Tint: num_workgroups use free binding group if not specified

In this patch NumWorkgroupsFromUniform::Config changed to storage
std::optional<sem::BindingPoint>, and if it has no value,
NumWorkgroupsFromUniform will choose a free binding group, i.e.
binding 0 of the largest used group plus 1 is used if at least one
resource is bound, otherwise group 0 binding 0 is used. Tint CLI
is also changed to provide a --hlsl-root-constant-binding-point
option allowing user to specify the binding point for num_workgroups
uniform buffer.

Bug: tint:1566
Change-Id: I3b8c22a4276bab722d901f5b07d23a268786c417
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91980
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Zhaoming Jiang <zhaoming.jiang@intel.com>
This commit is contained in:
Zhaoming Jiang 2022-06-01 10:23:51 +00:00 committed by Dawn LUCI CQ
parent cb6ddd2aa6
commit 84b48cf30c
6 changed files with 351 additions and 10 deletions

View File

@ -645,8 +645,8 @@ ResultOrError<std::string> TranslateToHLSL(dawn::platform::Platform* platform,
tint::writer::hlsl::Options options; tint::writer::hlsl::Options options;
options.disable_workgroup_init = request.disableWorkgroupInit; options.disable_workgroup_init = request.disableWorkgroupInit;
if (request.usesNumWorkgroups) { if (request.usesNumWorkgroups) {
options.root_constant_binding_point.group = request.numWorkgroupsRegisterSpace; options.root_constant_binding_point = tint::sem::BindingPoint{
options.root_constant_binding_point.binding = request.numWorkgroupsShaderRegister; request.numWorkgroupsRegisterSpace, request.numWorkgroupsShaderRegister};
} }
// TODO(dawn:549): HLSL generation outputs the indices into the // TODO(dawn:549): HLSL generation outputs the indices into the
// array_length_from_uniform buffer that were actually used. When the blob cache can // array_length_from_uniform buffer that were actually used. When the blob cache can

