Asahi¶
The Asahi driver aims to provide an OpenGL implementation for the Apple M1.
Wrap (macOS only)¶
Mesa includes a library that wraps the key IOKit entrypoints used in the macOS UABI for AGX. The wrapped routines print information about the kernel calls made and dump work submitted to the GPU using agxdecode. This facilitates reverse-engineering the hardware, as glue to get at the “interesting” GPU memory.
The library is only built if -Dtools=asahi is passed. It builds a single
wrap.dylib file, which should be inserted into a process with the
DYLD_INSERT_LIBRARIES environment variable.
For example, to trace an app ./app, run:
DYLD_INSERT_LIBRARIES=~/mesa/build/src/asahi/lib/libwrap.dylib ./app
Hardware varyings¶
At an API level, vertex shader outputs need to be interpolated to become fragment shader inputs. This process is logically pipelined in AGX, with a value traveling from a vertex shader to remapping hardware to coefficient register setup to the fragment shader to the iterator hardware. Each stage is described below.
Vertex shader¶
A vertex shader (running on the Unified Shader Cores) outputs varyings with the
st_var instruction. st_var takes a vertex output index and a 32-bit
value. The maximum number of vertex outputs is specified as the “output count”
of the shader in the “VDM State Vertex Outputs” structure. The value may be interpreted
consist of a single 32-bit value or an aligned 16-bit register pair, depending
on whether interpolation should happen at 32-bit or 16-bit. Vertex outputs are
indexed starting from 0, with the vertex position always coming first, the
32-bit user varyings coming next with perspective, flat, and linear interpolated
varyings grouped in that order, then 16-bit user varyings with the same groupings,
and finally point size, layer/viewport, and clip distances at the end if present. Note that
clip distances are not accessible from the fragment shader; if the fragment
shader needs to read the interpolated clip distance, the vertex shader must
also write the clip distance values to a user varying for the fragment shader
to interpolate. Also note there is no clip plane enable mask anywhere; that must
lowered for APIs that require this (OpenGL but not Vulkan).
Size (words) |
Value |
|---|---|
4 |
Vertex position |
1 |
32-bit smooth varying 0 |
… |
|
1 |
32-bit smooth varying m |
1 |
32-bit flat varying 0 |
… |
|
1 |
32-bit flat varying n |
1 |
32-bit linear varying 0 |
… |
|
1 |
32-bit linear varying o |
1 |
Packed pair of 16-bit smooth varyings 0 |
… |
|
1 |
Packed pair of 16-bit smooth varyings p |
1 |
Packed pair of 16-bit flat varyings 0 |
… |
|
1 |
Packed pair of 16-bit flat varyings q |
1 |
Packed pair of 16-bit linear varyings 0 |
… |
|
1 |
Packed pair of 16-bit linear varyings r |
1 |
Point size |
1 |
Layer/viewport |
1 |
Clip distance for plane 0 |
… |
|
1 |
Clip distance for plane 16 |
Remapping¶
Vertex outputs are remapped to varying slots to be interpolated. The output of remapping consists of the following items: the W fragment coordinate, the Z fragment coordinate, user varyings in the vertex output order. Z may be omitted, but W may not be. This remapping is configured by the “Output select” word.
Index |
Value |
|---|---|
0 |
Fragment coord W |
1 |
Fragment coord Z |
2 |
32-bit varying 0 |
… |
|
2 + m |
32-bit varying m |
2 + m + 1 |
Packed pair of 16-bit varyings 0 |
… |
|
2 + m + n + 1 |
Packed pair of 16-bit varyings n |
Coefficient registers¶
The fragment shader does not see the physical slots. Instead, it references varyings through coefficient registers. A coefficient register is a register allocated constant for all fragment shader invocations in a given polygon. Physically, it contains the values output by the vertex shader for each vertex of the polygon. Coefficient registers are preloaded with values from varying slots. This preloading appears to occur in fixed function hardware, a simplification from PowerVR which requires a specialized program for the programmable data sequencer to do the preload.
The “Fragment Shader” structure points to coefficient register bindings, preceded by a header. The header contains the number of 32-bit varying slots. As the W slot is always present, this field is always nonzero. Slots whose index is below this count are treated as 32-bit. The remaining slots are treated as 16-bits.
The header also contains the total number of coefficient registers bound.
Each binding that follows maps a (vector of) varying slots to a (consecutive) coefficient registers. Some details about the varying (perspective interpolation, flat shading, point sprites) are configured here.
