\_size, 3 \* block\_size \* block\_size] and kernel shape from