View File

@ -15,7 +15,9 @@
#include <cstdio> #include <cstdio>
#include <fstream> #include <fstream>
#include <iostream> #include <iostream>
#include <limits>
#include <memory> #include <memory>
#include <optional> // NOLINT(build/include_order)
#include <sstream> #include <sstream>
#include <string> #include <string>
#include <vector> #include <vector>
@ -86,6 +88,7 @@ struct Options {
std::string dxc_path; std::string dxc_path;
std::string xcrun_path; std::string xcrun_path;
std::vector<std::string> overrides; std::vector<std::string> overrides;
std::optional<tint::sem::BindingPoint> hlsl_root_constant_binding_point;
}; };
const char kUsage[] = R"(Usage: tint [options] <input-file> const char kUsage[] = R"(Usage: tint [options] <input-file>
@ -112,6 +115,11 @@ ${transforms}
Affects AST dumping, and text-based output languages. Affects AST dumping, and text-based output languages.
--dump-inspector-bindings -- Dump reflection data about bindins to stdout. --dump-inspector-bindings -- Dump reflection data about bindins to stdout.
-h -- This help text -h -- This help text
--hlsl-root-constant-binding-point <group>,<binding> -- Binding point for root constant.
Specify the binding point for generated uniform buffer
used for num_workgroups in HLSL. If not specified, then
default to binding 0 of the largest used group plus 1,
or group 0 if no resource bound.
--validate -- Validates the generated shader --validate -- Validates the generated shader
--fxc -- Ask to validate HLSL output using FXC instead of DXC. --fxc -- Ask to validate HLSL output using FXC instead of DXC.
When specified, automatically enables --validate When specified, automatically enables --validate
@ -220,6 +228,27 @@ std::vector<std::string> split_on_comma(std::string list) {
return res; return res;
} }
std::optional<uint64_t> parse_unsigned_number(std::string number) {
for (char c : number) {
if (!std::isdigit(c)) {
// Found a non-digital char, return nullopt
return std::nullopt;
}
}
errno = 0;
char* p_end;
uint64_t result;
// std::strtoull will not throw exception.
result = std::strtoull(number.c_str(), &p_end, 10);
if ((errno != 0) || (static_cast<size_t>(p_end - number.c_str()) != number.length())) {
// Unexpected conversion result
return std::nullopt;
}
return result;
}
std::string TextureDimensionToString(tint::inspector::ResourceBinding::TextureDimension dim) { std::string TextureDimensionToString(tint::inspector::ResourceBinding::TextureDimension dim) {
switch (dim) { switch (dim) {
case tint::inspector::ResourceBinding::TextureDimension::kNone: case tint::inspector::ResourceBinding::TextureDimension::kNone:
@ -402,6 +431,31 @@ bool ParseArgs(const std::vector<std::string>& args, Options* opts) {
return false; return false;
} }
opts->overrides = split_on_comma(args[i]); opts->overrides = split_on_comma(args[i]);
} else if (arg == "--hlsl-root-constant-binding-point") {
++i;
if (i >= args.size()) {
std::cerr << "Missing value for " << arg << std::endl;
return false;
}
auto binding_points = split_on_comma(args[i]);
if (binding_points.size() != 2) {
std::cerr << "Invalid binding point for " << arg << ": " << args[i] << std::endl;
return false;
}
auto group = parse_unsigned_number(binding_points[0]);
if ((!group.has_value()) || (group.value() > std::numeric_limits<uint32_t>::max())) {
std::cerr << "Invalid group for " << arg << ": " << binding_points[0] << std::endl;
return false;
}
auto binding = parse_unsigned_number(binding_points[1]);
if ((!binding.has_value()) ||
(binding.value() > std::numeric_limits<uint32_t>::max())) {
std::cerr << "Invalid binding for " << arg << ": " << binding_points[1]
<< std::endl;
return false;
}
opts->hlsl_root_constant_binding_point = tint::sem::BindingPoint{
static_cast<uint32_t>(group.value()), static_cast<uint32_t>(binding.value())};
} else if (!arg.empty()) { } else if (!arg.empty()) {
if (arg[0] == '-') { if (arg[0] == '-') {
std::cerr << "Unrecognized option: " << arg << std::endl; std::cerr << "Unrecognized option: " << arg << std::endl;
@ -723,6 +777,7 @@ bool GenerateHlsl(const tint::Program* program, const Options& options) {
tint::writer::hlsl::Options gen_options; tint::writer::hlsl::Options gen_options;
gen_options.disable_workgroup_init = options.disable_workgroup_init; gen_options.disable_workgroup_init = options.disable_workgroup_init;
gen_options.generate_external_texture_bindings = true; gen_options.generate_external_texture_bindings = true;
gen_options.root_constant_binding_point = options.hlsl_root_constant_binding_point;
auto result = tint::writer::hlsl::Generate(program, gen_options); auto result = tint::writer::hlsl::Generate(program, gen_options);
if (!result.success) { if (!result.success) {
PrintWGSL(std::cerr, *program); PrintWGSL(std::cerr, *program);

