griefith/test
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix #129571: Metal: Broken texture atomic workaround

Browse Source 
The refactor 9c0321ae9b
had the wrong mental model of the backing texture
layout for the atomic workaround.

For 3D textures, the layout is breaking the 3D texture
and reinterpreting the linear location as its 2D
linear location. This breaks the 3D texture Z slices
into non contiguous regions in 2D.

Comments have been added to avoid future confusion.

Pull Request: https://projects.blender.org/blender/blender/pulls/133830
This commit is contained in:
Clément Foucault
2025-01-31 16:10:59 +01:00
committed by Clément Foucault
parent 7397fda92b
commit 067f6767d4
1 changed files with 47 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
source/blender/gpu/shaders/metal/mtl_shader_defines.msl
View File

@@ -177,11 +177,13 @@ struct _mtl_sampler {
* in the texture. */
ushort aligned_width;
int2 to_internal_coord(IntCoord coord) const
/* Convert 2D coordinates to the backing 2D texture coordinates. */
int2 to_internal_coord(thread TextureT * /*texture*/, IntCoord coord) const
{
return coord.xy;
}
uint to_linear_coord(IntCoord coord) const
/* Convert 2D coordinates to the memory location in the 2D texture memory buffer. */
uint to_linear_coord(thread TextureT * /*texture*/, IntCoord coord) const
{
return coord.x + coord.y * aligned_width;
}
@@ -190,7 +192,20 @@ struct _mtl_sampler {
template<typename U> struct AtomicEmulation<U, 3, true> {
/** Buffer to do atomic operations on. */
device U *buffer;
/** Required to pixel location inside the backing texture 2D space, and texture size query. */
/**
* Required for pixel location inside the backing texture 2D space, and texture size query.
*
* IMPORTANT: Note that the slices are not contiguous in the backing texture as we simply use
* it for linear storage.
* For instance, for a [2,2,2] texture store in a [6x4] backing texture:
*
* 001122
* 334455
* 667788
* xxxxxx
*
* The numbers are rows in the 3D texture.
*/
ushort3 texture_size;
/* Aligned width matches the number of buffer elements in bytes_per_row. This may be greater
* than the texture's native width to satisfy device alignment rules. We need to use the padded
@@ -198,14 +213,26 @@ struct _mtl_sampler {
* in the texture. */
ushort aligned_width;
int2 to_internal_coord(IntCoord coord) const
/* Convert 3D coordinates to the backing 2D texture coordinates. */
int2 to_internal_coord(thread TextureT *texture, IntCoord coord) const
{
return int2(coord.x, coord.y + texture_size.y * coord.z);
/* Index of a pixel in the data array. Assuming data layout is scanline. */
uint linear_pixel = uint(coord.x) + uint(coord.y) * texture_size.x +
uint(coord.z) * (texture_size.x * texture_size.y);
uint backing_tex_width = texture->get_width();
/* Coordinate inside the backing texture. */
uint y = linear_pixel / backing_tex_width;
/* Avoid use of modulo operator for speed. */
uint x = linear_pixel - y * backing_tex_width;
return int2(x, y);
}
uint to_linear_coord(IntCoord coord) const
/* Convert 3D coordinates to the memory location in the 2D texture memory buffer. */
uint to_linear_coord(thread TextureT *texture, IntCoord coord) const
{
uint row = coord.y + coord.z * texture_size.y;
return coord.x + row * aligned_width;
uint2 co = uint2(to_internal_coord(texture, coord));
return co.x + co.y * uint(aligned_width);
}
};
@@ -350,29 +377,29 @@ struct _mtl_sampler {
#undef NON_ARRAY_FN
#ifndef MTL_SUPPORTS_TEXTURE_ATOMICS
/* Atomic samplers only support `textureFetch` as the texture layout doesn't allow filtering. */
/* Atomic samplers only support `texelFetch` as the texture layout doesn't allow filtering. */
# define ATOMIC_FN template<bool At = Atomic, ENABLE_IF(At == true)>
ATOMIC_FN DataVec fetch(IntCoord coord) const
{
int2 coord_2d = atomic.to_internal_coord(coord);
int2 coord_2d = atomic.to_internal_coord(texture, coord);
return texture->sample(_mtl_fetch_samp, float2(coord_2d));
}
ATOMIC_FN DataVec fetch(IntCoord coord, level lod, IntDeriv offset = {0}) const
{
int2 coord_2d = atomic.to_internal_coord(coord);
int2 coord_2d = atomic.to_internal_coord(texture, coord);
return texture->sample(_mtl_fetch_samp, float2(coord_2d), lod, offset);
}
ATOMIC_FN DataVec load(IntCoord coord) const
{
int2 coord_2d = atomic.to_internal_coord(coord);
int2 coord_2d = atomic.to_internal_coord(texture, coord);
return texture->read(uint2(coord_2d), 0);
}
ATOMIC_FN void store(DataVec data, IntCoord coord) const
{
int2 coord_2d = atomic.to_internal_coord(coord);
int2 coord_2d = atomic.to_internal_coord(texture, coord);
texture->write(data, uint2(coord_2d), 0);
}
@@ -380,31 +407,31 @@ struct _mtl_sampler {
AtomicT atomic_min(IntCoord coord, AtomicT data) const
{
return atomicMin(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicMin(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_max(IntCoord coord, AtomicT data) const
{
return atomicMax(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicMax(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_add(IntCoord coord, AtomicT data) const
{
return atomicAdd(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicAdd(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_and(IntCoord coord, AtomicT data) const
{
return atomicAnd(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicAnd(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_or(IntCoord coord, AtomicT data) const
{
return atomicOr(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicOr(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_xor(IntCoord coord, AtomicT data) const
{
return atomicXor(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicXor(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
AtomicT atomic_exchange(IntCoord coord, AtomicT data) const
{
return atomicExchange(atomic.buffer[atomic.to_linear_coord(coord)], data);
return atomicExchange(atomic.buffer[atomic.to_linear_coord(texture, coord)], data);
}
#else