3dbrew

GSP Shared Memory: Difference between revisions

← Older edit
Kynex7510 (talk | contribs)
88 edits
PSC + small edits
Kynex7510 (talk | contribs)
88 edits
m Use conventional names
 
(5 intermediate revisions by the same user not shown)
Line 34: Line 34:
GSP fills the interrupt list, then triggers the event set with [[GSPGPU:RegisterInterruptRelayQueue|RegisterInterruptRelayQueue]] for the specified process(es).
GSP fills the interrupt list, then triggers the event set with [[GSPGPU:RegisterInterruptRelayQueue|RegisterInterruptRelayQueue]] for the specified process(es).


PDC interrupts are sent to all processes; other interrupts are only sent to the process with GPU rights.
PDC interrupts are sent to all processes; other interrupts are only sent to the process with rendering rights.
 
When issuing a [[GSP_Shared_Memory#Trigger_Memory_Fill|Memory Fill]] command with both buffers set GSP will only dispatch PSC0.


= Framebuffer Info =
= Framebuffer Info =
Line 96: Line 98:
See [[Configuration Memory]].
See [[Configuration Memory]].


= Command Queue =
= GX Command Queue =


The command queue is located at sharedMemBase + 0x800 + (clientID * 0x200). It consists of an header followed by at most 15 command entries. Each command entry is of size 0x20 and has an header followed by command specific parameters.
This command queue is located at sharedMemBase + 0x800 + (clientID * 0x200). It consists of an header followed by at most 15 command entries.
 
After adding a command, [[GSPGPU:TriggerCmdReqQueue|TriggerCmdReqQueue]] must be used to start command processing (official code does so when the total commands field is 1).


== Command Queue Header ==
The queue header has the following structure:


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 116: Line 116:
|-
|-
| 2
| 2
| Status (bit0 = halted, bit7 = fatal error)
| Status (0x1 = halted, 0x80 = error)
|-
|-
| 3
| 3
Line 125: Line 125:
|}
|}


GSP checks for status.bit0 and optionally avoids handling further commands, however the check is done by equality, which means it will always fail if status.bit7 is also set (and thus other commands will be processed). This bug prevents the halting logic from working propertly, but can be worked around by keeping bit0 of word3 set, as that will force halting on each iteration.
After adding a command, [[GSPGPU:TriggerCmdReqQueue|TriggerCmdReqQueue]] must be used to start command processing (official code does so when the total commands field is 1).
 
== Commands ==
 
A command entry is made of 8 words. The first word is the command header, subsequent words represent command specific parameters.


== Command Header ==
The command header has the following structure:


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 147: Line 151:
|}
|}


== Commands ==
Addresses specified in command parameters are virtual addresses. Depending on the command, there might be constraints on the accepted parameters. In general, some commands require parameters to be aligned, and addresses are expected to be on [[Memory_Management#Memory_Mapping|linear]], [[Memory_layout#0x1F000000_.28New_3DS_only.29|QTM]] or VRAM memory.
 
Addresses specified in parameters are virtual addresses. Address and size parameters except for command 0 and command 5 must be 8-byte aligned.


=== Trigger DMA Request ===
=== RequestDMA ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 177: Line 179:
|}
|}


This command issues a [[Corelink_DMA_Engines|DMA request]] as the calling process. When the destination address is within VRAM, GSP places itself as the destination process, this makes it possible to transfer data in VRAM without needing it listed in the destination process [[NCCH/Extended_Header#ARM11_Kernel_Capabilities|exheader mappings]]. Otherwise, both source and destination of the DMA request are the calling process.
This command issues a [[Corelink_DMA_Engines|DMA request]] as the process with [[GSPGPU:AcquireRight|rendering rights]]. When the destination address is within VRAM, GSP places itself as the destination process: this makes it possible to transfer data in VRAM without needing it listed in the destination process [[NCCH/Extended_Header#ARM11_Kernel_Capabilities|exheader mappings]]. Otherwise, both source and destination of the DMA request are the process with rendering rights.
 
