Difference between revisions of "DSP"

From xboxdevwiki
Jump to: navigation, search
(→‎Related links)
(Add explicit link to family manual, because people don't seem to find it)
 
(31 intermediate revisions by 2 users not shown)
Line 1: Line 1:
−
The DSPs in the APU are probably "Parthus MediaStream" DSP core (DSP24210/DSP2420?).
+
The DSPs in the APU are probably "Parthus MediaStream" DSP cores (probably the 24-bit DSP2420 "Mozart" core).
  
−
Those are similar to Motorola DSP56362 (DSP56300 Family). If so, the datasheet can be found at http://www.nxp.com/docs/en/data-sheet/DSP56362.pdf (Also see "Documentation" section in said datasheet for the related documentation)
+
Those are similar to Motorola DSP56362 (DSP56300 Family). If so, the datasheet can be found at http://www.nxp.com/docs/en/data-sheet/DSP56362.pdf (Also see "Documentation" section in said datasheet for the related documentation, like [https://www.nxp.com/docs/en/reference-manual/DSP56300FM.pdf the family manual])
 +
 
 +
== Memory Size ==
 +
 
 +
{| class="wikitable"
 +
|-
 +
!
 +
! Program RAM Size
 +
! X Data RAM Size
 +
! Y Data RAM Size
 +
! MIXBUF Size
 +
|-
 +
! GP
 +
| 4096 x 24-bit
 +
| 4096 x 24-bit
 +
| 2048 x 24-bit
 +
| 992{{FIXME|reason=This is probably 1024 words, but the last 32 words were not working as intended during my testing}} x 24-bit
 +
|-
 +
! EP
 +
| 4096 x 24-bit
 +
| 3072 x 24-bit
 +
| 256 x 24-bit
 +
| n/a
 +
|}
 +
 
 +
MIXBUF is accessible at X:$001400 in the GP.
 +
 
 +
Other datasheets for similar DSPs suggest that the memory sizes might be different if instruction cache or switch mode are toggled.
 +
It is currently unknown if the DSPs in the Xbox APU support a similar feature{{FIXME}}.
  
 
== DMA ==
 
== DMA ==
Line 7: Line 35:
 
''This section is very incomplete and not much was tested on hardware either''
 
''This section is very incomplete and not much was tested on hardware either''
  
−
Sample formats:
+
DMA is controlled using peripheral registers:
 +
 
 +
* 0xFFFFD4: Memory address of next command block
 +
* 0xFFFFD5: DMA_START_BLOCK{{FIXME|reason=How is this used?}}
 +
* 0xFFFFD6: DMA_CONTROL{{FIXME|reason=Explain the bits}}
 +
* 0xFFFFD7: DMA_CONFIGURATION{{FIXME|reason=Explain the bits}}
 +
 
 +
Additionally, bit 7 in the interrupt register at 0xFFFFC5 is set if a DMA End-Of-List has been encountered.
 +
 
 +
=== Command blocks ===
  
−
* 0x0 = 8 bit
+
DSP command blocks are loaded from X-Memory.
−
* 0x1 = 16 bit
 
−
* 0x2 = 24 bit in MSB
 
−
* 0x3 = 32 bit
 
−
* 0x4
 
−
* 0x5
 
−
* 0x6 = 24 bit in LSB (also endianess switched?{{citation needed}})
 
−
* 0x7
 
  
−
Buffers:
+
{| class="wikitable"
 +
! Word
 +
! Meaning
 +
! Notes
 +
|-
 +
! 0
 +
| Next command block address || Memory address of next command block.
 +
Bit 14 is used as End-Of-List marker.
 +
|-
 +
! 1
 +
| Transfer control word || Controls the DMA transfer:
  
