Implement register spilling

[doe300]

We need to implement register spilling to be able to support more complex kernels.

If the size of spilled registers (times 12 QPUs!) is small enough, we could store them in VPM and save from accessing memory. Access to the spilled registers would still need to be synchronized via the hardware-mutex.

The actual problem of this implementation is not the spilling/loading of locals, but in determining the minimum number of registers to spill.

(see doe300/VC4CL#24)

[nomaddo]

Notice:
To spill registers, cache incoherence will be a problem between VPM and TMU.
DMA Load & Store doesn't use L2-cache, but TMU does.

Example (assembled by py-videocore):

import numpy as np

from videocore.assembler import qpu
from videocore.driver import Driver

@qpu
def dma_store(asm, preload):
    mov(ra0, uniform)
    mov(ra1, uniform)

    shl(r0, element_number, 2)
    iadd(r0, r0, ra0)

    if preload: # ATTENTION: difference of of the value `preload` chenge the result
        # we don't use the value loaded here
        mov(tmu0_s, r0)
        nop(sig='load tmu0')

    setup_vpm_write()
    mov(vpm, element_number)

    setup_dma_store(nrows=1)
    start_dma_store(ra0) # store the value
    wait_dma_store()

    mov(tmu0_s, r0)
    nop(sig='load tmu0') # load the value, which is just store into the buffer

    setup_vpm_write()
    mov(vpm, r4)

    setup_dma_store(nrows=1)
    start_dma_store(ra1) # resultにstore
    wait_dma_store()

    exit()
	
with Driver() as drv:
    print('----- enable preload -----')
    buffer = drv.alloc(16, 'uint32')
    buffer[:] = 0
    result = drv.alloc(16, 'uint32')
    result[:] = 0

    print('[Before]')
    print(buffer)
    print(result)

    drv.execute(
        n_threads=1,
        program=drv.program(dma_store, True),
        uniforms=[buffer.address, result.address]
    )

    print('[After]')
    print(buffer)
    print(result)
	
with Driver() as drv:
    print('----- disable preload -----')
    buffer = drv.alloc(16, 'uint32')
    buffer[:] = 0
    result = drv.alloc(16, 'uint32')
    result[:] = 0

    print('[Before]')
    print(buffer)
    print(result)

    drv.execute(
        n_threads=1,
        program=drv.program(dma_store, False),
        uniforms=[buffer.address, result.address]
    )

    print('[After]')
    print(buffer)
    print(result)

[doe300]

Yeah, I think this is also the problem in #30