mesa/src/intel/vulkan/anv_nir_apply_pipeline_layo...

957 lines
34 KiB
C
Raw Normal View History

/*
* Copyright © 2015 Intel Corporation
*
* 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 "anv_nir.h"
#include "program/prog_parameter.h"
#include "nir/nir_builder.h"
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
#include "compiler/brw_nir.h"
#include "util/set.h"
/* Sampler tables don't actually have a maximum size but we pick one just so
* that we don't end up emitting too much state on-the-fly.
*/
#define MAX_SAMPLER_TABLE_SIZE 128
#define BINDLESS_OFFSET 255
struct apply_pipeline_layout_state {
const struct anv_physical_device *pdevice;
nir_shader *shader;
nir_builder builder;
struct anv_pipeline_layout *layout;
bool add_bounds_checks;
/* Place to flag lowered instructions so we don't lower them twice */
struct set *lowered_instrs;
bool uses_constants;
uint8_t constants_offset;
struct {
bool desc_buffer_used;
uint8_t desc_offset;
uint8_t *use_count;
uint8_t *surface_offsets;
uint8_t *sampler_offsets;
} set[MAX_SETS];
};
static void
add_binding(struct apply_pipeline_layout_state *state,
uint32_t set, uint32_t binding)
{
const struct anv_descriptor_set_binding_layout *bind_layout =
&state->layout->set[set].layout->binding[binding];
if (state->set[set].use_count[binding] < UINT8_MAX)
state->set[set].use_count[binding]++;
/* Only flag the descriptor buffer as used if there's actually data for
* this binding. This lets us be lazy and call this function constantly
* without worrying about unnecessarily enabling the buffer.
*/
if (anv_descriptor_size(bind_layout))
state->set[set].desc_buffer_used = true;
}
static void
add_deref_src_binding(struct apply_pipeline_layout_state *state, nir_src src)
{
nir_deref_instr *deref = nir_src_as_deref(src);
nir_variable *var = nir_deref_instr_get_variable(deref);
add_binding(state, var->data.descriptor_set, var->data.binding);
}
static void
add_tex_src_binding(struct apply_pipeline_layout_state *state,
nir_tex_instr *tex, nir_tex_src_type deref_src_type)
{
int deref_src_idx = nir_tex_instr_src_index(tex, deref_src_type);
if (deref_src_idx < 0)
return;
add_deref_src_binding(state, tex->src[deref_src_idx].src);
}
static void
get_used_bindings_block(nir_block *block,
struct apply_pipeline_layout_state *state)
{
nir_foreach_instr_safe(instr, block) {
switch (instr->type) {
case nir_instr_type_intrinsic: {
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
case nir_intrinsic_vulkan_resource_index:
add_binding(state, nir_intrinsic_desc_set(intrin),
nir_intrinsic_binding(intrin));
break;
case nir_intrinsic_image_deref_load:
case nir_intrinsic_image_deref_store:
case nir_intrinsic_image_deref_atomic_add:
case nir_intrinsic_image_deref_atomic_min:
case nir_intrinsic_image_deref_atomic_max:
case nir_intrinsic_image_deref_atomic_and:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
case nir_intrinsic_image_deref_atomic_exchange:
case nir_intrinsic_image_deref_atomic_comp_swap:
case nir_intrinsic_image_deref_size:
case nir_intrinsic_image_deref_samples:
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
case nir_intrinsic_image_deref_load_param_intel:
case nir_intrinsic_image_deref_load_raw_intel:
case nir_intrinsic_image_deref_store_raw_intel:
add_deref_src_binding(state, intrin->src[0]);
break;
case nir_intrinsic_load_constant:
state->uses_constants = true;
break;
default:
break;
}
break;
}
case nir_instr_type_tex: {
nir_tex_instr *tex = nir_instr_as_tex(instr);
add_tex_src_binding(state, tex, nir_tex_src_texture_deref);
add_tex_src_binding(state, tex, nir_tex_src_sampler_deref);
break;
}
default:
continue;
}
}
}
static bool
find_descriptor_for_index_src(nir_src src,
struct apply_pipeline_layout_state *state)
{
nir_intrinsic_instr *intrin = nir_src_as_intrinsic(src);
while (intrin && intrin->intrinsic == nir_intrinsic_vulkan_resource_reindex)
intrin = nir_src_as_intrinsic(intrin->src[0]);
if (!intrin || intrin->intrinsic != nir_intrinsic_vulkan_resource_index)
