freedreno: Move a6xx's setup_slices() to a shareable helper function.

We pass in all the parameters for setting up the layout, though freedreno
still sets a few of them up early (since it uses layout helpers in making
some decisions about the layout setup parameters that will be cleaned up
once krh's blitter work lands).
This commit is contained in:
Eric Anholt 2019-11-26 12:02:34 -08:00
parent 67258a44d2
commit d9cf3e76bd
10 changed files with 301 additions and 148 deletions

View File

@ -15,6 +15,8 @@ drm_SOURCES := \
drm/msm_ringbuffer.c
ir3_SOURCES := \
fdl/fd6_layout.c \
fdl/freedreno_layout.c \
ir3/disasm-a3xx.c \
ir3/instr-a3xx.h \
ir3/ir3.c \

View File

@ -0,0 +1,185 @@
/*
* Copyright (C) 2018 Rob Clark <robclark@freedesktop.org>
* Copyright © 2018-2019 Google, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
* Authors:
* Rob Clark <robclark@freedesktop.org>
*/
#include <stdio.h>
#include "freedreno_layout.h"
/* indexed by cpp, including msaa 2x and 4x: */
static const struct {
unsigned pitchalign;
unsigned heightalign;
} tile_alignment[] = {
[1] = { 128, 32 },
[2] = { 64, 32 },
[3] = { 64, 32 },
[4] = { 64, 16 },
[6] = { 64, 16 },
[8] = { 64, 16 },
[12] = { 64, 16 },
[16] = { 64, 16 },
[24] = { 64, 16 },
[32] = { 64, 16 },
[48] = { 64, 16 },
[64] = { 64, 16 },
/* special cases for r16: */
[0] = { 128, 16 },
};
/* NOTE: good way to test this is: (for example)
* piglit/bin/texelFetch fs sampler3D 100x100x8
*/
void
fdl6_layout(struct fdl_layout *layout,
enum pipe_format format, uint32_t nr_samples,
uint32_t width0, uint32_t height0, uint32_t depth0,
uint32_t mip_levels, uint32_t array_size, bool is_3d)
{
assert(nr_samples > 0);
layout->width0 = width0;
layout->height0 = height0;
layout->depth0 = depth0;
layout->cpp = util_format_get_blocksize(format);
layout->cpp *= nr_samples;
const struct util_format_description *format_desc =
util_format_description(format);
uint32_t level;
uint32_t depth = depth0;
/* linear dimensions: */
uint32_t lwidth = width0;
uint32_t lheight = height0;
/* tile_mode dimensions: */
uint32_t twidth = util_next_power_of_two(lwidth);
uint32_t theight = util_next_power_of_two(lheight);
int ta = layout->cpp;
/* The z16/r16 formats seem to not play by the normal tiling rules: */
if ((layout->cpp == 2) && (util_format_get_nr_components(format) == 1))
ta = 0;
uint32_t alignment;
if (is_3d) {
layout->layer_first = false;
alignment = 4096;
} else {
layout->layer_first = true;
alignment = 1;
}
/* in layer_first layout, the level (slice) contains just one
* layer (since in fact the layer contains the slices)
*/
uint32_t layers_in_level = layout->layer_first ? 1 : array_size;
debug_assert(ta < ARRAY_SIZE(tile_alignment));
debug_assert(tile_alignment[ta].pitchalign);
for (level = 0; level < mip_levels; level++) {
struct fdl_slice *slice = &layout->slices[level];
uint32_t tile_mode = fdl_tile_mode(layout, level);
uint32_t width, height;
/* tiled levels of 3D textures are rounded up to PoT dimensions: */
if (is_3d && tile_mode) {
width = twidth;
height = theight;
} else {
width = lwidth;
height = lheight;
}
uint32_t aligned_height = height;
uint32_t pitchalign;
if (tile_mode) {
pitchalign = tile_alignment[ta].pitchalign;
aligned_height = align(aligned_height,
tile_alignment[ta].heightalign);
} else {
pitchalign = 64;
}
/* The blits used for mem<->gmem work at a granularity of
* 32x32, which can cause faults due to over-fetch on the
* last level. The simple solution is to over-allocate a
* bit the last level to ensure any over-fetch is harmless.
* The pitch is already sufficiently aligned, but height
* may not be:
*/
if (level == mip_levels - 1)
aligned_height = align(aligned_height, 32);
if (format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC)
slice->pitch =
util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
else
slice->pitch = align(width, pitchalign);
slice->offset = layout->size;
uint32_t blocks = util_format_get_nblocks(format,
slice->pitch, aligned_height);
/* 1d array and 2d array textures must all have the same layer size
* for each miplevel on a6xx. 3d textures can have different layer
* sizes for high levels, but the hw auto-sizer is buggy (or at least
* different than what this code does), so as soon as the layer size
* range gets into range, we stop reducing it.
*/
if (is_3d) {
if (level < 1 || layout->slices[level - 1].size0 > 0xf000) {
slice->size0 = align(blocks * layout->cpp, alignment);
} else {
slice->size0 = layout->slices[level - 1].size0;
}
} else {
slice->size0 = align(blocks * layout->cpp, alignment);
}
layout->size += slice->size0 * depth * layers_in_level;
if (false) {
fprintf(stderr, "%s: %ux%ux%u@%u:\t%2u: stride=%4u, size=%6u,%7u, aligned_height=%3u, blocks=%u, offset=0x%x tiling=%d\n",
util_format_name(format),
width, height, depth, layout->cpp,
level, slice->pitch * layout->cpp,
slice->size0, layout->size, aligned_height, blocks,
slice->offset, tile_mode);
}
depth = u_minify(depth, 1);
lwidth = u_minify(lwidth, 1);
lheight = u_minify(lheight, 1);
twidth = u_minify(twidth, 1);
theight = u_minify(theight, 1);
}
if (layout->layer_first) {
layout->layer_size = align(layout->size, 4096);
layout->size = layout->layer_size * array_size;
}
}