View File

@ -122,10 +122,31 @@ void NumWorkgroupsFromUniform::Run(CloneContext& ctx, const DataMap& inputs, Dat
auto* num_workgroups_struct = ctx.dst->Structure( auto* num_workgroups_struct = ctx.dst->Structure(
ctx.dst->Sym(), ctx.dst->Sym(),
{ctx.dst->Member(kNumWorkgroupsMemberName, ctx.dst->ty.vec3(ctx.dst->ty.u32()))}); {ctx.dst->Member(kNumWorkgroupsMemberName, ctx.dst->ty.vec3(ctx.dst->ty.u32()))});
uint32_t group, binding;
if (cfg->ubo_binding.has_value()) {
// If cfg->ubo_binding holds a value, use the specified binding point.
group = cfg->ubo_binding->group;
binding = cfg->ubo_binding->binding;
} else {
// If cfg->ubo_binding holds no value, use the binding 0 of the largest used group
// plus 1, or group 0 if no resource bound.
group = 0;
for (auto* var : ctx.src->AST().GlobalVariables()) {
if (auto binding_point = var->BindingPoint()) {
if (binding_point.group->value >= group) {
group = binding_point.group->value + 1;
}
}
}
binding = 0;
}
num_workgroups_ubo = ctx.dst->Global( num_workgroups_ubo = ctx.dst->Global(
ctx.dst->Sym(), ctx.dst->ty.Of(num_workgroups_struct), ast::StorageClass::kUniform, ctx.dst->Sym(), ctx.dst->ty.Of(num_workgroups_struct), ast::StorageClass::kUniform,
ast::AttributeList{ ast::AttributeList{ctx.dst->GroupAndBinding(group, binding)});
ctx.dst->GroupAndBinding(cfg->ubo_binding.group, cfg->ubo_binding.binding)});
} }
return num_workgroups_ubo; return num_workgroups_ubo;
}; };
@ -151,7 +172,8 @@ void NumWorkgroupsFromUniform::Run(CloneContext& ctx, const DataMap& inputs, Dat
ctx.Clone(); ctx.Clone();
} }
NumWorkgroupsFromUniform::Config::Config(sem::BindingPoint ubo_bp) : ubo_binding(ubo_bp) {} NumWorkgroupsFromUniform::Config::Config(std::optional<sem::BindingPoint> ubo_bp)
: ubo_binding(ubo_bp) {}
NumWorkgroupsFromUniform::Config::Config(const Config&) = default; NumWorkgroupsFromUniform::Config::Config(const Config&) = default;
NumWorkgroupsFromUniform::Config::~Config() = default; NumWorkgroupsFromUniform::Config::~Config() = default;

View File

@ -15,6 +15,8 @@
#ifndef SRC_TINT_TRANSFORM_NUM_WORKGROUPS_FROM_UNIFORM_H_ #ifndef SRC_TINT_TRANSFORM_NUM_WORKGROUPS_FROM_UNIFORM_H_
#define SRC_TINT_TRANSFORM_NUM_WORKGROUPS_FROM_UNIFORM_H_ #define SRC_TINT_TRANSFORM_NUM_WORKGROUPS_FROM_UNIFORM_H_
#include <optional> // NOLINT(build/include_order)
#include "src/tint/sem/binding_point.h" #include "src/tint/sem/binding_point.h"
#include "src/tint/transform/transform.h" #include "src/tint/transform/transform.h"
@ -52,8 +54,11 @@ class NumWorkgroupsFromUniform : public Castable<NumWorkgroupsFromUniform, Trans
/// Configuration options for the NumWorkgroupsFromUniform transform. /// Configuration options for the NumWorkgroupsFromUniform transform.
struct Config : public Castable<Data, transform::Data> { struct Config : public Castable<Data, transform::Data> {
/// Constructor /// Constructor
/// @param ubo_bp the binding point to use for the generated uniform buffer. /// @param ubo_bp the binding point to use for the generated uniform buffer. If ubo_bp
explicit Config(sem::BindingPoint ubo_bp); /// contains no value, a free binding point will be used to ensure the generated program is
/// valid. Specifically, binding 0 of the largest used group plus 1 is used if at least one
/// resource is bound, otherwise group 0 binding 0 is used.
explicit Config(std::optional<sem::BindingPoint> ubo_bp);
/// Copy constructor /// Copy constructor
Config(const Config&); Config(const Config&);
@ -61,8 +66,10 @@ class NumWorkgroupsFromUniform : public Castable<NumWorkgroupsFromUniform, Trans
/// Destructor /// Destructor
~Config() override; ~Config() override;
/// The binding point to use for the generated uniform buffer. /// The binding point to use for the generated uniform buffer. If ubo_bp contains no value,
sem::BindingPoint ubo_binding; /// a free binding point will be used. Specifically, binding 0 of the largest used group
/// plus 1 is used if at least one resource is bound, otherwise group 0 binding 0 is used.
std::optional<sem::BindingPoint> ubo_binding;
}; };
/// @param program the program to inspect /// @param program the program to inspect