The source buffer must be mapped as readable in the source process, while the destination address must be mapped as writable in the destination process, otherwise GSP calls [[SVC|svcBreak]]. When flushing is enabled and the source address is above VRAM, svcFlushProcessDataCache is used to flush the source buffer.


The source buffer must be mapped as readable in the source process, while the destination address must be mapped as writable in the destination process. When flushing is enabled and the source address is above VRAM, [[SVC|svcFlushProcessDataCache]] is used to flush the source buffer.
Any process must have acquired rendering rights, otherwise the command does nothing.


=== Trigger Command List Processing ===
=== ProcessCommandList ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 209: Line 213:
This command sets the [[GPU/External_Registers#Command_List|Command List registers]], and optionally updates gas additive blend results after command processing has ended.
This command sets the [[GPU/External_Registers#Command_List|Command List registers]], and optionally updates gas additive blend results after command processing has ended.


No error checking is performed on the parameters. Address and size should be both aligned to 8 bytes, and the address should be in linear, QTM or VRAM memory, otherwise PA 0 is used. When flushing is enabled, svcFlushProcessDataCache is used to flush the buffer.
No error checking is performed on the parameters. Address and size should be both aligned to 8 bytes, and the address should be in linear, QTM or VRAM memory, otherwise PA 0 is used. When flushing is enabled, [[SVC|svcFlushProcessDataCache]] is used to flush the buffer on the process that has acquired rendering rights.


=== Trigger Memory Fill ===
Any process must have acquired rendering rights, otherwise the command does nothing.
 
=== MemoryFill ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 243: Line 249:
|}
|}


This command sets the GPU [[GPU/External_Registers#Memory_Fill|Memory Fill registers]].
This command sets the [[GPU/External_Registers#Memory_Fill|Memory Fill registers]].


Addresses should be aligned to 8 bytes and must be in linear, QTM or VRAM memory, otherwise error 0xE0E02BF5 (GSP_INVALID_ADDRESS) is returned. The start address for a buffer must be < its end address, else the same error is returned. If the start address for a buffer is 0, that buffer is skipped; otherwise, its relative PSC unit is used for the fill operation.
Addresses should be aligned to 8 bytes and must be in linear, QTM or VRAM memory, otherwise error 0xE0E02BF5 (GSP_INVALID_ADDRESS) is returned. The start address for a buffer must be below its end address, else the same error is returned. If the start address for a buffer is 0, that buffer is skipped; otherwise, its relative PSC unit is used for the fill operation.


=== Trigger Display Transfer ===
=== DisplayTransfer ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 258: Line 264:
|-
|-
| 1
| 1
| Input framebuffer address
| Source address
|-
|-
| 2
| 2
| Output framebuffer address
| Destination address
|-
|-
| 3
| 3
| Input framebuffer [[GPU|dimensions]]
| Source dimensions
|-
|-
| 4
| 4
| Output framebuffer dimensions
| Output dimensions
|-
|-
| 5
| 5
| [[GPU|Flags]], for applications this is 0x1001000 for the main screen, and 0x1000 for the sub screen.
| Flags
|-
|-
| 7-6
| 7-6
Line 276: Line 282:
|}
|}


This command converts the specified addresses to physical addresses, then writes these physical addresses and parameters to the [[GPU]] registers at 0x1EF00C00. This GPU command copies the already rendered framebuffer data from the input GPU framebuffer address to the specified output LCD framebuffer. The input framebuffer is normally located in VRAM.
This command sets the [[GPU/External_Registers#Transfer_Engine|Display Transfer registers]].
 
The GPU color buffer is stored in the same Z-curve (tiled) format as textures. By default, SetDisplayTransfer converts the given buffer from the tiled format to a linear format adapted to the LCD framebuffers.
 