−
* 0x0 = FIFO0
+
'''DSP address interleave (Bit-offset 0; 1-bit):'''{{FIXME|reason=How is addressing affected with sample format 0x3 = 32 bit words?}}
−
* 0x1 = FIFO1
+
<ul>
−
* 0x2 = FIFO2{{citation needed}}
+
<li>0 = Addressing using:<pre>
−
* 0x3 = FIFO3{{citation needed}}
+
for (sample_index = 0; sample_index < sample_count; sample_index++) {
−
* 0x4
+
  transfer_dsp_word(dsp_address + sample_index * dsp_step_size)
−
* 0x5
+
}</pre></li>
−
* 0x6
+
<li>1 = Addressing using:<pre>
−
* 0x7
+
for (block_index = 0; block_index < block_count; block_index++) {
−
* 0x8
+
  for (channel_index = 0; channel_index < channel_count; channel_index++) {
−
* 0x9
+
    transfer_dsp_word(dsp_address + block_index + channel_index * dsp_step_size)
−
* 0xA
+
  }
−
* 0xB
+
}</pre></li>
−
* 0xC
+
</ul>
−
* 0xD
+
 
−
* 0xE = Circular
+
'''Direction (Bit-offset 1; 1-bit):'''
−
* 0xF
+
* 0 = Buffer to DSP (In)
 +
* 1 = DSP to buffer (Out)
 +
 
 +
'''Unknown (Bit-offset 2; 2-bits):'''
 +
* 0 = ?
 +
* 1 = ?
 +
* 2 = ?
 +
* 3 = ?
 +
 
 +
'''Buffer offset writeback (Bit-offset 4; 1-bit):'''
 +
''Only used for scratch buffers, behaves like 0 otherwise.''{{FIXME|reason=Assumption; Didn't do anything for FIFO0. Not sure if I documented this for 0xE or 0xF}}
 +
* 0 = Don't update buffer offset (Word 4)
 +
* 1 = Update buffer offset (Word 4); this respects address wrapping of circular buffers
 +
 
 +
'''Buffer (Bit-offset 5; 4-bits):'''
 +
* 0x0 = FIFO0 (In / Out)
 +
* 0x1 = FIFO1 (In / Out)
 +
* 0x2 = FIFO2 (Out only; Buffer to DSP is ignored)
 +
* 0x3 = FIFO3 (Out only; Buffer to DSP is ignored)
 +
* 0x4 = ?
 +
* 0x5 = ?
 +
* 0x6 = ?
 +
* 0x7 = ?
 +
* 0x8 = ?
 +
* 0x9 = ?
 +
* 0xA = ?
 +
* 0xB = ?
 +
* 0xC = ?
 +
* 0xD = ?
 +
* 0xE = Scratch (Circular)
 +
* 0xF = Scratch
 +
 
