S-SMP

The SPC700 is the CPU portion of the S-SMP processor used to run the sound and music of the SNES. It is an 8-bit CPU with capabilities similar to the 6502.

Memory Layout
64 kilobytes of RAM are mapped across the 16-bit memory space of the SPC-700. Some regions of this space are overlaid with special hardware functions.

The region at $FFC0-FFFF will normally read from the 64-byte IPL ROM, but the underlying RAM can always be written to, and the high bit of the Control register $F1 can be cleared to unmap the IPL ROM and allow read access to this RAM.

Registers
Write-only registers will read back as $00.

$F0 Test
This undocumented register responds to writes only when the P flag is clear.

7 bit  0 .I.E TRWH


 * I (bit 6) - Internal wait state.
 * E (bit 4) - External wait state.
 * T (bit 3) - Timers enable.
 * R (bit 2) - Ram disable.
 * W (bit 1) - Ram writable.
 * H (bit 0) - Timers disable.

$F1 Control
This provides a way for the SPC-700 to reset the read ports ($F4-F7) without the SNES CPU having to write them externally. It also starts and stops the 3 timers.

7 bit  0 I.CC .210


 * 0 (bit 0) - Enables timer 0 when set. A transition from clear to set resets the internal interval counter to 0.
 * 1 (bit 1) - Enables timer 1.
 * 2 (bit 2) - Enables timer 2.
 * C (bit 4) - If set will reset the value the SPC will read from ports 0 and 1 ($F4, $F5) to $00.
 * C (bit 5) - If set will reset ports 2 and 3 ($F6, $F7).
 * I (bit 7) - Will enable the IPL ROM if set.

At reset this register is initialized as if $80 was written to it.

The function of bit 7 enabling the IPL ROM is not documented in the SNES Development Book.

$F2-F3 DSP
Write $F2 to select a DSP register, then a value can be read or written to that DSP register via $F3.


 * Writing $F2 with the high bit set will select a DSP register according to the lower 7 bits, but it will be read-only.
 * The high bit of $F2 will always read back as 0.

$F4-F7 Port 0-3
These 4 ports allow communication with the SNES CPU. There are 8 stored values, each is a one-way communication written from one side, and readable only from the other side. Each port therefore has two separate one-way values, each seen from only either the SNES CPU or the SPC-700.

If a port is read on the same cycle it is written, an incorrect value will result. For this reason, common practice is to read a port in a loop until the value changes, and then read it once more to ensure the correct value is read. (A single port can be used this way to indicate that a message is ready, and the other 3 ports could be safely read only once with the assumption that the other CPU will not write to them once the ready indication was given.)

At reset these registers are initialized to $00.

$FA-FC Timer 0-2
When enabled via $F1, the 3 timers will internally count at a rate of 8 KHz (timers 0,1) or 64 KHz (timer 2), and when this interval value has been exceeded, they will increment their external counter result ($FD-FF) and begin again.

$FD-FF Counter 0-2
The 4-bit result of the three timers counts up every time the interval is reached.

Reading these registers resets each counter to 0 immediately after the read. The upper 4 bits will always read as 0.

DSP Registers
A DSP register can be selected with $F2, after which it can be read or written at $F3. Often it is useful to load the register address into A, and the value to send in Y, so that MOV $F2, YA can be used to do both in one 16-bit instruction.

The DSP register address space only has 7 bits. The high bit of $F2, if set, will make the selected register read-only via $F3.

When initializing the DSP registers for the first time, take care not to accidentally enable echo writeback via FLG, because it will immediately begin overwriting values in RAM.

Voices
There are 8 voices, numbered 0 to 7. Each voice X has 10 registers in the range $X0-$X9.

P
Sample pitch is a 14-bit value controlling the rate the BRR sound sample will be played back.

Rate: P x 32,000 Hz / $1000

A pitch of $1000 will play back the sample at the SNES native samplerate of 32,000 Hz.

The pitch can go as high as $3FFF, almost two octaves above $1000. Pitches above $1000 will be subject to some aliasing from samples that are skipped over.

The pitch can go all the way down to 0, where it is halted. Pitches below $1000 will be lacking in higher frequencies, and there is not very much precision as the pitch value approaches 0.

