[Impeller] Switch from glBlitFramebuffer to implicit MSAA resolution. (#47282)

Closes flutter/flutter#137093.

This widens supports from Open GLES 3.x to Open GLES 2.x, and uses [ARM
GPU Best
Practices](https://developer.arm.com/documentation/101897/0301/Fragment-shading/Multisampling-for-OpenGL-ES):

> Do not use `glBlitFramebuffer()` to implement a multisample resolve.

This PR:

- Removes usage of `glBlitFramebuffer`
- Adds the capability check, `SupportsImplicitResolvingMSAA`, which is
`false` outside of GLES
- Does not discard color attachments resolved by
`EXT_multisampled_render_to_texture` (done implicitly)

I spoke to @jonahwilliams about the changes to the HAL, who I believe
also talked to @bdero about it. The short explantation is that, with the
`EXT_multisampled_render_to_texture`, we can be more efficient by
letting GLES perform multisampled rendering for us (no per-sample data
is written out).

See also
https://registry.khronos.org/OpenGL/extensions/EXT/EXT_multisampled_render_to_texture.txt
for details.

Author: matanlurey

impeller/renderer/backend/gles/capabilities_gles.cc

@@ -19,6 +19,10 @@ static const constexpr char* kNvidiaTextureBorderClampExt =
 static const constexpr char* kOESTextureBorderClampExt =
     "GL_OES_texture_border_clamp";
 
+// https://www.khronos.org/registry/OpenGL/extensions/EXT/EXT_multisampled_render_to_texture.txt
+static const constexpr char* kMultisampledRenderToTextureExt =
+    "GL_EXT_multisampled_render_to_texture2";
+
 CapabilitiesGLES::CapabilitiesGLES(const ProcTableGLES& gl) {
   {
     GLint value = 0;
@@ -32,7 +36,9 @@ CapabilitiesGLES::CapabilitiesGLES(const ProcTableGLES& gl) {
     max_cube_map_texture_size = value;
   }
 
-  if (gl.GetDescription()->IsES()) {
+  auto const desc = gl.GetDescription();
+
+  if (desc->IsES()) {
     GLint value = 0;
     gl.GetIntegerv(GL_MAX_FRAGMENT_UNIFORM_VECTORS, &value);
     max_fragment_uniform_vectors = value;
@@ -56,7 +62,7 @@ CapabilitiesGLES::CapabilitiesGLES(const ProcTableGLES& gl) {
     max_texture_size = ISize{value, value};
   }
 
-  if (gl.GetDescription()->IsES()) {
+  if (desc->IsES()) {
     GLint value = 0;
     gl.GetIntegerv(GL_MAX_VARYING_VECTORS, &value);
     max_varying_vectors = value;
@@ -74,7 +80,7 @@ CapabilitiesGLES::CapabilitiesGLES(const ProcTableGLES& gl) {
     max_vertex_texture_image_units = value;
   }
 
-  if (gl.GetDescription()->IsES()) {
+  if (desc->IsES()) {
     GLint value = 0;
     gl.GetIntegerv(GL_MAX_VERTEX_UNIFORM_VECTORS, &value);
     max_vertex_uniform_vectors = value;
@@ -92,31 +98,26 @@ CapabilitiesGLES::CapabilitiesGLES(const ProcTableGLES& gl) {
     num_compressed_texture_formats = value;
   }
 
-  if (gl.GetDescription()->IsES()) {
+  if (desc->IsES()) {
     GLint value = 0;
     gl.GetIntegerv(GL_NUM_SHADER_BINARY_FORMATS, &value);
     num_shader_binary_formats = value;
   }
 
-  supports_framebuffer_fetch_ =
-      gl.GetDescription()->HasExtension(kFramebufferFetchExt);
+  supports_framebuffer_fetch_ = desc->HasExtension(kFramebufferFetchExt);
 
-  if (gl.GetDescription()->HasExtension(kTextureBorderClampExt) ||
-      gl.GetDescription()->HasExtension(kNvidiaTextureBorderClampExt) ||
-      gl.GetDescription()->HasExtension(kOESTextureBorderClampExt)) {
+  if (desc->HasExtension(kTextureBorderClampExt) ||
+      desc->HasExtension(kNvidiaTextureBorderClampExt) ||
+      desc->HasExtension(kOESTextureBorderClampExt)) {
     supports_decal_sampler_address_mode_ = true;
   }
 
