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>

src/tint/transform/num_workgroups_from_uniform.cc

@@ -122,10 +122,31 @@ void NumWorkgroupsFromUniform::Run(CloneContext& ctx, const DataMap& inputs, Dat
             auto* num_workgroups_struct = ctx.dst->Structure(
                 ctx.dst->Sym(),
                 {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(
                 ctx.dst->Sym(), ctx.dst->ty.Of(num_workgroups_struct), ast::StorageClass::kUniform,
-                ast::AttributeList{
-                    ctx.dst->GroupAndBinding(cfg->ubo_binding.group, cfg->ubo_binding.binding)});
+                ast::AttributeList{ctx.dst->GroupAndBinding(group, binding)});
         }
         return num_workgroups_ubo;
     };
@@ -151,7 +172,8 @@ void NumWorkgroupsFromUniform::Run(CloneContext& ctx, const DataMap& inputs, Dat
     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() = default;
 

src/tint/transform/num_workgroups_from_uniform.h

@@ -15,6 +15,8 @@
 #ifndef 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/transform/transform.h"
 
@@ -52,8 +54,11 @@ class NumWorkgroupsFromUniform : public Castable<NumWorkgroupsFromUniform, Trans
     /// Configuration options for the NumWorkgroupsFromUniform transform.
     struct Config : public Castable<Data, transform::Data> {
         /// Constructor
-        /// @param ubo_bp the binding point to use for the generated uniform buffer.
-        explicit Config(sem::BindingPoint ubo_bp);
+        /// @param ubo_bp the binding point to use for the generated uniform buffer. If 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
         Config(const Config&);
@@ -61,8 +66,10 @@ class NumWorkgroupsFromUniform : public Castable<NumWorkgroupsFromUniform, Trans
         /// Destructor
         ~Config() override;
 
-        /// The binding point to use for the generated uniform buffer.
-        sem::BindingPoint ubo_binding;
+        /// The binding point to use for the generated uniform buffer. If ubo_bp contains no value,
+        /// 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

src/tint/transform/num_workgroups_from_uniform_test.cc

@@ -434,5 +434,261 @@ fn main(tint_symbol : tint_symbol_1) {
     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 tint::transform