View File

@ -0,0 +1,40 @@
/*
* Copyright (C) 2018 Rob Clark <robclark@freedesktop.org>
* Copyright © 2018 Google, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
* Authors:
* Rob Clark <robclark@freedesktop.org>
*/
#include <stdio.h>
#include "freedreno_layout.h"
void
fdl_layout_buffer(struct fdl_layout *layout, uint32_t size)
{
layout->width0 = size;
layout->height0 = 1;
layout->depth0 = 1;
layout->cpp = 1;
layout->size = size;
}

View File

@ -24,6 +24,13 @@
#ifndef FREEDRENO_LAYOUT_H_
#define FREEDRENO_LAYOUT_H_
#include <stdbool.h>
#include <stdint.h>
#include "util/u_debug.h"
#include "util/u_math.h"
#include "util/format/u_format.h"
/* Shared freedreno mipmap layout helper
*
* It does *not* attempt to track surface transitions, in particular
@ -102,6 +109,8 @@ struct fdl_layout {
uint32_t width0, height0, depth0;
uint32_t size; /* Size of the whole image, in bytes. */
uint32_t ubwc_size;
};
@ -128,8 +137,8 @@ fdl_ubwc_offset(const struct fdl_layout *layout, unsigned level, unsigned layer)
* for multi layer/level images, it will.
*/
if (layout->ubwc_size) {
debug_assert(level == 0);
debug_assert(layer == 0);
assert(level == 0);
assert(layer == 0);
}
return layout->ubwc_slices[0].offset;
}
@ -158,4 +167,13 @@ fdl_ubwc_enabled(const struct fdl_layout *layout, int level)
return layout->ubwc_size && fdl_tile_mode(layout, level);
}
void
fdl_layout_buffer(struct fdl_layout *layout, uint32_t size);
void
fdl6_layout(struct fdl_layout *layout,
enum pipe_format format, uint32_t nr_samples,
uint32_t width0, uint32_t height0, uint32_t depth0,
uint32_t mip_levels, uint32_t array_size, bool is_3d);
#endif /* FREEDRENO_LAYOUT_H_ */