return false;
return true;
}
static bool
nir_deref_find_descriptor(nir_deref_instr *deref,
struct apply_pipeline_layout_state *state)
{
while (1) {
/* Nothing we will use this on has a variable */
assert(deref->deref_type != nir_deref_type_var);
nir_deref_instr *parent = nir_src_as_deref(deref->parent);
if (!parent)
break;
deref = parent;
}
assert(deref->deref_type == nir_deref_type_cast);
nir_intrinsic_instr *intrin = nir_src_as_intrinsic(deref->parent);
if (!intrin || intrin->intrinsic != nir_intrinsic_load_vulkan_descriptor)
return false;
return find_descriptor_for_index_src(intrin->src[0], state);
}
static nir_ssa_def *
build_index_for_res_reindex(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
if (intrin->intrinsic == nir_intrinsic_vulkan_resource_reindex) {
nir_ssa_def *bti =
build_index_for_res_reindex(nir_src_as_intrinsic(intrin->src[0]), state);
b->cursor = nir_before_instr(&intrin->instr);
return nir_iadd(b, bti, nir_ssa_for_src(b, intrin->src[1], 1));
}
assert(intrin->intrinsic == nir_intrinsic_vulkan_resource_index);
uint32_t set = nir_intrinsic_desc_set(intrin);
uint32_t binding = nir_intrinsic_binding(intrin);
const struct anv_descriptor_set_binding_layout *bind_layout =
&state->layout->set[set].layout->binding[binding];
uint32_t surface_index = state->set[set].surface_offsets[binding];
uint32_t array_size = bind_layout->array_size;
b->cursor = nir_before_instr(&intrin->instr);
nir_ssa_def *array_index = nir_ssa_for_src(b, intrin->src[0], 1);
if (nir_src_is_const(intrin->src[0]) || state->add_bounds_checks)
array_index = nir_umin(b, array_index, nir_imm_int(b, array_size - 1));
return nir_iadd_imm(b, array_index, surface_index);
}
static nir_ssa_def *
build_index_offset_for_deref(nir_deref_instr *deref,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
nir_deref_instr *parent = nir_deref_instr_parent(deref);
if (parent) {
nir_ssa_def *addr = build_index_offset_for_deref(parent, state);
b->cursor = nir_before_instr(&deref->instr);
return nir_explicit_io_address_from_deref(b, deref, addr,
nir_address_format_32bit_index_offset);
}
nir_intrinsic_instr *load_desc = nir_src_as_intrinsic(deref->parent);
assert(load_desc->intrinsic == nir_intrinsic_load_vulkan_descriptor);
nir_ssa_def *index =
build_index_for_res_reindex(nir_src_as_intrinsic(load_desc->src[0]), state);
/* Return a 0 offset which will get picked up by the recursion */
b->cursor = nir_before_instr(&deref->instr);
return nir_vec2(b, index, nir_imm_int(b, 0));
}
static bool
try_lower_direct_buffer_intrinsic(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
if (deref->mode != nir_var_mem_ssbo)
return false;
if (!nir_deref_find_descriptor(deref, state))
return false;
nir_ssa_def *addr = build_index_offset_for_deref(deref, state);
b->cursor = nir_before_instr(&intrin->instr);
nir_lower_explicit_io_instr(b, intrin, addr,
nir_address_format_32bit_index_offset);
return true;
}
static void
lower_direct_buffer_access(nir_function_impl *impl,
struct apply_pipeline_layout_state *state)
{
nir_foreach_block(block, impl) {
nir_foreach_instr_safe(instr, block) {
if (instr->type != nir_instr_type_intrinsic)
continue;
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
case nir_intrinsic_load_deref:
case nir_intrinsic_store_deref:
case nir_intrinsic_deref_atomic_add:
case nir_intrinsic_deref_atomic_imin:
case nir_intrinsic_deref_atomic_umin:
case nir_intrinsic_deref_atomic_imax:
case nir_intrinsic_deref_atomic_umax:
case nir_intrinsic_deref_atomic_and:
case nir_intrinsic_deref_atomic_or:
case nir_intrinsic_deref_atomic_xor:
case nir_intrinsic_deref_atomic_exchange:
case nir_intrinsic_deref_atomic_comp_swap:
case nir_intrinsic_deref_atomic_fmin:
case nir_intrinsic_deref_atomic_fmax:
case nir_intrinsic_deref_atomic_fcomp_swap:
try_lower_direct_buffer_intrinsic(intrin, state);
break;
case nir_intrinsic_get_buffer_size: {
/* The get_buffer_size intrinsic always just takes a
* index/reindex intrinsic.