Display transfers are performed asynchronously, so after requesting a display transfer you should wait for the PPF interrupt to fire before reading the output data.


The minimum supported dimension for output is 64x64, anything lower will hang the engine.
No error checking is performed on the parameters. Addresses should be aligned to 8 bytes and should be in linear, QTM or VRAM memory, otherwise PA 0 is used.


=== Trigger Texture Copy ===
=== TextureCopy ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 295: Line 297:
|-
|-
| 1
| 1
| Input buffer address.
| Source address
|-
|-
| 2
| 2
| Output buffer address.
| Destination address
|-
|-
| 3
| 3
| Total bytes to copy, not including gaps.
| Size
|-
|-
| 4
| 4
| Bits 0-15: Size of input line, in bytes. Bits 16-31: Gap between input lines, in bytes.
| Line width <nowiki>|</nowiki> (gap << 16)
|-
|-
| 5
| 5
| Same as 4, but for the output.
| Same as above, for the destination
|-
|-
| 6
| 6
| Flags, corresponding to the [[GPU/External_Registers#Transfer_Engine|Transfer Engine flags]]. However, for TextureCopy commands, bit 3 is always set, bit 2 is set if any output dimension is smaller than the input, and other bits are always 0.
| Flags
|-
|-
| 7
| 7
Line 316: Line 318:
|}
|}


This command is similar to cmd3. It also triggers the [[GPU/External_Registers#Transfer_Engine|GPU Transfer Engine]], but setting the TextureCopy parameters.
This command sets the [[GPU/External_Registers#TextureCopy|Texture Copy registers]]. Note that GSP doesn't enforce bit3 of the flags to be set.


=== Flush Cache Regions ===
No error checking is performed on the parameters. Addresses and size should be aligned to 8 bytes, and the addresses should be in linear, QTM or VRAM memory, otherwise PA 0 is used.
 
=== FlushCacheRegions ===


{| class="wikitable" border="1"
{| class="wikitable" border="1"
Line 329: Line 333:
|-
|-
| 1
| 1
| Buf0 address
| Buffer 0 address
|-
|-
| 2
| 2
| Buf0 size
| Buffer 0 size
|-
|-
| 3
| 3
| Buf1 address
| Buffer 1 address
|-
|-
| 4
| 4
| Buf1 size
| Buffer 1 size
|-
|-
| 5
| 5
| Buf2 address
| Buffer 2 address
|-
|-
| 6
| 6
| Buf2 size
| Buffer 2 size
|-
|-
| 7
| 7
Line 350: Line 354:
|}
|}


The application buffer addresses specified in the parameters are used with [[SVC|svcFlushProcessDataCache]]. The input buf0 size must not be zero. When buf1 size is zero, svcFlushProcessDataCache() for buf1 and buf2 are skipped. When buf2 size is zero, svcFlushProcessDataCache() for buf2 is skipped.
This command calls svcFlushProcessDataCache for each buffer on the process that has acquired rendering rights.
 
If any call fails, its error is returned; If any buffer has size 0, the buffer is skipped. In both cases, subsequent buffers are not processed.
 
Any process must have acquired rendering rights, otherwise the error 0xD8202A06 (GSP_NO_RIGHT) is returned.
 
== Bugs ==
 
* When issuing a DMA request, GSP attempts to acquire an internal semaphore that rules CDMA access; this semaphore is never released on failure paths. While this is generally not an issue, as GSP breaks on DMA failures, it becomes a problem if the DMA request is done with cache flushing: in that case, GSP will error silently, causing a deadlock in DMA code.
* When handling GX commands apart from RequestDMA and ProcessCommandList, GSP sets the relative busy flags in internal state before executing the commands. This means that, if the relevant interrupts are never triggered (eg. on invalid parameters), the busy flags never get reset, preventing execution of future commands of the same kind.
Retrieved from "https://www.3dbrew.org/wiki/GSP_Shared_Memory"