[Impeller][Compute] Guard subgroups with feature detection

impeller/fixtures/stroke.comp

@@ -17,6 +17,7 @@ layout(binding = 0) buffer Polyline {
 polyline;
 
 layout(binding = 1) buffer VertexBuffer {
+  uint count;
   vec2 position[];
 }
 vertex_buffer;
@@ -40,6 +41,8 @@ void main() {
     return;
   }
 
+  atomicAdd(vertex_buffer.count, 4);
+
   vec2 offset = compute_offset(ident);
   uint index = ident - 1;
   vertex_buffer.position[index * 4 + 0] = polyline.data[ident - 1] + offset;

impeller/renderer/backend/gles/context_gles.cc

@@ -81,6 +81,7 @@ ContextGLES::ContextGLES(std::unique_ptr<ProcTableGLES> gl,
             .SetSupportsFramebufferFetch(false)
             .SetDefaultColorFormat(PixelFormat::kB8G8R8A8UNormInt)
             .SetDefaultStencilFormat(PixelFormat::kS8UInt)
+            .SetSupportsCompute(false, false)
             .Build();
   }
 

impeller/renderer/backend/metal/context_mtl.mm

@@ -90,6 +90,14 @@
 #endif
 
   {
+    bool supports_subgroups = false;
+    // Refer to the "SIMD-scoped reduction operations" feature in the table
+    // below: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+    if (@available(ios 13.0, tvos 13.0, macos 10.15, *)) {
+      supports_subgroups = [device supportsFamily:MTLGPUFamilyApple7] ||
+                           [device supportsFamily:MTLGPUFamilyMac2];
+    }
+
     device_capabilities_ =
         DeviceCapabilitiesBuilder()
             .SetHasThreadingRestrictions(false)
@@ -99,6 +107,7 @@
             .SetSupportsFramebufferFetch(SupportsFramebufferFetch())
             .SetDefaultColorFormat(PixelFormat::kB8G8R8A8UNormInt)
             .SetDefaultStencilFormat(PixelFormat::kS8UInt)
+            .SetSupportsCompute(true, supports_subgroups)
             .Build();
   }
 

impeller/renderer/backend/vulkan/context_vk.cc

@@ -557,6 +557,8 @@ ContextVK::ContextVK(
           .SetSupportsFramebufferFetch(false)
           .SetDefaultColorFormat(PixelFormat::kB8G8R8A8UNormInt)
           .SetDefaultStencilFormat(PixelFormat::kS8UInt)
+          // TODO(110622): detect this and enable.
+          .SetSupportsCompute(false, false)
           .Build();
   graphics_command_pool_ = std::move(graphics_command_pool.value);
   descriptor_pool_ = std::move(descriptor_pool.value);

impeller/renderer/compute_subgroup_unittests.cc

@@ -35,6 +35,7 @@ TEST_P(ComputeTest, HeartCubicsToStrokeVertices) {
 
   auto context = GetContext();
   ASSERT_TRUE(context);
+  ASSERT_TRUE(context->GetDeviceCapabilities().SupportsComputeSubgroups());
 
   auto cmd_buffer = context->CreateCommandBuffer();
   auto pass = cmd_buffer->CreateComputePass();
@@ -182,6 +183,7 @@ TEST_P(ComputeTest, HeartCubicsToStrokeVertices) {
 
     auto* v = reinterpret_cast<SS::VertexBuffer<kCubicCount * 10 * 10 * 4>*>(
         vertex_buffer->AsBufferView().contents);
+    EXPECT_EQ(v->count, golden_heart_vertices.size());
     for (size_t i = 0; i < golden_heart_vertices.size(); i += 1) {
       EXPECT_LT(std::abs(golden_heart_vertices[i].x - v->position[i].x), 1e-3);
       EXPECT_LT(std::abs(golden_heart_vertices[i].y - v->position[i].y), 1e-3);
@@ -197,6 +199,7 @@ TEST_P(ComputeTest, QuadsToPolyline) {
   using QS = QuadPolylineComputeShader;
   auto context = GetContext();
   ASSERT_TRUE(context);
+  ASSERT_TRUE(context->GetDeviceCapabilities().SupportsComputeSubgroups());
 
   auto cmd_buffer = context->CreateCommandBuffer();
   auto pass = cmd_buffer->CreateComputePass();
@@ -259,7 +262,6 @@ TEST_P(ComputeTest, QuadsToPolyline) {
           EXPECT_LT(std::abs(p->data[i].x - golden_heart_points[i].x), 1e-3);
           EXPECT_LT(std::abs(p->data[i].y - golden_heart_points[i].y), 1e-3);
         }
-
         latch.Signal();
       }));
 

impeller/renderer/compute_unittests.cc

@@ -28,6 +28,8 @@ TEST_P(ComputeTest, CanCreateComputePass) {
   using CS = SampleComputeShader;
   auto context = GetContext();
   ASSERT_TRUE(context);
+  ASSERT_TRUE(context->GetDeviceCapabilities().SupportsCompute());
+
   using SamplePipelineBuilder = ComputePipelineBuilder<CS>;
   auto pipeline_desc =
       SamplePipelineBuilder::MakeDefaultPipelineDescriptor(*context);
@@ -114,6 +116,7 @@ TEST_P(ComputeTest, MultiStageInputAndOutput) {
 
   auto context = GetContext();
   ASSERT_TRUE(context);
+  ASSERT_TRUE(context->GetDeviceCapabilities().SupportsCompute());
 
   auto pipeline_desc_1 =
       Stage1PipelineBuilder::MakeDefaultPipelineDescriptor(*context);

impeller/renderer/device_capabilities.cc

@@ -3,6 +3,7 @@
 // found in the LICENSE file.
 