*/
if (!find_descriptor_for_index_src(intrin->src[0], state))
break;
nir_ssa_def *index =
build_index_for_res_reindex(nir_src_as_intrinsic(intrin->src[0]),
state);
nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],
nir_src_for_ssa(index));
_mesa_set_add(state->lowered_instrs, intrin);
}
default:
break;
}
}
}
}
static void
lower_res_index_intrinsic(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
uint32_t set = nir_intrinsic_desc_set(intrin);
uint32_t binding = nir_intrinsic_binding(intrin);
const struct anv_descriptor_set_binding_layout *bind_layout =
&state->layout->set[set].layout->binding[binding];
uint32_t surface_index = state->set[set].surface_offsets[binding];
uint32_t array_size = bind_layout->array_size;
nir_ssa_def *array_index = nir_ssa_for_src(b, intrin->src[0], 1);
if (nir_src_is_const(intrin->src[0]) || state->add_bounds_checks)
array_index = nir_umin(b, array_index, nir_imm_int(b, array_size - 1));
nir_ssa_def *index;
if (bind_layout->data & ANV_DESCRIPTOR_INLINE_UNIFORM) {
/* This is an inline uniform block. Just reference the descriptor set
* and use the descriptor offset as the base.
*/
index = nir_imm_ivec2(b, state->set[set].desc_offset,
bind_layout->descriptor_offset);
} else {
/* We're using nir_address_format_32bit_index_offset */
index = nir_vec2(b, nir_iadd_imm(b, array_index, surface_index),
nir_imm_int(b, 0));
}
assert(intrin->dest.is_ssa);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(index));
nir_instr_remove(&intrin->instr);
}
static void
lower_res_reindex_intrinsic(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
/* For us, the resource indices are just indices into the binding table and
* array elements are sequential. A resource_reindex just turns into an
* add of the two indices.
*/
assert(intrin->src[0].is_ssa && intrin->src[1].is_ssa);
nir_ssa_def *old_index = intrin->src[0].ssa;
nir_ssa_def *offset = intrin->src[1].ssa;
nir_ssa_def *new_index =
nir_vec2(b, nir_iadd(b, nir_channel(b, old_index, 0), offset),
nir_channel(b, old_index, 1));
assert(intrin->dest.is_ssa);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(new_index));
nir_instr_remove(&intrin->instr);
}
static void
lower_load_vulkan_descriptor(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
/* We follow the nir_address_format_32bit_index_offset model */
assert(intrin->src[0].is_ssa);
nir_ssa_def *index = intrin->src[0].ssa;
assert(intrin->dest.is_ssa);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(index));
nir_instr_remove(&intrin->instr);
}
static void
lower_get_buffer_size(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
if (_mesa_set_search(state->lowered_instrs, intrin))
return;
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
assert(intrin->src[0].is_ssa);
nir_ssa_def *index = intrin->src[0].ssa;
/* We're following the nir_address_format_32bit_index_offset model so the
* binding table index is the first component of the address. The
* back-end wants a scalar binding table index source.