 +
'''Unknown (Bit-offset 9; 1-bit):'''
 +
* 0 = ?
 +
* 1 = ?
 +
 
 +
'''Sample format (Bit-offset 10; 3-bits):''' ''{{FIXME|reason=Some of these need tests with signed datatypes}}''
 +
* 0x0 = 8 bit (1 DSP word / 1 byte); ''sample-count must be multiple of 4, or transfer is skipped; rounded; byte MSB is flipped''{{FIXME|reason=How are negative input words affected by rounding? 0x123000 etc.}}
 +
** Buffer to DSP: bytes: <code>0x12,0x34,0x56,0x78</code> &rarr; words: <code>0x920000, 0xB40000, 0xD60000, 0xF80000</code>
 +
** Buffer to DSP: bytes: <code>0x92,0xB4,0xD6,0xF8</code> &rarr; words: <code>0x120000, 0x340000, 0x560000, 0x780000</code>
 +
** DSP to buffer: words: <code>0x927FFF, 0xB47FFF, 0xD67FFF, 0xF87FFF</code> &rarr; bytes: <code>0x12,0x34,0x56,0x78</code> ''(Rounded down)''
 +
** DSP to buffer: words: <code>0x928000, 0xB48000, 0xD68000, 0xF88000</code> &rarr; bytes: <code>0x13,0x35,0x57,0x79</code> ''(Rounded up)''
 +
** DSP to buffer: words: <code>0x800000, 0x7E7FFF, 0x7E8000, 0x7FFFFF</code> &rarr; bytes: <code>0x00,0xFE,0xFF,0xFF</code> ''(Saturated)''
 +
* 0x1 = 16 bit (1 DSP word / 2 bytes) ''sample-count must be multiple of 2, or transfer is skipped; truncated''
 +
** Buffer to DSP: bytes: <code>0x34,0x12</code> &rarr; word: <code>0x123400</code>
 +
** DSP to buffer: word: <code>0x1234FF</code> &rarr; bytes: <code>0x34,0x12</code> ''(Truncated)''
 +
** DSP to buffer: word: <code>0x123400</code> &rarr; bytes: <code>0x34,0x12</code> ''(Truncated)''
 +
* 0x2 = 24 bit in MSB (1 DSP word / 4 bytes); ''padded''
 +
** Buffer to DSP: bytes: <code>0x00,0x56,0x34,0x12</code> &rarr; word: <code>0x123456</code> ''(Padding ignored)''
 +
** Buffer to DSP: bytes: <code>0xFF,0x56,0x34,0x12</code> &rarr; word: <code>0x123456</code> ''(Padding ignored)''
 +
** DSP to buffer: word: <code>0x123456</code> &rarr; bytes: <code>0x00,0x56,0x34,0x12</code> ''(Zero padding)''
 +
* 0x3 = 32 bit (2 DSP words / 4 bytes); ''trunacted''
 +
** Buffer to DSP: bytes: <code>0x12,0xBC,0x9A,0x78</code> &rarr; words: <code>0x120000, 0x789ABC</code>
 +
** DSP to buffer: words: <code>0x12FFFF, 0x789ABC</code> &rarr; bytes: <code>0x12,0xBC,0x9A,0x78</code> ''(Truncated)''
 +
** DSP to buffer: words: <code>0x120000, 0x789ABC</code> &rarr; bytes: <code>0x12,0xBC,0x9A,0x78</code> ''(Truncated)''
 +
* 0x4 = ''Transfer skipped''
 +
* 0x5 = ''Transfer skipped''
 +
* 0x6 = 24 bit in LSB (1 DSP word / 4 bytes); ''padded''
 +
** Buffer to DSP: bytes: <code>0x56,0x34,0x12,0x00</code> &rarr; word: <code>0x123456</code> ''(Padding ignored)''
 +
** Buffer to DSP: bytes: <code>0x56,0x34,0x12,0xFF</code> &rarr; word: <code>0x123456</code> ''(Padding ignored)''
 +
** DSP to buffer: word: <code>0x123456</code> &rarr; bytes: <code>0x56,0x34,0x12,0x00</code> ''(Zero padding)''
 +
* 0x7 = ''Transfer skipped''
 +
 
 +
'''Unknown (Bit-offset 13; 1-bit):'''
 +
* 0 = ?
 +
* 1 = ?
 +
 
 +
'''DSP address step size (Bit-offset 14; 10-bits):'''
 +
 
 +
Used in DSP address calculation
 +
 
 +
|-
 +
! 2
 +
| Sample count || Behaviour depends on interleave setting
 +
 
 +
* Non-interleaved mode:
 +
:
 +
: '''Sample count (Bit-offset 0; 24-bits):'''
 +
: The number of samples to be transferred
 +
:
 +
:
 +
* Interleaved mode:
 +
:
 +
: '''Channel count (Bit-offset 0; 4-bits):'''
 +
: The number of channels minus 1. For 3 channels, this has to be 0x2.
 +
:
 +
: '''Block count (Bit-offset 4; 20-bits):'''
 +
: The number of blocks to be transferred.
 +
:
 +
:
 +
|-
 +
! 3
 +
| DSP address || This is the address in the DSP:
 +
* 0x0000 - 0x17FF = X-Memory
 +
* 0x1800 - 0x27FF = Y-Memory
 +
* 0x2800 - 0x37FF = P-Memory
 +
|-
 +
! 4
 +
| Buffer offset || ''Only used for scratch buffers, ignored otherwise.''{{FIXME|reason=Assumption; effect in buffers 0xE / 0xF; didn't seem to do anything for buffer 0x0}}
 +
This is the offset within the buffer where the first sample is accessed.
 +
 