-  if (gl.GetDescription()->HasExtension(
-          "GL_EXT_multisampled_render_to_texture2") &&
-      // The current implementation of MSAA support in Impeller GLES requires
-      // the use of glBlitFramebuffer, which is not available on all GLES
-      // implementations. We can't use MSAA on these platforms yet.
-      gl.BlitFramebuffer.IsAvailable()) {
+  if (desc->HasExtension(kMultisampledRenderToTextureExt)) {
+    supports_implicit_msaa_ = true;
+
     // We hard-code 4x MSAA, so let's make sure it's supported.
     GLint value = 0;
     gl.GetIntegerv(GL_MAX_SAMPLES_EXT, &value);
-
     supports_offscreen_msaa_ = value >= 4;
   }
 }
@@ -140,6 +141,10 @@ bool CapabilitiesGLES::SupportsOffscreenMSAA() const {
   return supports_offscreen_msaa_;
 }
 
+bool CapabilitiesGLES::SupportsImplicitResolvingMSAA() const {
+  return supports_implicit_msaa_;
+}
+
 bool CapabilitiesGLES::SupportsSSBO() const {
   return false;
 }

impeller/renderer/backend/gles/capabilities_gles.h

@@ -76,6 +76,9 @@ class CapabilitiesGLES final
   // |Capabilities|
   bool SupportsOffscreenMSAA() const override;
 
+  // |Capabilities|
+  bool SupportsImplicitResolvingMSAA() const override;
+
   // |Capabilities|
   bool SupportsSSBO() const override;
 
@@ -119,6 +122,7 @@ class CapabilitiesGLES final
   bool supports_framebuffer_fetch_ = false;
   bool supports_decal_sampler_address_mode_ = false;
   bool supports_offscreen_msaa_ = false;
+  bool supports_implicit_msaa_ = false;
 };
 
 }  // namespace impeller

impeller/renderer/backend/gles/render_pass_gles.cc

@@ -4,6 +4,7 @@
 
 #include "impeller/renderer/backend/gles/render_pass_gles.h"
 
+#include "GLES3/gl3.h"
 #include "flutter/fml/trace_event.h"
 #include "fml/closure.h"
 #include "fml/logging.h"
@@ -127,7 +128,6 @@ struct RenderPassData {
   Scalar clear_depth = 1.0;
 
   std::shared_ptr<Texture> color_attachment;
-  std::shared_ptr<Texture> resolve_attachment;
   std::shared_ptr<Texture> depth_attachment;
   std::shared_ptr<Texture> stencil_attachment;
 
@@ -474,52 +474,6 @@ struct RenderPassData {
     }
   }
 
-  // When we have a resolve_attachment, MSAA is being used. We blit from the
-  // MSAA FBO to the resolve FBO, otherwise the resolve FBO ends up being
-  // incomplete (because it has no attachments).
-  //
-  // Note that this only works on OpenGLES 3.0+, or put another way, in older
-  // versions of OpenGLES, MSAA is not currently supported by Impeller. It's
-  // possible to work around this issue a few different ways (not yet done).
-  //
-  // TODO(matanlurey): See https://github.com/flutter/flutter/issues/137093.
-  if (!is_default_fbo && pass_data.resolve_attachment) {
-    // MSAA should not be enabled if BlitFramebuffer is not available.
-    FML_DCHECK(gl.BlitFramebuffer.IsAvailable());
-
-    GLuint draw_fbo = GL_NONE;
-    fml::ScopedCleanupClosure delete_draw_fbo([&gl, &draw_fbo, fbo]() {
-      if (draw_fbo != GL_NONE) {
-        gl.BindFramebuffer(GL_FRAMEBUFFER, fbo);
-        gl.DeleteFramebuffers(1u, &draw_fbo);
-      }
-    });
-
-    gl.GenFramebuffers(1u, &draw_fbo);
-    gl.BindFramebuffer(GL_FRAMEBUFFER, draw_fbo);
-
-    auto resolve = TextureGLES::Cast(pass_data.resolve_attachment.get());
-    if (!resolve->SetAsFramebufferAttachment(
-            GL_FRAMEBUFFER, TextureGLES::AttachmentPoint::kColor0)) {
-      return false;
-    }
-
-    gl.BindFramebuffer(GL_DRAW_FRAMEBUFFER, draw_fbo);
-    gl.BindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
-    auto size = pass_data.resolve_attachment->GetSize();
-    gl.BlitFramebuffer(0,                    // srcX0
-                       0,                    // srcY0
-                       size.width,           // srcX1
-                       size.height,          // srcY1
-                       0,                    // dstX0
-                       0,                    // dstY0
-                       size.width,           // dstX1
-                       size.height,          // dstY1
-                       GL_COLOR_BUFFER_BIT,  // mask
-                       GL_NEAREST            // filter
-    );
-  }
-
   if (gl.DiscardFramebufferEXT.IsAvailable()) {
     std::vector<GLenum> attachments;
 