*/
nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],
nir_src_for_ssa(nir_channel(b, index, 0)));
}
static nir_ssa_def *
build_descriptor_load(nir_deref_instr *deref, unsigned offset,
unsigned num_components, unsigned bit_size,
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
struct apply_pipeline_layout_state *state)
{
nir_variable *var = nir_deref_instr_get_variable(deref);
unsigned set = var->data.descriptor_set;
unsigned binding = var->data.binding;
unsigned array_size =
state->layout->set[set].layout->binding[binding].array_size;
const struct anv_descriptor_set_binding_layout *bind_layout =
&state->layout->set[set].layout->binding[binding];
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
nir_builder *b = &state->builder;
nir_ssa_def *desc_buffer_index =
nir_imm_int(b, state->set[set].desc_offset);
nir_ssa_def *desc_offset =
nir_imm_int(b, bind_layout->descriptor_offset + offset);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
if (deref->deref_type != nir_deref_type_var) {
assert(deref->deref_type == nir_deref_type_array);
const unsigned descriptor_size = anv_descriptor_size(bind_layout);
nir_ssa_def *arr_index = nir_ssa_for_src(b, deref->arr.index, 1);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
if (state->add_bounds_checks)
arr_index = nir_umin(b, arr_index, nir_imm_int(b, array_size - 1));
desc_offset = nir_iadd(b, desc_offset,
nir_imul_imm(b, arr_index, descriptor_size));
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
}
nir_intrinsic_instr *desc_load =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo);
desc_load->src[0] = nir_src_for_ssa(desc_buffer_index);
desc_load->src[1] = nir_src_for_ssa(desc_offset);
desc_load->num_components = num_components;
nir_ssa_dest_init(&desc_load->instr, &desc_load->dest,
num_components, bit_size, NULL);
nir_builder_instr_insert(b, &desc_load->instr);
return &desc_load->dest.ssa;
}
static void
lower_image_intrinsic(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
if (intrin->intrinsic == nir_intrinsic_image_deref_load_param_intel) {
b->cursor = nir_instr_remove(&intrin->instr);
const unsigned param = nir_intrinsic_base(intrin);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
nir_ssa_def *desc =
build_descriptor_load(deref, param * 16,
intrin->dest.ssa.num_components,
intrin->dest.ssa.bit_size, state);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(desc));
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
} else {
nir_variable *var = nir_deref_instr_get_variable(deref);
unsigned set = var->data.descriptor_set;
unsigned binding = var->data.binding;
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
unsigned binding_offset = state->set[set].surface_offsets[binding];
unsigned array_size =
state->layout->set[set].layout->binding[binding].array_size;
nir_ssa_def *index = NULL;
if (deref->deref_type != nir_deref_type_var) {
assert(deref->deref_type == nir_deref_type_array);
index = nir_ssa_for_src(b, deref->arr.index, 1);
if (state->add_bounds_checks)
index = nir_umin(b, index, nir_imm_int(b, array_size - 1));
} else {
index = nir_imm_int(b, 0);
}
index = nir_iadd_imm(b, index, binding_offset);
nir_rewrite_image_intrinsic(intrin, index, false);
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
}
}
static void
lower_load_constant(nir_intrinsic_instr *intrin,
struct apply_pipeline_layout_state *state)
{
nir_builder *b = &state->builder;
b->cursor = nir_before_instr(&intrin->instr);
nir_ssa_def *index = nir_imm_int(b, state->constants_offset);
nir_ssa_def *offset = nir_iadd(b, nir_ssa_for_src(b, intrin->src[0], 1),
nir_imm_int(b, nir_intrinsic_base(intrin)));
nir_intrinsic_instr *load_ubo =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo);
load_ubo->num_components = intrin->num_components;
load_ubo->src[0] = nir_src_for_ssa(index);
load_ubo->src[1] = nir_src_for_ssa(offset);
nir_ssa_dest_init(&load_ubo->instr, &load_ubo->dest,
intrin->dest.ssa.num_components,
intrin->dest.ssa.bit_size, NULL);
nir_builder_instr_insert(b, &load_ubo->instr);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
nir_src_for_ssa(&load_ubo->dest.ssa));
nir_instr_remove(&intrin->instr);
}
static void
lower_tex_deref(nir_tex_instr *tex, nir_tex_src_type deref_src_type,
unsigned *base_index,
struct apply_pipeline_layout_state *state)
{
int deref_src_idx = nir_tex_instr_src_index(tex, deref_src_type);
if (deref_src_idx < 0)
return;
nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
nir_variable *var = nir_deref_instr_get_variable(deref);
unsigned set = var->data.descriptor_set;
unsigned binding = var->data.binding;
unsigned array_size =
state->layout->set[set].layout->binding[binding].array_size;
nir_tex_src_type offset_src_type;
if (deref_src_type == nir_tex_src_texture_deref) {
offset_src_type = nir_tex_src_texture_offset;
*base_index = state->set[set].surface_offsets[binding];
} else {
assert(deref_src_type == nir_tex_src_sampler_deref);
offset_src_type = nir_tex_src_sampler_offset;
*base_index = state->set[set].sampler_offsets[binding];
}
nir_ssa_def *index = NULL;
if (deref->deref_type != nir_deref_type_var) {
assert(deref->deref_type == nir_deref_type_array);
if (nir_src_is_const(deref->arr.index)) {
unsigned arr_index = nir_src_as_uint(deref->arr.index);
*base_index += MIN2(arr_index, array_size - 1);
} else {
nir_builder *b = &state->builder;
/* From VK_KHR_sampler_ycbcr_conversion:
*
* If sampler YCBCR conversion is enabled, the combined image
* sampler must be indexed only by constant integral expressions when
* aggregated into arrays in shader code, irrespective of the
* shaderSampledImageArrayDynamicIndexing feature.