Coefficient registers may be ordered the same as the internal varying slots. However, this may be inconvenient for some APIs that require a separable shader model. For these APIs, the flexibility to mix-and-match slots and coefficient registers allows mixing shaders without shader variants. In that case, the bindings should be generated outside of the compiler. For simple APIs where the bindings are fixed and known at compile-time, the bindings could be generated within the compiler.
Mathematically, the value of the coefficient register is a vector in \(\mathbb{R}^3\). The X and Y components are the screen-space partial derivatives of the varying with respect to X and Y. The Z component is the interpolated value of the varying at the upper-left pixel in the 32x32 tile that the pixel belongs to.
Fragment shader¶
In the fragment shader, coefficient registers, identified by the prefix cf
followed by a decimal index, act as opaque handles to varyings. For flat
shading, coefficient registers may be loaded into general registers with the
ldcf instruction. For smooth shading, the coefficient register corresponding
to the desired varying is passed as an argument to the “iterate” instruction
iter in order to “iterate” (interpolate) a varying. As perspective correct
interpolation also requires the W component of the fragment coordinate, the
coefficient register for W is passed as a second argument. As an example, if
there’s a single varying to interpolate, an instruction like iter r0, cf1, cf0
is used.
It is occassionally useful to manipulate the raw coefficient registers, for
example to implement interpolation modes not natively supported by the hardware.
ldcf is used for this purpose.
Iterator¶
To actually interpolate varyings, AGX provides fixed-function iteration hardware to multiply the specified coefficient registers with the required barycentrics, producing an interpolated value, hence the name “coefficient register”. This operation is purely mathematical and does not require any memory access, as the required coefficients are preloaded before the shader begins execution. That means the iterate instruction executes in constant time, does not signal a data fence, and does not require the shader to wait on a data fence before using the value.
Image layouts¶
AGX supports several image layouts, described here. To work with image layouts
in the drivers, use the ail library, located in src/asahi/layout.
Strided linear¶
The simplest layout is strided linear. Pixels are stored in raster-order in memory with a software-controlled stride. Strided linear images are useful for working with modifier-unaware window systems, however performance will suffer. Strided linear images have numerous limitations:
Strides must be a multiple of 16 bytes.
Strides must be nonzero. For 1D images where the stride is logically irrelevant, ail will internally select the minimal stride.
Only 1D, 2D, and 2D Array images may be linear. In particular, no 3D or cubemaps.
2D images must not be mipmapped.
Block-compressed formats and multisampled images are unsupported. Elements of a strided linear image are simply pixels.
With these limitations, addressing into a strided linear image is as simple as
In practice, this suffices for window system integration and little else.
GPU-tiled¶
The most common uncompressed layout is GPU-tiled. The image is divided into power-of-two sized tiles. The tiles themselves are stored in raster-order. Within each tile, elements (pixels/blocks) are stored in Morton (Z) order.
The tile size used depends on both the image size and the block size of the image format. For large images, \(n \times n\) or \(2n \times n\) tiles are used (\(n\) power-of-two). \(n\) is such that each page contains exactly one tile. Only power-of-two block sizes are supported in hardware, ensuring such a tile size always exists. The hardware uses 16 KiB pages, so tile sizes are as follows:
Bytes per block |
Tile size |
|---|---|
1 |
128 x 128 |
2 |
128 x 64 |
4 |
64 x 64 |
8 |
64 x 32 |
16 |
32 x 32 |
The dimensions of large images are rounded up to be multiples of the tile size. In addition, non-power-of-two large images have extra padding tiles when mipmapping is used, see below.
That rounding would waste a great deal of memory for small images. If an image is smaller than this tile size, a smaller tile size is used to reduce the memory footprint. For small images, the tile size is \(m \times m\) where
In other words, small images use the smallest square power-of-two tile such that the image’s minor axis fits in one tile.
For mipmapped images, tile sizes are determined independently for each level. Typically, the first levels of an image are “large” and the remaining levels are “small”. This scheme reduces the memory footprint of mipmapping, compared to a fixed tile size for the whole image. Each mip level are padded to fill at least one cache line (128 bytes), ensure no cache line contains multiple mip levels.
There is a wrinkle: the dimensions of large mip levels in tiles are determined by the dimensions of level 0. For power-of-two images, the two calculations are equivalent. However, they differ subtly for non-power-of-two images. To determine the number of tiles to allocate for level \(l\), the number of tiles for level 0 should be right-shifted by \(2l\). That appears to divide by \(2^l\) in both width and height, matching the definition of mipmapping, however it rounds down incorrectly. To compensate, the level contains one extra row, column, or both (with the corner) as required if any of the first \(l\) levels were rounded down. This hurt the memory footprint. However, it means non-power-of-two integer multiplication is only required for level 0. Calculating the sizes for subsequent levels requires only addition and bitwise math. That simplifies the hardware (but complicates software).