SCRN
This points to an entry in the sample source directory (DIR). Changing this will not immediately change the voice's sample without a key on (KON), but if a looping sample is playing it can be used to change the loop point without a key on.

ADSR
This controls an Attack-Decay-Sustain-Release envelope that automatically adjusts the sample's envelope volume (ENVX) over time.


 * E: 1 to enable ADSR envelope, otherwise 0 uses GAIN instead.
 * A: Attack speed. $F for instant.
 * D: Decay speed, time to decay from full volume to the sustain level after the initial attack.
 * L: Sustain level.
 * R: Sustain release, speed of decay to 0 after note off.

GAIN
This register has 5 modes:


 * 0VVV VVVV sets ENVX directly.
 * 110V VVVV Linear slide up to 100% volume with rate V.
 * 111V VVVV Bent-line (fast to 75%, then slower to 100%) slide up with rate V.
 * 100V VVVV Linear slide down to 0% volume with rate V.
 * 101V VVVV Exponential slide down to 0% volume with rate V.

Global
Other DSP registers apply globally, rather than to a specific voice.

KON
This immediately starts the sample from its beginning, as designated by SCRN. If ADSR is enabled, its envelope will also start from the beginning.

FLG

 * R: soft reset prevents KON and mutes all voices.
 * M: mutes all voices.
 * E: set to 1 to disable echo, if 0 the echo unit will actively overwrite memory in the echo region (ESA). Do not write 0 to this bit unless ESA/EDL have been prepared.
 * N: sets the noise generator frequency.

At reset the high 3 bits of FLG are set.

ENDX
Each bit will read as set once a voice has finished playing a BRR sample block with the Source End flag set. Writing any value to this register will reset all the bits to 0, otherwise they remain set until the voice receives a new note on (KON).

Instruction Set
See: SPC-700 Instruction Set

IPL Boot ROM
See: Booting the SPC700

The IPL Boot ROM is a small built-in program responsible for initializing the SPC-700, and making it ready to transfer a program from the SNES CPU, then execute it. It normally resides at $FFC0, with its code beginning at this address, but if desired it can be unmapped using the $F1 control register, and replaced with the underlying RAM.

When the SNES is reset, the SPC-700 will also reset and begin executing the IPL.

IPL might stand for "Initial Program Load".

The high level process of the IPL is described below:


 * 1) Reset: Stack pointer = $EF. Zero-page from $00-$EF is set to $00. (Note: this leaves the top 16-bytes of the stack page unused by default.)
 * 2) Signal ready: Port 0 = $AA. Port 1 = $BB.
 * 3) Wait for signal: Loop until $CC is read from port 0.
 * 4) Read address: Read a 2 byte address from port 2 (low) and 3 (high).
 * 5) Acknowledge: Read value from port 0 and write it to port 0, confirming the signal was received.
 * 6) Begin: Read value from port 1, if 0 begin executing code at address read in step 4, otherwise begin reading data (step 7).
 * 7) Begin Transfer: Loop until read port 0 reads 0. Set 8-bit counter to 0, then proceed to step 8.
 * 8) Transfer Loop:
 * 9) * Read a byte from port 1 and write to the destination address.
 * 10) * Write the value read from port 0 to port 0 to acknowledge receipt of the byte.
 * 11) * Increment the destination address, and increment the 8-bit counter.
 * 12) * Wait until port 0 reads equal to the new counter (SNES will increment its own counter and send it to port 0 to signal the next byte), and repeat step 8, otherwise...
 * 13) * If port 0 reads greater than the new counter (i.e. SNES increments by 2 or more before writing port 0): write back port 0 to acknowledge, transfer ends and return to step 4.

On the SNES side, to load your program into the SPC:


 * 1) Wait for the ready signal: port 0 = $AA, port 1 = $BB (IPL 2).
 * 2) Write the first destination address to ports 2+3, write any non-zero value to port 1, then write $CC to port 0 to begin (IPL 3).
 * 3) Wait for port 0 to read back $CC (IPL 5).
 * 4) Set an 8-bit counter to $00. Write the first byte of data to port 1, then write the counter value ($00) to port 0 (IPL 7). Read port 0 until it is equal to the counter $00.
 * 5) For each byte of data: write the data to port 1, increment the 8-bit counter and write it to port 0, then read port 0 until is becomes equal to the counter. Repeat until finished. (IPL 8)
 * 6) After the last byte is written, write the next destination address to ports 2+3, write $00 to port 1, then increment your counter twice and write it to port 0. (IPL 4)
 * 7) The SPC-700 will begin executing code at the destination address now.