*/
assert(nir_tex_instr_src_index(tex, nir_tex_src_plane) == -1);
index = nir_ssa_for_src(b, deref->arr.index, 1);
if (state->add_bounds_checks)
index = nir_umin(b, index, nir_imm_int(b, array_size - 1));
}
}
if (index) {
nir_instr_rewrite_src(&tex->instr, &tex->src[deref_src_idx].src,
nir_src_for_ssa(index));
tex->src[deref_src_idx].src_type = offset_src_type;
} else {
nir_tex_instr_remove_src(tex, deref_src_idx);
}
}
static uint32_t
tex_instr_get_and_remove_plane_src(nir_tex_instr *tex)
{
int plane_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_plane);
if (plane_src_idx < 0)
return 0;
unsigned plane = nir_src_as_uint(tex->src[plane_src_idx].src);
nir_tex_instr_remove_src(tex, plane_src_idx);
return plane;
}
static void
lower_tex(nir_tex_instr *tex, struct apply_pipeline_layout_state *state)
{
state->builder.cursor = nir_before_instr(&tex->instr);
unsigned plane = tex_instr_get_and_remove_plane_src(tex);
lower_tex_deref(tex, nir_tex_src_texture_deref,
&tex->texture_index, state);
tex->texture_index += plane;
lower_tex_deref(tex, nir_tex_src_sampler_deref,
&tex->sampler_index, state);
tex->sampler_index += plane;
/* The backend only ever uses this to mark used surfaces. We don't care
* about that little optimization so it just needs to be non-zero.
*/
tex->texture_array_size = 1;
}
static void
apply_pipeline_layout_block(nir_block *block,
struct apply_pipeline_layout_state *state)
{
nir_foreach_instr_safe(instr, block) {
switch (instr->type) {
case nir_instr_type_intrinsic: {
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
case nir_intrinsic_vulkan_resource_index:
lower_res_index_intrinsic(intrin, state);
break;
case nir_intrinsic_vulkan_resource_reindex:
lower_res_reindex_intrinsic(intrin, state);
break;
case nir_intrinsic_load_vulkan_descriptor:
lower_load_vulkan_descriptor(intrin, state);
break;
case nir_intrinsic_get_buffer_size:
lower_get_buffer_size(intrin, state);
break;
anv,i965: Lower away image derefs in the driver Previously, the back-end compiler turn image access into magic uniform reads and there was a complex contract between back-end compiler and driver about setting up and filling out those params. As of this commit, both drivers now lower image_deref_load_param_intel intrinsics to load_uniform intrinsics controlled by the driver and lower the other image_deref_* intrinsics to image_* intrinsics which take an actual binding table index. There are still "magic" uniforms but they are now added and controlled entirely by the driver and that contract no longer spans components. This also has the side-effect of making most image use compile-time binding table indices. Previously, all image access pulled the binding table index from a uniform. Part of the reason for this was that the magic uniforms made it difficult to decouple binding table indices from the uniforms and, since they are indexed completely differently (especially in Vulkan), it was hard to pull them apart. Now that the driver is handling both, it's trivial to decouple the two and provide actual binding table indices. Shader-db results on Kaby Lake: total instructions in shared programs: 15166872 -> 15164293 (-0.02%) instructions in affected programs: 115834 -> 113255 (-2.23%) helped: 191 HURT: 0 total cycles in shared programs: 571311495 -> 571196465 (-0.02%) cycles in affected programs: 4757115 -> 4642085 (-2.42%) helped: 73 HURT: 67 total spills in shared programs: 10951 -> 10926 (-0.23%) spills in affected programs: 742 -> 717 (-3.37%) helped: 7 HURT: 0 total fills in shared programs: 22226 -> 22201 (-0.11%) fills in affected programs: 1146 -> 1121 (-2.18%) helped: 7 HURT: 0 Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