 #include "impeller/renderer/device_capabilities.h"
+#include "device_capabilities.h"
 
 namespace impeller {
 
@@ -12,14 +13,18 @@ IDeviceCapabilities::IDeviceCapabilities(bool has_threading_restrictions,
                                          bool supports_texture_to_texture_blits,
                                          bool supports_framebuffer_fetch,
                                          PixelFormat default_color_format,
-                                         PixelFormat default_stencil_format)
+                                         PixelFormat default_stencil_format,
+                                         bool supports_compute,
+                                         bool supports_compute_subgroups)
     : has_threading_restrictions_(has_threading_restrictions),
       supports_offscreen_msaa_(supports_offscreen_msaa),
       supports_ssbo_(supports_ssbo),
       supports_texture_to_texture_blits_(supports_texture_to_texture_blits),
       supports_framebuffer_fetch_(supports_framebuffer_fetch),
       default_color_format_(default_color_format),
-      default_stencil_format_(default_stencil_format) {}
+      default_stencil_format_(default_stencil_format),
+      supports_compute_(supports_compute),
+      supports_compute_subgroups_(supports_compute_subgroups) {}
 
 IDeviceCapabilities::~IDeviceCapabilities() = default;
 
@@ -51,6 +56,14 @@ PixelFormat IDeviceCapabilities::GetDefaultStencilFormat() const {
   return default_stencil_format_;
 }
 
+bool IDeviceCapabilities::SupportsCompute() const {
+  return supports_compute_;
+}
+
+bool IDeviceCapabilities::SupportsComputeSubgroups() const {
+  return supports_compute_subgroups_;
+}
+
 DeviceCapabilitiesBuilder::DeviceCapabilitiesBuilder() = default;
 
 DeviceCapabilitiesBuilder::~DeviceCapabilitiesBuilder() = default;
@@ -97,6 +110,14 @@ DeviceCapabilitiesBuilder& DeviceCapabilitiesBuilder::SetDefaultStencilFormat(
   return *this;
 }
 
+DeviceCapabilitiesBuilder& DeviceCapabilitiesBuilder::SetSupportsCompute(
+    bool value,
+    bool subgroups) {
+  supports_compute_ = value;
+  supports_compute_subgroups_ = subgroups;
+  return *this;
+}
+
 std::unique_ptr<IDeviceCapabilities> DeviceCapabilitiesBuilder::Build() {
   FML_CHECK(default_color_format_.has_value())
       << "Default color format not set";
@@ -110,7 +131,9 @@ std::unique_ptr<IDeviceCapabilities> DeviceCapabilitiesBuilder::Build() {
       supports_texture_to_texture_blits_,                       //
       supports_framebuffer_fetch_,                              //
       *default_color_format_,                                   //
-      *default_stencil_format_                                  //
+      *default_stencil_format_,                                 //
+      supports_compute_,                                        //
+      supports_compute_subgroups_                               //
   );
   return std::unique_ptr<IDeviceCapabilities>(capabilities);
 }

impeller/renderer/device_capabilities.h

@@ -29,14 +29,19 @@ class IDeviceCapabilities {
 
   PixelFormat GetDefaultStencilFormat() const;
 
+  bool SupportsCompute() const;
+  bool SupportsComputeSubgroups() const;
+
  private:
   IDeviceCapabilities(bool has_threading_restrictions,
                       bool supports_offscreen_msaa,
                       bool supports_ssbo,
                       bool supports_texture_to_texture_blits,
                       bool supports_framebuffer_fetch,
                       PixelFormat default_color_format,
-                      PixelFormat default_stencil_format);
+                      PixelFormat default_stencil_format,
+                      bool supports_compute,
+                      bool supports_compute_subgroups);
 
   friend class DeviceCapabilitiesBuilder;
 
@@ -47,6 +52,8 @@ class IDeviceCapabilities {
   bool supports_framebuffer_fetch_ = false;
   PixelFormat default_color_format_;
   PixelFormat default_stencil_format_;
+  bool supports_compute_ = false;
+  bool supports_compute_subgroups_ = false;
 
   FML_DISALLOW_COPY_AND_ASSIGN(IDeviceCapabilities);
 };
@@ -71,6 +78,8 @@ class DeviceCapabilitiesBuilder {
 
   DeviceCapabilitiesBuilder& SetDefaultStencilFormat(PixelFormat value);
 
+  DeviceCapabilitiesBuilder& SetSupportsCompute(bool value, bool subgroups);
+
   std::unique_ptr<IDeviceCapabilities> Build();
 
  private:
@@ -79,6 +88,8 @@ class DeviceCapabilitiesBuilder {
   bool supports_ssbo_ = false;
   bool supports_texture_to_texture_blits_ = false;
   bool supports_framebuffer_fetch_ = false;
+  bool supports_compute_ = false;
+  bool supports_compute_subgroups_ = false;
   std::optional<PixelFormat> default_color_format_ = std::nullopt;
   std::optional<PixelFormat> default_stencil_format_ = std::nullopt;