Most often only one data transfer is necessary, but by writing a non-zero value to port 1 in SNES step 6, we can instead initiate another transfer at the new address. When doing this, if your counter is 0 after incrementing twice, increment it a third time to be non-zero before writing it to port 0. This is because a value of 0 in port 0 will also signal the first byte of the transfer in the next step.

It is most common to place the data program at $0200, just above the stack page, and start execution from there, but this is not a requirement.

BRR Samples
Sound samples are stored in the BRR (bit-rate-reduction) data format.

BRR is composed of 16-sample blocks, each of which is stored in 9 bytes of data. This is a 1 byte control block, followed by 8 bytes containing 16 4-bit samples to be decoded (high nibble first).

When a block with the end (E) flag set finishes, the channel will automatically stop unless the loop (L) flag is also set for this block. If the loop flag is set, the loop point will be read based on the channel's current SCRN and sample playback will continue from there. Because of the block structure, loop points must lie on 16-sample boundaries. The loop flag has no effect on blocks without the end flag set.

The left-shift (S) value specifies the overall magnitude of values. This is simply a left shift applied to each sample nibble before BRR filter decoding.

Each sample nibble is a signed 4-bit value in the range of -8 to +7. After being shifted, one of four BRR decoding filters (F) is applied, which may include the last 0-2 decoded samples in the result. Filter 0 will just use the shifted nibble directly.

Sound samples should normally start with a filter-0 block, to prevent leftover BRR samples from the previously playing sample from having an effect on decoding.

Filters
The four BRR filters form an adaptive PCM (ADPCM) encoding scheme, using the previous few samples to predict the next one with a low amount of input data. By choosing the best fitting filter (i.e. adapting) for each block of 16 samples, the result can more closely match a desired waveform.

Each filter adds the shifted nibble to two previously decoded samples, each multiplied by a coefficient:

Filter 0 is the non-filter, which just uses the shifted nibble result directly. The previous samples are not used. Sound samples should normally begin with a filter 0 block to prevent leftover results from the previous sound affecting the new one.

Filter 1 includes some of the previously encoded sample.

Filters 2 and 3 include some of two previously encoded samples.

As the filter number increases, the high frequency bandwidth (sharpness) of the 4-bit sample encoding is traded for better low frequency (smoothness) encoding. Because most natural sounds have stronger low frequencies and weaker high frequencies, this allows the bandwidth allocation of each block to be adapted to better suit the frequency spectrum of the sound at that moment.

Gaussian Interpolation
Finally, after the BRR sample data is decoded, when a sample is played back, because the rate (P) is adjustable, an filter is used to smoothly interpolate values between the decoded samples. This is a gaussian interpolator which blends a combination of a neighbourhood of 4 adjacent samples according to a lookup table. This reduces high frequency aliasing across a range of playback rates.

Playback begins with the interpolator centred on the second sample. If your key-on does not begin with an ADSR attack envelope, you may wish to have three 0 samples at the beginning of your sound sample to ensure it begins at 0 and smoothly moves from there (avoiding clicks/pops).

Because of a bug in the DSP's gaussian interpolation filter, 3 maximum-negative values in a row can cause an overflow of the resulting sample, creating a loud pop. This can be intentionally exploited, but normally it is advised avoid this.

Links

 * SPC 700 Documentation - Article by Gau.
 * SPC700 Reference - Superfamicom.org wiki article.
 * Bit Rate Reduction - Superfamicom.org wiki article, documents BRR sample format.
 * SNES APU DSP BRR Samples - Fullsnes documentation of BRR sample format.
 * How to Write to DSP Registers Without any SPC-700 Code - Superfamicom.org wiki, demonstrating that the DSP can be written directly through the IPL.