2018-08-16 22:23:10 +01:00
case nir_intrinsic_image_deref_load:
case nir_intrinsic_image_deref_store:
case nir_intrinsic_image_deref_atomic_add:
case nir_intrinsic_image_deref_atomic_min:
case nir_intrinsic_image_deref_atomic_max:
case nir_intrinsic_image_deref_atomic_and:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
case nir_intrinsic_image_deref_atomic_exchange:
case nir_intrinsic_image_deref_atomic_comp_swap:
case nir_intrinsic_image_deref_size:
case nir_intrinsic_image_deref_samples:
case nir_intrinsic_image_deref_load_param_intel:
case nir_intrinsic_image_deref_load_raw_intel:
case nir_intrinsic_image_deref_store_raw_intel:
lower_image_intrinsic(intrin, state);
break;
case nir_intrinsic_load_constant:
lower_load_constant(intrin, state);
break;
default:
break;
}
break;
}
case nir_instr_type_tex:
lower_tex(nir_instr_as_tex(instr), state);
break;
default:
continue;
}
}
}
struct binding_info {
uint32_t binding;
uint8_t set;
uint16_t score;
};
static int
compare_binding_infos(const void *_a, const void *_b)
{
const struct binding_info *a = _a, *b = _b;
if (a->score != b->score)
return b->score - a->score;
if (a->set != b->set)
return a->set - b->set;
return a->binding - b->binding;
}
void
anv_nir_apply_pipeline_layout(const struct anv_physical_device *pdevice,
bool robust_buffer_access,
struct anv_pipeline_layout *layout,
nir_shader *shader,
struct brw_stage_prog_data *prog_data,
struct anv_pipeline_bind_map *map)
{
void *mem_ctx = ralloc_context(NULL);
struct apply_pipeline_layout_state state = {
.pdevice = pdevice,
.shader = shader,
.layout = layout,
.add_bounds_checks = robust_buffer_access,
.lowered_instrs = _mesa_pointer_set_create(mem_ctx),
};
for (unsigned s = 0; s < layout->num_sets; s++) {
const unsigned count = layout->set[s].layout->binding_count;
state.set[s].use_count = rzalloc_array(mem_ctx, uint8_t, count);
state.set[s].surface_offsets = rzalloc_array(mem_ctx, uint8_t, count);
state.set[s].sampler_offsets = rzalloc_array(mem_ctx, uint8_t, count);
}
nir_foreach_function(function, shader) {
if (!function->impl)
continue;
nir_foreach_block(block, function->impl)
get_used_bindings_block(block, &state);
}
for (unsigned s = 0; s < layout->num_sets; s++) {
if (state.set[s].desc_buffer_used) {
map->surface_to_descriptor[map->surface_count] =
(struct anv_pipeline_binding) {
.set = ANV_DESCRIPTOR_SET_DESCRIPTORS,
.binding = s,
};
state.set[s].desc_offset = map->surface_count;
map->surface_count++;
}
}
if (state.uses_constants) {
state.constants_offset = map->surface_count;
map->surface_to_descriptor[map->surface_count].set =
ANV_DESCRIPTOR_SET_SHADER_CONSTANTS;
map->surface_count++;
}
unsigned used_binding_count = 0;
for (uint32_t set = 0; set < layout->num_sets; set++) {
struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
for (unsigned b = 0; b < set_layout->binding_count; b++) {
if (state.set[set].use_count[b] == 0)
continue;
used_binding_count++;
}
}
struct binding_info *infos =
rzalloc_array(mem_ctx, struct binding_info, used_binding_count);
used_binding_count = 0;
for (uint32_t set = 0; set < layout->num_sets; set++) {
struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
for (unsigned b = 0; b < set_layout->binding_count; b++) {
if (state.set[set].use_count[b] == 0)
continue;
struct anv_descriptor_set_binding_layout *binding =
&layout->set[set].layout->binding[b];
/* Do a fixed-point calculation to generate a score based on the
* number of uses and the binding array size. We shift by 7 instead
* of 8 because we're going to use the top bit below to make
* everything which does not support bindless super higher priority
* than things which do.