View File

@ -0,0 +1,33 @@
# Copyright © 2018 Rob Clark
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
libfreedreno_layout = static_library(
'freedreno_layout',
[
'fd6_layout.c',
'freedreno_layout.c',
],
include_directories : [inc_freedreno, inc_common],
c_args : [c_vis_args, no_override_init_args],
cpp_args : [cpp_vis_args],
dependencies : idep_nir_headers,
build_by_default : false,
)

View File

@ -21,6 +21,7 @@
inc_freedreno = include_directories(['.', './registers'])
subdir('drm')
subdir('fdl')
subdir('ir3')
subdir('registers')
subdir('perfcntrs')

View File

@ -110,6 +110,7 @@ libvulkan_freedreno = shared_library(
libvulkan_wsi,
libfreedreno_drm, # required by ir3_shader_get_variant, which we don't use
libfreedreno_ir3,
libfreedreno_layout,
],
dependencies : [
dep_dl,

View File

@ -30,139 +30,6 @@
#include "a6xx.xml.h"
/* indexed by cpp, including msaa 2x and 4x: */
static const struct {
unsigned pitchalign;
unsigned heightalign;
} tile_alignment[] = {
[1] = { 128, 32 },
[2] = { 64, 32 },
[3] = { 64, 32 },
[4] = { 64, 16 },
[6] = { 64, 16 },
[8] = { 64, 16 },
[12] = { 64, 16 },
[16] = { 64, 16 },
[24] = { 64, 16 },
[32] = { 64, 16 },
[48] = { 64, 16 },
[64] = { 64, 16 },
/* special cases for r16: */
[0] = { 128, 16 },
};
/* NOTE: good way to test this is: (for example)
* piglit/bin/texelFetch fs sampler3D 100x100x8
*/
static uint32_t
setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format format)
{
struct pipe_resource *prsc = &rsc->base;
struct fd_screen *screen = fd_screen(prsc->screen);
enum util_format_layout layout = util_format_description(format)->layout;
uint32_t pitchalign = screen->gmem_alignw;
uint32_t level, size = 0;
uint32_t depth = prsc->depth0;
/* linear dimensions: */
uint32_t lwidth = prsc->width0;
uint32_t lheight = prsc->height0;
/* tile_mode dimensions: */
uint32_t twidth = util_next_power_of_two(lwidth);
uint32_t theight = util_next_power_of_two(lheight);
/* in layer_first layout, the level (slice) contains just one
* layer (since in fact the layer contains the slices)
*/
uint32_t layers_in_level = rsc->layout.layer_first ? 1 : prsc->array_size;
int ta = rsc->layout.cpp;
/* The z16/r16 formats seem to not play by the normal tiling rules: */
if ((rsc->layout.cpp == 2) && (util_format_get_nr_components(format) == 1))
ta = 0;
debug_assert(ta < ARRAY_SIZE(tile_alignment));
debug_assert(tile_alignment[ta].pitchalign);
for (level = 0; level <= prsc->last_level; level++) {
struct fdl_slice *slice = fd_resource_slice(rsc, level);
uint32_t tile_mode = fd_resource_tile_mode(prsc, level);
uint32_t width, height;
/* tiled levels of 3D textures are rounded up to PoT dimensions: */
if ((prsc->target == PIPE_TEXTURE_3D) && tile_mode) {
width = twidth;
height = theight;
} else {
width = lwidth;
height = lheight;
}
uint32_t aligned_height = height;
uint32_t blocks;
if (tile_mode) {
pitchalign = tile_alignment[ta].pitchalign;
aligned_height = align(aligned_height,
tile_alignment[ta].heightalign);
} else {
pitchalign = 64;
}
/* The blits used for mem<->gmem work at a granularity of
* 32x32, which can cause faults due to over-fetch on the
* last level. The simple solution is to over-allocate a
* bit the last level to ensure any over-fetch is harmless.
* The pitch is already sufficiently aligned, but height
* may not be:
*/