View File

@ -434,5 +434,261 @@ fn main(tint_symbol : tint_symbol_1) {
EXPECT_EQ(expect, str(got)); EXPECT_EQ(expect, str(got));
} }
// Test that group 0 binding 0 is used if no bound resource in the program and binding point is not
// specified in NumWorkgroupsFromUniform::Config.
TEST_F(NumWorkgroupsFromUniformTest, UnspecifiedBindingPoint_NoResourceBound) {
auto* src = R"(
struct Builtins1 {
@builtin(num_workgroups) num_wgs : vec3<u32>,
};
struct Builtins2 {
@builtin(global_invocation_id) gid : vec3<u32>,
@builtin(num_workgroups) num_wgs : vec3<u32>,
@builtin(workgroup_id) wgid : vec3<u32>,
};
@stage(compute) @workgroup_size(1)
fn main1(in : Builtins1) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main2(in : Builtins2) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main3(@builtin(num_workgroups) num_wgs : vec3<u32>) {
let groups_x = num_wgs.x;
let groups_y = num_wgs.y;
let groups_z = num_wgs.z;
}
)";
auto* expect = R"(
struct tint_symbol_6 {
num_workgroups : vec3<u32>,
}
@group(0) @binding(0) var<uniform> tint_symbol_7 : tint_symbol_6;
struct Builtins1 {
num_wgs : vec3<u32>,
}
struct Builtins2 {
gid : vec3<u32>,
num_wgs : vec3<u32>,
wgid : vec3<u32>,
}
fn main1_inner(in : Builtins1) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main1() {
main1_inner(Builtins1(tint_symbol_7.num_workgroups));
}
struct tint_symbol_3 {
@builtin(global_invocation_id)
gid : vec3<u32>,
@builtin(workgroup_id)
wgid : vec3<u32>,
}
fn main2_inner(in : Builtins2) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main2(tint_symbol_2 : tint_symbol_3) {
main2_inner(Builtins2(tint_symbol_2.gid, tint_symbol_7.num_workgroups, tint_symbol_2.wgid));
}
fn main3_inner(num_wgs : vec3<u32>) {
let groups_x = num_wgs.x;
let groups_y = num_wgs.y;
let groups_z = num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main3() {
main3_inner(tint_symbol_7.num_workgroups);
}
)";
DataMap data;
data.Add<CanonicalizeEntryPointIO::Config>(CanonicalizeEntryPointIO::ShaderStyle::kHlsl);
// Make binding point unspecified.
data.Add<NumWorkgroupsFromUniform::Config>(std::nullopt);
auto got = Run<Unshadow, CanonicalizeEntryPointIO, NumWorkgroupsFromUniform>(src, data);
EXPECT_EQ(expect, str(got));
}
// Test that binding 0 of the largest used group plus 1 is used if at least one resource is bound in
// the program and binding point is not specified in NumWorkgroupsFromUniform::Config.
TEST_F(NumWorkgroupsFromUniformTest, UnspecifiedBindingPoint_MultipleResourceBound) {
auto* src = R"(
struct Builtins1 {
@builtin(num_workgroups) num_wgs : vec3<u32>,
};
struct Builtins2 {
@builtin(global_invocation_id) gid : vec3<u32>,
@builtin(num_workgroups) num_wgs : vec3<u32>,
@builtin(workgroup_id) wgid : vec3<u32>,
};
struct S0 {
@size(4)
m0 : u32,
m1 : array<u32>,
};
struct S1 {
@size(4)
m0 : u32,
m1 : array<u32, 6>,
};
@group(0) @binding(0) var g2 : texture_2d<f32>;
@group(1) @binding(0) var g3 : texture_depth_2d;
@group(1) @binding(1) var g4 : texture_storage_2d<rg32float, write>;
@group(3) @binding(0) var g5 : texture_depth_cube_array;
@group(4) @binding(0) var g6 : texture_external;
@group(0) @binding(1) var<storage, write> g8 : S0;
@group(1) @binding(3) var<storage, read> g9 : S0;
@group(3) @binding(2) var<storage, read_write> g10 : S0;
@stage(compute) @workgroup_size(1)
fn main1(in : Builtins1) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