@@ -576,12 +530,19 @@ bool RenderPassGLES::OnEncodeCommands(const Context& context) const {
   /// Setup color data.
   ///
   pass_data->color_attachment = color0.texture;
-  pass_data->resolve_attachment = color0.resolve_texture;
   pass_data->clear_color = color0.clear_color;
   pass_data->clear_color_attachment = CanClearAttachment(color0.load_action);
   pass_data->discard_color_attachment =
       CanDiscardAttachmentWhenDone(color0.store_action);
 
+  // When we are using EXT_multisampled_render_to_texture, it is implicitly
+  // resolved when we bind the texture to the framebuffer. We don't need to
+  // discard the attachment when we are done.
+  if (color0.resolve_texture) {
+    FML_DCHECK(context.GetCapabilities()->SupportsImplicitResolvingMSAA());
+    pass_data->discard_color_attachment = false;
+  }
+
   //----------------------------------------------------------------------------
   /// Setup depth data.
   ///

impeller/renderer/backend/vulkan/capabilities_vk.cc

@@ -409,6 +409,11 @@ bool CapabilitiesVK::SupportsOffscreenMSAA() const {
   return true;
 }
 
+// |Capabilities|
+bool CapabilitiesVK::SupportsImplicitResolvingMSAA() const {
+  return false;
+}
+
 // |Capabilities|
 bool CapabilitiesVK::SupportsSSBO() const {
   return true;

impeller/renderer/backend/vulkan/capabilities_vk.h

@@ -60,6 +60,9 @@ class CapabilitiesVK final : public Capabilities,
   // |Capabilities|
   bool SupportsOffscreenMSAA() const override;
 
+  // |Capabilities|
+  bool SupportsImplicitResolvingMSAA() const override;
+
   // |Capabilities|
   bool SupportsSSBO() const override;
 

impeller/renderer/capabilities.cc

@@ -20,6 +20,9 @@ class StandardCapabilities final : public Capabilities {
     return supports_offscreen_msaa_;
   }
 
+  // |Capabilities|
+  bool SupportsImplicitResolvingMSAA() const override { return false; }
+
   // |Capabilities|
   bool SupportsSSBO() const override { return supports_ssbo_; }
 

impeller/renderer/capabilities.h

@@ -19,6 +19,11 @@ class Capabilities {
   ///         color/stencil textures.
   virtual bool SupportsOffscreenMSAA() const = 0;
 
+  /// @brief  Whether the context backend supports multisampled rendering to
+  ///         the on-screen surface without requiring an explicit resolve of
+  ///         the MSAA color attachment.
+  virtual bool SupportsImplicitResolvingMSAA() const = 0;
+
   /// @brief  Whether the context backend supports binding Shader Storage Buffer
   ///         Objects (SSBOs) to pipelines.
   virtual bool SupportsSSBO() const = 0;

impeller/renderer/render_target.cc

@@ -286,6 +286,11 @@ RenderTarget RenderTarget::CreateOffscreenMSAA(
   color0_tex_desc.size = size;
   color0_tex_desc.usage = static_cast<uint64_t>(TextureUsage::kRenderTarget);
 
+  if (context.GetCapabilities()->SupportsImplicitResolvingMSAA()) {
+    // See below ("SupportsImplicitResolvingMSAA") for more details.
+    color0_tex_desc.storage_mode = StorageMode::kDevicePrivate;
+  }
+
   auto color0_msaa_tex = allocator.CreateTexture(color0_tex_desc);
   if (!color0_msaa_tex) {
     VALIDATION_LOG << "Could not create multisample color texture.";
@@ -322,6 +327,17 @@ RenderTarget RenderTarget::CreateOffscreenMSAA(
   color0.texture = color0_msaa_tex;
   color0.resolve_texture = color0_resolve_tex;
 
+  if (context.GetCapabilities()->SupportsImplicitResolvingMSAA()) {
+    // If implicit MSAA is supported, then the resolve texture is not needed
+    // because the multisample texture is automatically resolved. We instead
+    // provide a view of the multisample texture as the resolve texture (because
+    // the HAL does expect a resolve texture).
+    //
+    // In practice, this is used for GLES 2.0 EXT_multisampled_render_to_texture
+    // https://registry.khronos.org/OpenGL/extensions/EXT/EXT_multisampled_render_to_texture.txt
+    color0.resolve_texture = color0_msaa_tex;
+  }
+
   target.SetColorAttachment(color0, 0u);
 
   // Create MSAA stencil texture.