if ((level == prsc->last_level) && (prsc->target != PIPE_BUFFER))
aligned_height = align(aligned_height, 32);
if (layout == UTIL_FORMAT_LAYOUT_ASTC)
slice->pitch =
util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
else
slice->pitch = align(width, pitchalign);
slice->offset = size;
blocks = util_format_get_nblocks(format, slice->pitch, aligned_height);
/* 1d array and 2d array textures must all have the same layer size
* for each miplevel on a6xx. 3d textures can have different layer
* sizes for high levels, but the hw auto-sizer is buggy (or at least
* different than what this code does), so as soon as the layer size
* range gets into range, we stop reducing it.
*/
if (prsc->target == PIPE_TEXTURE_3D) {
if (level < 1 || fd_resource_slice(rsc, level - 1)->size0 > 0xf000) {
slice->size0 = align(blocks * rsc->layout.cpp, alignment);
} else {
slice->size0 = fd_resource_slice(rsc, level - 1)->size0;
}
} else {
slice->size0 = align(blocks * rsc->layout.cpp, alignment);
}
size += slice->size0 * depth * layers_in_level;
#if 0
fprintf(stderr, "%s: %ux%ux%u@%u:\t%2u: stride=%4u, size=%6u,%7u, aligned_height=%3u, blocks=%u, offset=0x%x tiling=%d\n",
util_format_name(prsc->format),
width, height, depth, rsc->layout.cpp,
level, slice->pitch * rsc->layout.cpp,
slice->size0, size, aligned_height, blocks,
slice->offset, fd_resource_tile_mode(prsc, level));
#endif
depth = u_minify(depth, 1);
lwidth = u_minify(lwidth, 1);
lheight = u_minify(lheight, 1);
twidth = u_minify(twidth, 1);
theight = u_minify(theight, 1);
}
return size;
}
/* A subset of the valid tiled formats can be compressed. We do
* already require tiled in order to be compressed, but just because
* it can be tiled doesn't mean it can be compressed.
@ -325,21 +192,19 @@ setup_lrz(struct fd_resource *rsc)
uint32_t
fd6_setup_slices(struct fd_resource *rsc)
{
uint32_t alignment;
struct pipe_resource *prsc = &rsc->base;
if (!(fd_mesa_debug & FD_DBG_NOLRZ) && has_depth(rsc->base.format))
setup_lrz(rsc);
switch (rsc->base.target) {
case PIPE_TEXTURE_3D:
rsc->layout.layer_first = false;
alignment = 4096;
break;
default:
rsc->layout.layer_first = true;
alignment = 1;
break;
}
fdl6_layout(&rsc->layout, prsc->format, fd_resource_nr_samples(prsc),
prsc->width0, prsc->height0, prsc->depth0,
prsc->last_level + 1, prsc->array_size,
prsc->target == PIPE_TEXTURE_3D);
return setup_slices(rsc, alignment, rsc->base.format);
/* The caller does this bit of layout setup again. */
if (rsc->layout.layer_first)
return rsc->layout.size / prsc->array_size;
else
return rsc->layout.size;
}

View File

@ -991,7 +991,13 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
rsc->internal_format = format;
size = screen->setup_slices(rsc);
if (prsc->target == PIPE_BUFFER) {
assert(prsc->format == PIPE_FORMAT_R8_UNORM);
size = prsc->width0;
fdl_layout_buffer(&rsc->layout, size);
} else {
size = screen->setup_slices(rsc);
}
if (allow_ubwc && screen->fill_ubwc_buffer_sizes && rsc->layout.tile_mode)
size += screen->fill_ubwc_buffer_sizes(rsc);

View File

@ -243,6 +243,7 @@ driver_freedreno = declare_dependency(
libfreedreno,
libfreedreno_drm,
libfreedreno_ir3,
libfreedreno_layout,
libfreedreno_perfcntrs
],
dependencies : idep_nir,
@ -260,6 +261,7 @@ ir3_compiler = executable(
libfreedreno,
libfreedreno_drm,
libfreedreno_ir3,
libfreedreno_layout,
libgallium,
libglsl_standalone,
],