 +
If this is above or equal to the buffer end (buffer base + buffer size), then the write will behave like a non circular write.
 +
|-
 +
! 5
 +
| Buffer base || ''Only used for circular scratch buffer, ignored otherwise.''{{FIXME|reason=Assumption; it happens for buffer 0xE, but not for buffer 0xF}}
 +
The start address of the buffer.
 +
|-
 +
! 6
 +
| Buffer size || ''Only used for circular scratch buffer, ignored otherwise.''{{FIXME|reason=Assumption; it happens for buffer 0xE, but not for buffer 0xF}}
 +
Size of buffer minus 1. For a buffer with 0x1000 bytes, this has to be 0xFFF.
 +
|}
 +
 
 +
=== FIFO ===
 +
 
 +
{{FIXME|reason=Document register names? The FIFO registers are already defined in the XQEMU source}}
 +
 
 +
Each of the DSP has 4 output FIFOs, and 2 input FIFOs.
 +
 
 +
All FIFOs belonging to the same DSP share the memory view. The memory view is controlled in the DSPs FIFO Scatter-Gather entries.
 +
 
 +
Each FIFO has 3 address registers, which define a ringbuffer:
 +
 
 +
* Base address: The address of the first byte in the FIFO.
 +
* End address: The address of the first byte that's no longer part of the FIFO (so this address is exclusive).
 +
* Current address: The absolute address of the next byte that will be written.
 +
 
 +
During a transfer, the current address is incremented.
 +
If the current address hits the end address during a transfer, the current address is moved to the base, where the transfer continues.
 +
 
 +
If the transfer length is greater than the FIFO length when using the output FIFO, the transfer will continue normally and overwrite previously transferred bytes.
 +
If the transfer length is greater than the FIFO length when using the input FIFO, the transfer will continue normally and read previously transferred bytes again.
 +
 
 +
If the current address is below the base address when starting a transfer, it is moved to the base address.
 +
If the current address is above or equal to the end address when starting a transfer, the DSP hangs{{FIXME|reason=Needed to reboot Xbox to start new DMA, even DSP reset did not work}}.
  
 
== Related links ==
 
== Related links ==
  
−
* [http://www.zdomain.com/a56.html A56, open-source assembler for the similar 56000 architecture]
+
* [https://github.com/XboxDev/a56 Modernized fork of a56, open-source assembler for the similar 56000 architecture]
 +
* [https://github.com/JayFoxRox/xbox-tools/pull/70 Script to analyze DSP instructions]
 +
* [https://github.com/JayFoxRox/xbox-tools/pull/60 Script to control DSP DMA]
 +
* [https://web.archive.org/web/20010212122052/http://www.parthus.com:80/platforms/parthus_mediastream/index.html MediaStream product page]
 +
* [https://web.archive.org/web/20011130084353/http://www.fs2.com:80/parthus_download/ First Silicon MediaStream tools]
 +
* [https://web.archive.org/web/20020213235936/http://www.tasking.com/products/MediaStream/index.html Tasking MediaStream compiler]
 +
 
 +
[[Category:APU]]

Latest revision as of 17:17, 3 February 2019

The DSPs in the APU are probably "Parthus MediaStream" DSP cores (probably the 24-bit DSP2420 "Mozart" core).

Those are similar to Motorola DSP56362 (DSP56300 Family). If so, the datasheet can be found at http://www.nxp.com/docs/en/data-sheet/DSP56362.pdf (Also see "Documentation" section in said datasheet for the related documentation, like the family manual)

Memory Size

Program RAM Size X Data RAM Size Y Data RAM Size MIXBUF Size
GP 4096 x 24-bit 4096 x 24-bit 2048 x 24-bit 992[FIXME] x 24-bit
EP 4096 x 24-bit 3072 x 24-bit 256 x 24-bit n/a

MIXBUF is accessible at X:$001400 in the GP.

Other datasheets for similar DSPs suggest that the memory sizes might be different if instruction cache or switch mode are toggled. It is currently unknown if the DSPs in the Xbox APU support a similar feature[FIXME].