g8.m0 = 1u;
}
@stage(compute) @workgroup_size(1)
fn main2(in : Builtins2) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main3(@builtin(num_workgroups) num_wgs : vec3<u32>) {
let groups_x = num_wgs.x;
let groups_y = num_wgs.y;
let groups_z = num_wgs.z;
}
)";
auto* expect = R"(
struct tint_symbol_6 {
num_workgroups : vec3<u32>,
}
@group(5) @binding(0) var<uniform> tint_symbol_7 : tint_symbol_6;
struct Builtins1 {
num_wgs : vec3<u32>,
}
struct Builtins2 {
gid : vec3<u32>,
num_wgs : vec3<u32>,
wgid : vec3<u32>,
}
struct S0 {
@size(4)
m0 : u32,
m1 : array<u32>,
}
struct S1 {
@size(4)
m0 : u32,
m1 : array<u32, 6>,
}
@group(0) @binding(0) var g2 : texture_2d<f32>;
@group(1) @binding(0) var g3 : texture_depth_2d;
@group(1) @binding(1) var g4 : texture_storage_2d<rg32float, write>;
@group(3) @binding(0) var g5 : texture_depth_cube_array;
@group(4) @binding(0) var g6 : texture_external;
@group(0) @binding(1) var<storage, write> g8 : S0;
@group(1) @binding(3) var<storage, read> g9 : S0;
@group(3) @binding(2) var<storage, read_write> g10 : S0;
fn main1_inner(in : Builtins1) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
g8.m0 = 1u;
}
@stage(compute) @workgroup_size(1)
fn main1() {
main1_inner(Builtins1(tint_symbol_7.num_workgroups));
}
struct tint_symbol_3 {
@builtin(global_invocation_id)
gid : vec3<u32>,
@builtin(workgroup_id)
wgid : vec3<u32>,
}
fn main2_inner(in : Builtins2) {
let groups_x = in.num_wgs.x;
let groups_y = in.num_wgs.y;
let groups_z = in.num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main2(tint_symbol_2 : tint_symbol_3) {
main2_inner(Builtins2(tint_symbol_2.gid, tint_symbol_7.num_workgroups, tint_symbol_2.wgid));
}
fn main3_inner(num_wgs : vec3<u32>) {
let groups_x = num_wgs.x;
let groups_y = num_wgs.y;
let groups_z = num_wgs.z;
}
@stage(compute) @workgroup_size(1)
fn main3() {
main3_inner(tint_symbol_7.num_workgroups);
}
)";
DataMap data;
data.Add<CanonicalizeEntryPointIO::Config>(CanonicalizeEntryPointIO::ShaderStyle::kHlsl);
// Make binding point unspecified.
data.Add<NumWorkgroupsFromUniform::Config>(std::nullopt);
auto got = Run<Unshadow, CanonicalizeEntryPointIO, NumWorkgroupsFromUniform>(src, data);
EXPECT_EQ(expect, str(got));
}
} // namespace } // namespace
} // namespace tint::transform } // namespace tint::transform

View File

@ -16,6 +16,7 @@
#define SRC_TINT_WRITER_HLSL_GENERATOR_H_ #define SRC_TINT_WRITER_HLSL_GENERATOR_H_
#include <memory> #include <memory>
#include <optional> // NOLINT(build/include_order)
#include <string> #include <string>
#include <unordered_set> #include <unordered_set>
#include <utility> #include <utility>
@ -46,7 +47,7 @@ struct Options {
Options& operator=(const Options&); Options& operator=(const Options&);
/// The binding point to use for information passed via root constants. /// The binding point to use for information passed via root constants.
sem::BindingPoint root_constant_binding_point; std::optional<sem::BindingPoint> root_constant_binding_point;
/// Set to `true` to disable workgroup memory zero initialization /// Set to `true` to disable workgroup memory zero initialization
bool disable_workgroup_init = false; bool disable_workgroup_init = false;
/// Set to 'true' to generates binding mappings for external textures /// Set to 'true' to generates binding mappings for external textures