freedreno/ir3: Add descriptor set lowering

Add support to lower IBO (image/SSBO) and fb-read to use bindless
descriptors.  This will be used by a6xx to avoid having to merge image
and SSBO state into a single compact IBO descriptor, and also simplify
enabling image and SSBO support for additional shader stages (since each
stage can use it's own descriptor set).

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20687>
This commit is contained in:
Rob Clark
2022-12-31 11:35:15 -08:00
committed by Marge Bot
parent 066ff5c759
commit a45d111c18
3 changed files with 203 additions and 0 deletions
@@ -0,0 +1,131 @@
/*
* Copyright © 2022 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.
*/
#include "nir.h"
#include "nir_builder.h"
#include "ir3/ir3_descriptor.h"
static bool
lower_intrinsic(nir_builder *b, nir_intrinsic_instr *intr)
{
unsigned desc_offset;
switch (intr->intrinsic) {
case nir_intrinsic_load_ssbo:
case nir_intrinsic_store_ssbo:
case nir_intrinsic_ssbo_atomic_add:
case nir_intrinsic_ssbo_atomic_imin:
case nir_intrinsic_ssbo_atomic_umin:
case nir_intrinsic_ssbo_atomic_imax:
case nir_intrinsic_ssbo_atomic_umax:
case nir_intrinsic_ssbo_atomic_and:
case nir_intrinsic_ssbo_atomic_or:
case nir_intrinsic_ssbo_atomic_xor:
case nir_intrinsic_ssbo_atomic_exchange:
case nir_intrinsic_ssbo_atomic_comp_swap:
case nir_intrinsic_ssbo_atomic_fadd:
case nir_intrinsic_ssbo_atomic_fmin:
case nir_intrinsic_ssbo_atomic_fmax:
case nir_intrinsic_ssbo_atomic_fcomp_swap:
case nir_intrinsic_get_ssbo_size:
desc_offset = IR3_BINDLESS_SSBO_OFFSET;
break;
case nir_intrinsic_image_load:
case nir_intrinsic_image_store:
case nir_intrinsic_image_atomic_add:
case nir_intrinsic_image_atomic_imin:
case nir_intrinsic_image_atomic_umin:
case nir_intrinsic_image_atomic_imax:
case nir_intrinsic_image_atomic_umax:
case nir_intrinsic_image_atomic_and:
case nir_intrinsic_image_atomic_or:
case nir_intrinsic_image_atomic_xor:
case nir_intrinsic_image_atomic_exchange:
case nir_intrinsic_image_atomic_comp_swap:
case nir_intrinsic_image_size:
case nir_intrinsic_image_samples:
desc_offset = IR3_BINDLESS_IMAGE_OFFSET;
break;
default:
return false;
}
unsigned buffer_src;
if (intr->intrinsic == nir_intrinsic_store_ssbo) {
/* store_ssbo has the value first, and ssbo src as 2nd src: */
buffer_src = 1;
} else {
/* the rest have ssbo src as 1st src: */
buffer_src = 0;
}
unsigned set = ir3_shader_descriptor_set(b->shader->info.stage);
nir_ssa_def *src = nir_ssa_for_src(b, intr->src[buffer_src], 1);
src = nir_iadd(b, src, nir_imm_int(b, desc_offset));
/* An out-of-bounds index into an SSBO/image array can cause a GPU fault
* on access to the descriptor (I don't see any hw mechanism to bound the
* access). We could just allow the resulting iova fault (it is a read
* fault, so shouldn't corrupt anything), but at the cost of one extra
* instruction (as long as IR3_BINDLESS_DESC_COUNT is a power-of-two) we
* can avoid the dmesg spam and users thinking this is a driver bug:
*/
src = nir_umod(b, src, nir_imm_int(b, IR3_BINDLESS_DESC_COUNT));
nir_ssa_def *bindless = nir_bindless_resource_ir3(b, 32, src, set);
nir_instr_rewrite_src_ssa(&intr->instr, &intr->src[buffer_src], bindless);
return true;
}
static bool
lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
{
b->cursor = nir_before_instr(instr);
switch (instr->type) {
case nir_instr_type_intrinsic:
return lower_intrinsic(b, nir_instr_as_intrinsic(instr));
default:
return false;
}
}
/**
* Lower bindful image/SSBO to bindless
*/
bool
ir3_nir_lower_io_to_bindless(nir_shader *shader)
{
/* Note: We don't currently support API level bindless, as we assume we
* can remap bindful images/SSBOs to bindless while controlling the entire
* descriptor set space.
*
* If we needed to support API level bindless, we could probably just remap
* bindful ops to a range of the descriptor set space that does not conflict
* with what we advertise for bindless descriptors? But I'm not sure that
* ARB_bindless_texture is of too much value to care about, especially for
* GLES
*/
assert(!shader->info.uses_bindless);
return nir_shader_instructions_pass(shader, lower_instr, nir_metadata_none, NULL);
}
@@ -0,0 +1,70 @@
/*
* Copyright © 2022 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.
*/
#ifndef IR3_DESCRIPTOR_H_
#define IR3_DESCRIPTOR_H_
#include "ir3/ir3_shader.h"
/*
* When using bindless descriptor sets for image/SSBO (and fb-read) state,
* since the descriptor sets are large, layout the descriptor set with the
* first IR3_BINDLESS_SSBO_COUNT slots for SSBOs followed by
* IR3_BINDLESS_IMAGE_COUNT slots for images. (For fragment shaders, the
* last image slot is reserved for fb-read tex descriptor.)
*
* Note that these limits are more or less arbitrary. But the enable_mask
* in fd_shaderbuf_stateobj / fd_shaderimg_stateobj would need to be more
* than uint32_t to support more than 32.
*/
#define IR3_BINDLESS_SSBO_OFFSET 0
#define IR3_BINDLESS_SSBO_COUNT 32
#define IR3_BINDLESS_IMAGE_OFFSET IR3_BINDLESS_SSBO_COUNT
#define IR3_BINDLESS_IMAGE_COUNT 32
#define IR3_BINDLESS_DESC_COUNT (IR3_BINDLESS_IMAGE_OFFSET + IR3_BINDLESS_IMAGE_COUNT)
/**
* When using bindless descriptor sets for IBO/etc, each shader stage gets
* it's own descriptor set, avoiding the need to merge image/ssbo state
* across shader stages.
*/
static inline unsigned
ir3_shader_descriptor_set(enum pipe_shader_type shader)
{
switch (shader) {
case PIPE_SHADER_VERTEX: return 0;
case PIPE_SHADER_TESS_CTRL: return 1;
case PIPE_SHADER_TESS_EVAL: return 2;
case PIPE_SHADER_GEOMETRY: return 3;
case PIPE_SHADER_FRAGMENT: return 4;
case PIPE_SHADER_COMPUTE: return 0;
default:
unreachable("bad shader stage");
return ~0;
}
}
bool ir3_nir_lower_io_to_bindless(nir_shader *shader);
#endif /* IR3_DESCRIPTOR_H_ */
@@ -217,6 +217,8 @@ files_libfreedreno = files(
'ir3/ir3_cache.c',
'ir3/ir3_cache.h',
'ir3/ir3_const.h',
'ir3/ir3_descriptor.c',
'ir3/ir3_descriptor.h',
'ir3/ir3_gallium.c',
'ir3/ir3_gallium.h',
)