DMA

This section is very incomplete and not much was tested on hardware either

DMA is controlled using peripheral registers:

  • 0xFFFFD4: Memory address of next command block
  • 0xFFFFD5: DMA_START_BLOCK[FIXME]
  • 0xFFFFD6: DMA_CONTROL[FIXME]
  • 0xFFFFD7: DMA_CONFIGURATION[FIXME]

Additionally, bit 7 in the interrupt register at 0xFFFFC5 is set if a DMA End-Of-List has been encountered.

Command blocks

DSP command blocks are loaded from X-Memory.

Word Meaning Notes
0 Next command block address Memory address of next command block.

Bit 14 is used as End-Of-List marker.

1 Transfer control word Controls the DMA transfer:

DSP address interleave (Bit-offset 0; 1-bit):[FIXME]

  • 0 = Addressing using:
    for (sample_index = 0; sample_index < sample_count; sample_index++) {
  transfer_dsp_word(dsp_address + sample_index * dsp_step_size)
}
  • 1 = Addressing using:
    for (block_index = 0; block_index < block_count; block_index++) {
  for (channel_index = 0; channel_index < channel_count; channel_index++) {
    transfer_dsp_word(dsp_address + block_index + channel_index * dsp_step_size)
  }
}

Direction (Bit-offset 1; 1-bit):

  • 0 = Buffer to DSP (In)
  • 1 = DSP to buffer (Out)

Unknown (Bit-offset 2; 2-bits):

  • 0 = ?
  • 1 = ?
  • 2 = ?
  • 3 = ?

Buffer offset writeback (Bit-offset 4; 1-bit): Only used for scratch buffers, behaves like 0 otherwise.[FIXME]

  • 0 = Don't update buffer offset (Word 4)
  • 1 = Update buffer offset (Word 4); this respects address wrapping of circular buffers

Buffer (Bit-offset 5; 4-bits):

  • 0x0 = FIFO0 (In / Out)
  • 0x1 = FIFO1 (In / Out)
  • 0x2 = FIFO2 (Out only; Buffer to DSP is ignored)
  • 0x3 = FIFO3 (Out only; Buffer to DSP is ignored)
  • 0x4 = ?
  • 0x5 = ?
  • 0x6 = ?
  • 0x7 = ?
  • 0x8 = ?
  • 0x9 = ?
  • 0xA = ?
  • 0xB = ?
  • 0xC = ?
  • 0xD = ?
  • 0xE = Scratch (Circular)
  • 0xF = Scratch

Unknown (Bit-offset 9; 1-bit):

  • 0 = ?
  • 1 = ?

Sample format (Bit-offset 10; 3-bits): [FIXME]