*/
uint16_t score = ((uint16_t)state.set[set].use_count[b] << 7) /
binding->array_size;
/* If the descriptor type doesn't support bindless then put it at the
* beginning so we guarantee it gets a slot.
*/
if (!anv_descriptor_supports_bindless(pdevice, binding, true) ||
!anv_descriptor_supports_bindless(pdevice, binding, false))
score |= 1 << 15;
infos[used_binding_count++] = (struct binding_info) {
.set = set,
.binding = b,
.score = score,
};
}
}
/* Order the binding infos based on score with highest scores first. If
* scores are equal we then order by set and binding.
*/
qsort(infos, used_binding_count, sizeof(struct binding_info),
compare_binding_infos);
for (unsigned i = 0; i < used_binding_count; i++) {
unsigned set = infos[i].set, b = infos[i].binding;
struct anv_descriptor_set_binding_layout *binding =
&layout->set[set].layout->binding[b];
const uint32_t array_size = binding->array_size;
if (binding->data & ANV_DESCRIPTOR_SURFACE_STATE) {
if (map->surface_count + array_size > MAX_BINDING_TABLE_SIZE ||
anv_descriptor_requires_bindless(pdevice, binding, false)) {
/* If this descriptor doesn't fit in the binding table or if it
* requires bindless for some reason, flag it as bindless.
*/
assert(anv_descriptor_supports_bindless(pdevice, binding, false));
state.set[set].surface_offsets[b] = BINDLESS_OFFSET;
} else {
state.set[set].surface_offsets[b] = map->surface_count;
struct anv_sampler **samplers = binding->immutable_samplers;
for (unsigned i = 0; i < binding->array_size; i++) {
uint8_t planes = samplers ? samplers[i]->n_planes : 1;
for (uint8_t p = 0; p < planes; p++) {
map->surface_to_descriptor[map->surface_count++] =
(struct anv_pipeline_binding) {
.set = set,
.binding = b,
.index = i,
.plane = p,
};
}
}
}
assert(map->surface_count <= MAX_BINDING_TABLE_SIZE);
}
if (binding->data & ANV_DESCRIPTOR_SAMPLER_STATE) {
if (map->sampler_count + array_size > MAX_SAMPLER_TABLE_SIZE ||
anv_descriptor_requires_bindless(pdevice, binding, true)) {
/* If this descriptor doesn't fit in the binding table or if it
* requires bindless for some reason, flag it as bindless.
*/
assert(anv_descriptor_supports_bindless(pdevice, binding, true));
state.set[set].sampler_offsets[b] = BINDLESS_OFFSET;
} else {
state.set[set].sampler_offsets[b] = map->sampler_count;
struct anv_sampler **samplers = binding->immutable_samplers;
for (unsigned i = 0; i < binding->array_size; i++) {
uint8_t planes = samplers ? samplers[i]->n_planes : 1;
for (uint8_t p = 0; p < planes; p++) {
map->sampler_to_descriptor[map->sampler_count++] =
(struct anv_pipeline_binding) {
.set = set,
.binding = b,
.index = i,
.plane = p,
};
}
}
}
}
}
nir_foreach_variable(var, &shader->uniforms) {
spirv: Get rid of vtn_variable_mode_image/sampler vtn_variable_mode_image and _sampler are instead replaced with vtn_variable_mode_uniform which encompasses both of them. In the few places where it was neccessary to distinguish between the two, the GLSL type of the pointer is used instead. The main reason to do this is that on OpenGL it is permitted to put images and samplers into structs and declare a uniform with them. That means that variables can now have a mix of uniform, sampler and image modes so picking a single one of those modes for a variable no longer makes sense. This fixes OpLoad on a sampler within a struct which was previously using the variable mode to determine whether it was a sampler or not. The type of the variable is a struct so it was not being considered to be uniform mode even though the member being loaded should be sampler mode. The previous code appeared to be using var->interface_type as a place to store the type of the variable without the enclosing array for images and samplers. I guess this worked because opaque types can not appear in interfaces so the interface_type is sort of unused. This patch removes the overloading of var->interface_type and any places that needed the type without the array can now just deduce it from var->type. v2: squash in this patch the changes to anv/nir (Timothy) Signed-off-by: Eduardo Lima <elima@igalia.com> Signed-off-by: Neil Roberts <nroberts@igalia.com Signed-off-by: Alejandro Piñeiro <apinheiro@igalia.com> Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2018-03-01 16:51:58 +00:00
const struct glsl_type *glsl_type = glsl_without_array(var->type);