A 2D image consists of a full miptree (constructed as above) rounded up to the page size (16 KiB).
3D images consist simply of an array of 2D layers (constructed as above). That
means cube maps, 2D arrays, cube map arrays, and 3D images all use the same
layout. The only difference is the number of layers. Notably, 3D images (like
GL_TEXTURE_3D) reserve space even for mip levels that do not exist
logically. These extra levels pad out layers of 3D images to the size of the
first layer, simplifying layout calculations for both software and hardware.
Although the padding is logically unnecessary, it wastes little space compared
to the sizes of large mipmapped 3D textures.
Twiddled¶
In addition to GPU-tiled images, AGX also has a fully twiddled layout. The image is rounded up to power-of-two dimensions, then all elements are stored in Morton (Z) order.
A twiddled image is equivalent to a GPU-tiled image with a single square power-of-two tile spanning the entire image. That means GPU-tiled and twiddled images may share address calculation code, as long as everything is parametrized in terms of the tile size.
In general, twiddled images require more memory than GPU-tiled images due to the extra padding required. Because GPU tiles are page-sized, twiddling beyond that does not offer any cache locality benefit either. The twiddled layout is mostly vestigial at this point, but the PBE requires it for sparse mapping.
Sparse page tables¶
The hardware has native support for sparse images. If an texture/PBE descriptor is configured in sparse mode, the specified address does not point to the image itself. Rather, it points to a sparse page table. The extra indirection enables on-device sparse binding (e.g. by updating the page table from a compute kernel).
At the top level, the sparse page table is an array of folios. Each folio describes 256 pages. The folio itself has two halves. The first half is the page table itself, containing 4-byte “Sparse Block” data structures mapping image pages to GPU virtual addresses. The second half is probably sparse texture counters, again 4-bytes per page. Each folio therefore consumes \(256 \cdot 4 \cdot 2 = 2 \mathrm{KiB}\) in order to describe \(256 \cdot 16384 = 4 \mathrm{MiB}\).
In a layered (array or 3D) image, a given folio only describes a single layer. That implies extra padding between layers.
Within a layer, the table works purely at an address level. It maps pages to pages, rather than tiles to tiles. It is not a spatial data structure in itself. Rather, it inherits the tiling of the image by virtue of the addresses mapped. This presumably simplifies the hardware implementation.
drm-shim (Linux only)¶
Mesa includes a library that mocks out the DRM UABI used by the Asahi driver stack, allowing the Mesa driver to run on non-M1 Linux hardware. This can be useful for exercising the compiler. To build, use options:
-Dgallium-drivers=asahi -Dtools=drm-shim
Then run an OpenGL workload with environment variable:
LD_PRELOAD=~/mesa/build/src/asahi/drm-shim/libasahi_noop_drm_shim.so
For example to compile a shader with shaderdb and print some statistics along with the IR:
~/shader-db$ AGX_MESA_DEBUG=shaders,shaderdb ASAHI_MESA_DEBUG=precompile LD_PRELOAD=~/mesa/build/src/asahi/drm-shim/libasahi_noop_drm_shim.so ./run shaders/glmark/1-12.shader_test
The drm-shim implementation for Asahi is located in src/asahi/drm-shim. The
drm-shim implementation there should be updated as new UABI is added.
Hardware glossary¶
AGX is a tiled renderer descended from the PowerVR architecture. Some hardware concepts used in PowerVR GPUs appear in AGX.
- CDM¶
- Compute Data Master¶
Dispatches compute kernels.
- ISP¶
- Image Synthesis Processor¶
The Image Synthesis Processor is responsible for the rasterization stage of the rendering pipeline.
- PBE¶
- Pixel BackEnd¶
Hardware unit which writes to color attachments and images. Also the name for a descriptor passed to PBE instructions.
- PDM¶
- Pixel Data Master¶
Dispatches pixel shaders.
- PPP¶
- Primitive Processing Pipeline¶
The Primitive Processing Pipeline is a hardware unit that does primitive assembly. The PPP is between the VDM and ISP.
- USC¶
- Unified Shader Cores¶
A unified shader core is a small CPU that runs shader code. The core is unified because a single ISA is used for vertex, pixel and compute shaders. This differs from older GPUs where the vertex, fragment and compute have separate ISAs for shader stages.
- UVS¶
- Unified Vertex Store¶
Hardware unit which buffers the outputs of the vertex shader (varyings).
- VDM¶
- Vertex Data Master¶
Dispatches vertex shaders.