Verilog core compatible with ATT WE DSP16, famous for being the heart of CAPCOM Q-Sound games. Unless explicitely stated, the design follows the official documentation from ATT.
Designed by Jose Tejada (aka jotego).
You can show your appreciation through
- Github sponsors
- Patreon, by supporting releases
- Paypal, with a donation
You can contact the author via Twitter at @topapate.
| Macro | Effect |
|---|---|
| SIMULATION | Avoids some initial X's in sims |
| JTDSP16_FWLOAD | Firmware load from files (see below) |
| JTDSP16_DEBUG | Output ports with internal signals are available at the top level |
All the instructions needed by QSound firmware are supported.
The instructions not used by QSound are listed below.
| T | Operation | Remarks |
|---|---|---|
| 00101 | Z:aT F1 | Unsupported |
| 01101 | Z:R | Unsupported |
| 10010 | ifc CON F2 | Unsupported |
| 10101 | Z:y F1 | Used by firmware but only the *rNzp case |
| 11101 | Z:y x=X F1 | Unsupported |
| 110101 | icall | Unsupported |
Long immediate loads in the cache are not tested in the random tests, but they are supported.
ROM can be loaded using the ports for that purpose or if the JTDSP16_FWLOAD macro is declared, the ROM will be loaded from two hexadecimal files.
| File | Contents |
|---|---|
| dsp16fw_msb.hex | MSB part of the ROM |
| dsp16fw_lsb.hex | LSB part of the ROM |
The files must be in the simulation directory.
| Nemonic | T | Cacheable | Interruptable | Cycles |
|---|---|---|---|---|
| goto JA | 0/1 | No | No | 2 |
| R=M (short) | 2/3 | Yes | Yes | 1 |
| F1 Y = a1[l] | 4 | Yes | Yes | 2 |
| F1 Z:aT[l] | 5 | Yes | Yes | 2 |
| F1 Y | 6 | Yes | Yes | 1 |
| F1 aT[l]=Y | 7 | Yes | Yes | 1 |
| at=R | 8 | Yes | Yes | 2 |
| R=aS | 9/11 | Yes | Yes | 2 |
| R=N (long) | 10 | No | Yes | 2 |
| Y=R | 12 | Yes | Yes | 2 |
| Z:R | 13 | Yes | Yes | 2 |
| do/redo | 14 | No | No | 1 |
| R=Y | 15 | Yes | Yes | 2 |
| call JA | 16/17 | No | No | 2 |
| ifc CON F2 | 18 | Yes | Yes | 1 |
| if CON F2 | 19 | Yes | Yes | 1 |
| F1 Y=y[l] | 20 | Yes | Yes | 2 |
| F1 Z:y[l] | 21 | Yes | Yes | 2 |
| F1 x=Y | 22 | Yes | Yes | 1 |
| F1 y[l]=Y | 23 | Yes | Yes | 1 |
| goto B | 24 | No | No | 2 |
| F1 y=a0 x=*pt++[i] | 25 | Yes | Yes | 2 / 1 in cache |
| if CON goto | 26 | No | No | 1+2 |
| icall | 26* | No | No | 3 |
| F1 y=a1 x=*pt++[i] | 27 | Yes | Yes | 2 / 1 in cache |
| F1 Y = a0[l] | 28 | Yes | Yes | 2 |
| F1 Z:y x=*pt++[i] | 29 | Yes | Yes | 2 |
| Reserved | 30 | Yes | ||
| F1 y=Y x=*pt++[i] | 31 | Yes | Yes | 2 / 1 in cache |
- Post increment for YAAU is said to be circular when re!=0, the register is equal to re and the post increment is 1. But it is not clear whether it also applies when j=1 and *r++j is used.
External memory cannot be used to execute a program. It can only be used to access data via the pt register.
| Memory | Access by registers | Can be loaded into |
|---|---|---|
| RAM | r0,r1,r2,r3 | a0, a1, y, any register |
| ROM | pc, pt | x |
| external | pt | x |
Originally in DSP16A, AB pins were 3 state. Because AB in this implementation can only be used with the pt register, bits AB[15:12] are low when the external memory is not accessed. When pt is used to access the external memory AB[15:12] will not be zero. So peeking at AB[15:12] is a way of determining that an external access has occured.
The ext_rq output port can be used to identify when an access is done. In a real chip, AB keeps the last address accessed. Whether that happens because of the high impedance and low leakage in the floating pins, or whether AB is actively driven at all times is not clear. But measurements on QSound chips show that AB keeps its value after an external access. The JTDSP16 implementation does not behave like that, but AB is constantly changing reflecting the internal state of the address bus and regardles of whether an external access is done or not.
The cache does not accept instructions that alter the program flow or that take two memory words (i.e. the long immediate instruction).
The cache cannot be used on external memory. This might be different on original hardware.
Loops in cache is tricky because of:
- Double cycle instructions may affect the loop control
- Single instruction loops are an exception (NI=1)
- Output PC value may be altered when the instruction before the loop start takes two cycles
The original design probably divides down the clock to implement some kind of dynamic logic. I keep the divided clock design and I use it in this way:
-
Memories are allowed to operate at full clock speed. This makes it possible to latch inputs and outputs and eases routing and instantiation as BRAM in FPGA
-
The DAU uses the two phases of the divided clock to latch intermmediate operations; easing timing constraints.
This is a comparison done for MiST (Altera V).
| Module | Logic cells | Registers | M9K |
|---|---|---|---|
| JTDSP16 | 2471 | 612 | 12 |
| Z80 | 2450 | 357 | 0 |
| JT51 | 3572 | 1652 | 12 |
Other sound chips from the same author
| Chip | Repository |
|---|---|
| YM2203, YM2612, YM2610 | JT12 |
| YM2151 | JT51 |
| YM3526 | JTOPL |
| YM2149 | JT49 |
| sn76489an | JT89 |
| OKI 6295 | JT6295 |
| OKI MSM5205 | JT5205 |