if (!glsl_type_is_image(glsl_type))
continue;
spirv: Get rid of vtn_variable_mode_image/sampler vtn_variable_mode_image and _sampler are instead replaced with vtn_variable_mode_uniform which encompasses both of them. In the few places where it was neccessary to distinguish between the two, the GLSL type of the pointer is used instead. The main reason to do this is that on OpenGL it is permitted to put images and samplers into structs and declare a uniform with them. That means that variables can now have a mix of uniform, sampler and image modes so picking a single one of those modes for a variable no longer makes sense. This fixes OpLoad on a sampler within a struct which was previously using the variable mode to determine whether it was a sampler or not. The type of the variable is a struct so it was not being considered to be uniform mode even though the member being loaded should be sampler mode. The previous code appeared to be using var->interface_type as a place to store the type of the variable without the enclosing array for images and samplers. I guess this worked because opaque types can not appear in interfaces so the interface_type is sort of unused. This patch removes the overloading of var->interface_type and any places that needed the type without the array can now just deduce it from var->type. v2: squash in this patch the changes to anv/nir (Timothy) Signed-off-by: Eduardo Lima <elima@igalia.com> Signed-off-by: Neil Roberts <nroberts@igalia.com Signed-off-by: Alejandro Piñeiro <apinheiro@igalia.com> Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
2018-03-01 16:51:58 +00:00
enum glsl_sampler_dim dim = glsl_get_sampler_dim(glsl_type);
const uint32_t set = var->data.descriptor_set;
const uint32_t binding = var->data.binding;
const uint32_t array_size =
layout->set[set].layout->binding[binding].array_size;
if (state.set[set].use_count[binding] == 0)
continue;
if (state.set[set].surface_offsets[binding] >= MAX_BINDING_TABLE_SIZE)
continue;
struct anv_pipeline_binding *pipe_binding =
&map->surface_to_descriptor[state.set[set].surface_offsets[binding]];
for (unsigned i = 0; i < array_size; i++) {
assert(pipe_binding[i].set == set);
assert(pipe_binding[i].binding == binding);
assert(pipe_binding[i].index == i);
if (dim == GLSL_SAMPLER_DIM_SUBPASS ||
dim == GLSL_SAMPLER_DIM_SUBPASS_MS)
pipe_binding[i].input_attachment_index = var->data.index + i;
pipe_binding[i].write_only =
(var->data.image.access & ACCESS_NON_READABLE) != 0;
}
}
nir_foreach_function(function, shader) {
if (!function->impl)
continue;
/* Before we do the normal lowering, we look for any SSBO operations
* that we can lower to the BTI model and lower them up-front. The BTI
* model can perform better than the A64 model for a couple reasons:
*
* 1. 48-bit address calculations are potentially expensive and using
* the BTI model lets us simply compute 32-bit offsets and the
* hardware adds the 64-bit surface base address.
*
* 2. The BTI messages, because they use surface states, do bounds
* checking for us. With the A64 model, we have to do our own
* bounds checking and this means wider pointers and extra
* calculations and branching in the shader.
*
* The solution to both of these is to convert things to the BTI model
* opportunistically. The reason why we need to do this as a pre-pass
* is for two reasons:
*
* 1. The BTI model requires nir_address_format_32bit_index_offset
* pointers which are not the same type as the pointers needed for
* the A64 model. Because all our derefs are set up for the A64
* model (in case we have variable pointers), we have to crawl all
* the way back to the vulkan_resource_index intrinsic and build a
* completely fresh index+offset calculation.
*
* 2. Because the variable-pointers-capable lowering that we do as part
* of apply_pipeline_layout_block is destructive (It really has to
* be to handle variable pointers properly), we've lost the deref
* information by the time we get to the load/store/atomic
* intrinsics in that pass.
*/
lower_direct_buffer_access(function->impl, &state);
nir_builder_init(&state.builder, function->impl);
nir_foreach_block(block, function->impl)
apply_pipeline_layout_block(block, &state);
nir_metadata_preserve(function->impl, nir_metadata_block_index |
nir_metadata_dominance);
}
ralloc_free(mem_ctx);
}