  • 0x0 = 8 bit (1 DSP word / 1 byte); sample-count must be multiple of 4, or transfer is skipped; rounded; byte MSB is flipped[FIXME]
    • Buffer to DSP: bytes: 0x12,0x34,0x56,0x78 → words: 0x920000, 0xB40000, 0xD60000, 0xF80000
    • Buffer to DSP: bytes: 0x92,0xB4,0xD6,0xF8 → words: 0x120000, 0x340000, 0x560000, 0x780000
    • DSP to buffer: words: 0x927FFF, 0xB47FFF, 0xD67FFF, 0xF87FFF → bytes: 0x12,0x34,0x56,0x78 (Rounded down)
    • DSP to buffer: words: 0x928000, 0xB48000, 0xD68000, 0xF88000 → bytes: 0x13,0x35,0x57,0x79 (Rounded up)
    • DSP to buffer: words: 0x800000, 0x7E7FFF, 0x7E8000, 0x7FFFFF → bytes: 0x00,0xFE,0xFF,0xFF (Saturated)
  • 0x1 = 16 bit (1 DSP word / 2 bytes) sample-count must be multiple of 2, or transfer is skipped; truncated
    • Buffer to DSP: bytes: 0x34,0x12 → word: 0x123400
    • DSP to buffer: word: 0x1234FF → bytes: 0x34,0x12 (Truncated)
    • DSP to buffer: word: 0x123400 → bytes: 0x34,0x12 (Truncated)
  • 0x2 = 24 bit in MSB (1 DSP word / 4 bytes); padded
    • Buffer to DSP: bytes: 0x00,0x56,0x34,0x12 → word: 0x123456 (Padding ignored)
    • Buffer to DSP: bytes: 0xFF,0x56,0x34,0x12 → word: 0x123456 (Padding ignored)
    • DSP to buffer: word: 0x123456 → bytes: 0x00,0x56,0x34,0x12 (Zero padding)
  • 0x3 = 32 bit (2 DSP words / 4 bytes); trunacted
    • Buffer to DSP: bytes: 0x12,0xBC,0x9A,0x78 → words: 0x120000, 0x789ABC
    • DSP to buffer: words: 0x12FFFF, 0x789ABC → bytes: 0x12,0xBC,0x9A,0x78 (Truncated)
    • DSP to buffer: words: 0x120000, 0x789ABC → bytes: 0x12,0xBC,0x9A,0x78 (Truncated)
  • 0x4 = Transfer skipped
  • 0x5 = Transfer skipped
  • 0x6 = 24 bit in LSB (1 DSP word / 4 bytes); padded
    • Buffer to DSP: bytes: 0x56,0x34,0x12,0x00 → word: 0x123456 (Padding ignored)
    • Buffer to DSP: bytes: 0x56,0x34,0x12,0xFF → word: 0x123456 (Padding ignored)
    • DSP to buffer: word: 0x123456 → bytes: 0x56,0x34,0x12,0x00 (Zero padding)
  • 0x7 = Transfer skipped

Unknown (Bit-offset 13; 1-bit):

  • 0 = ?
  • 1 = ?

DSP address step size (Bit-offset 14; 10-bits):

Used in DSP address calculation

2 Sample count Behaviour depends on interleave setting
  • Non-interleaved mode:
Sample count (Bit-offset 0; 24-bits):
The number of samples to be transferred
  • Interleaved mode:
Channel count (Bit-offset 0; 4-bits):
The number of channels minus 1. For 3 channels, this has to be 0x2.
Block count (Bit-offset 4; 20-bits):
The number of blocks to be transferred.
3 DSP address This is the address in the DSP:
  • 0x0000 - 0x17FF = X-Memory
  • 0x1800 - 0x27FF = Y-Memory
  • 0x2800 - 0x37FF = P-Memory
4 Buffer offset Only used for scratch buffers, ignored otherwise.[FIXME]

This is the offset within the buffer where the first sample is accessed.

If this is above or equal to the buffer end (buffer base + buffer size), then the write will behave like a non circular write.

5 Buffer base Only used for circular scratch buffer, ignored otherwise.[FIXME]

The start address of the buffer.

6 Buffer size Only used for circular scratch buffer, ignored otherwise.[FIXME]

Size of buffer minus 1. For a buffer with 0x1000 bytes, this has to be 0xFFF.

FIFO

[FIXME]

Each of the DSP has 4 output FIFOs, and 2 input FIFOs.

All FIFOs belonging to the same DSP share the memory view. The memory view is controlled in the DSPs FIFO Scatter-Gather entries.

Each FIFO has 3 address registers, which define a ringbuffer:

  • Base address: The address of the first byte in the FIFO.
  • End address: The address of the first byte that's no longer part of the FIFO (so this address is exclusive).
  • Current address: The absolute address of the next byte that will be written.

During a transfer, the current address is incremented. If the current address hits the end address during a transfer, the current address is moved to the base, where the transfer continues.

If the transfer length is greater than the FIFO length when using the output FIFO, the transfer will continue normally and overwrite previously transferred bytes. If the transfer length is greater than the FIFO length when using the input FIFO, the transfer will continue normally and read previously transferred bytes again.

If the current address is below the base address when starting a transfer, it is moved to the base address. If the current address is above or equal to the end address when starting a transfer, the DSP hangs[FIXME].

Related links

Retrieved from "https://xboxdevwiki.net/index.php?title=DSP&oldid=6690"