Fix unpacking overlapping unpack buffer rows on NVIDIA GL

gpu/command_buffer/service/texture_manager.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "gpu/command_buffer/service/texture_manager.h"
 
 #include <stddef.h>
 #include <stdint.h>
 
 #include <algorithm>
@@ skipping 2376 matching lines @@
     memset(zero.get(), 0, args.pixels_size);
     for (GLenum face : undefined_faces) {
       new_args.target = face;
       new_args.pixels = zero.get();
       DoTexImage(texture_state, state, framebuffer_state,
                  function_name, texture_ref, new_args);
       texture->MarkLevelAsInternalWorkaround(face, args.level);
     }
   }
 
+  if (texture_state->unpack_overlapping_rows_separately_unpack_buffer &&
+      buffer) {
+    ContextState::Dimension dimension =
+        (args.command_type == DoTexImageArguments::kTexImage3D)
+            ? ContextState::k3D
+            : ContextState::k2D;
+    const PixelStoreParams unpack_params(state->GetUnpackParams(dimension));
+    if (unpack_params.row_length != 0 &&
+        unpack_params.row_length < args.width) {
+      // The rows overlap in unpack memory. Upload the texture row by row to
+      // work around driver bug.
+
+      // Reserve memory for the texture and set its attributes so it can be
+      // filled with TexSubImage.

[oetuaho-nv, 2016/05/09 14:22:30]

This part here could maybe be refactored into a common helper function.

+      glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+      state->SetBoundBuffer(GL_PIXEL_UNPACK_BUFFER, nullptr);
+      DoTexImageArguments new_args = args;
+      new_args.pixels = nullptr;
+      // pixels_size might be incorrect, but it's not used in this case.
+      DoTexImage(texture_state, state, framebuffer_state, function_name,
+                 texture_ref, new_args);
+      glBindBuffer(GL_PIXEL_UNPACK_BUFFER, buffer->service_id());
+      state->SetBoundBuffer(GL_PIXEL_UNPACK_BUFFER, buffer);
+
+      DoTexSubImageArguments sub_args = {
+          args.target, args.level, 0, 0, 0, args.width, args.height, args.depth,
+          args.format, args.type, args.pixels, args.pixels_size, args.padding,
+          args.command_type == DoTexImageArguments::kTexImage3D
+              ? DoTexSubImageArguments::kTexSubImage3D
+              : DoTexSubImageArguments::kTexSubImage2D};
+      DoTexSubImageRowByRowWorkaround(texture_state, state, sub_args,
+                                      unpack_params);
+
+      SetLevelCleared(texture_ref, args.target, args.level, true);
+      return;
+    }
+  }
+
   if (texture_state->unpack_alignment_workaround_with_unpack_buffer && buffer) {
     uint32_t buffer_size = static_cast<uint32_t>(buffer->size());
     if (buffer_size - args.pixels_size - ToGLuint(args.pixels) < args.padding) {
       // In ValidateTexImage(), we already made sure buffer size is no less
       // than offset + pixels_size.
       glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
       state->SetBoundBuffer(GL_PIXEL_UNPACK_BUFFER, nullptr);
       DoTexImageArguments new_args = args;
       new_args.pixels = nullptr;
       // pixels_size might be incorrect, but it's not used in this case.
@@ skipping 158 matching lines @@
         return;
       }
     }
     full_image = false;
   } else {
     SetLevelCleared(texture_ref, args.target, args.level, true);
     full_image = true;
   }
 
   Buffer* buffer = state->bound_pixel_unpack_buffer.get();
+
+  if (texture_state->unpack_overlapping_rows_separately_unpack_buffer &&
+      buffer) {
+    ContextState::Dimension dimension =
+        (args.command_type == DoTexSubImageArguments::kTexSubImage3D)
+            ? ContextState::k3D
+            : ContextState::k2D;
+    const PixelStoreParams unpack_params(state->GetUnpackParams(dimension));
+    if (unpack_params.row_length != 0 &&
+        unpack_params.row_length < args.width) {
+      // The rows overlap in unpack memory. Upload the texture row by row to
+      // work around driver bug.
+      DoTexSubImageRowByRowWorkaround(texture_state, state, args,
+                                      unpack_params);
+      return;
+    }
+  }
+
   if (texture_state->unpack_alignment_workaround_with_unpack_buffer && buffer) {
     uint32_t buffer_size = static_cast<uint32_t>(buffer->size());
     if (buffer_size - args.pixels_size - ToGLuint(args.pixels) < args.padding) {
       DoTexSubImageWithAlignmentWorkaround(texture_state, state, args);
       return;
     }
   }
 
   if (full_image && !texture_state->texsubimage_faster_than_teximage &&
       !texture->IsImmutable() && !texture->HasImages()) {
@@ skipping 127 matching lines @@
                                           args.format, args.type,
                                           params,
                                           &size,
                                           nullptr, nullptr, nullptr, nullptr);
       offset += size;
     }
   }
   DCHECK_EQ(ToGLuint(args.pixels) + args.pixels_size, offset);
 }
 
+void TextureManager::DoTexSubImageRowByRowWorkaround(
+    DecoderTextureState* texture_state,
+    ContextState* state,
+    const DoTexSubImageArguments& args,
+    const PixelStoreParams& unpack_params) {
+  glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
+  glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+  DCHECK_EQ(0, state->unpack_skip_pixels);
+  DCHECK_EQ(0, state->unpack_skip_rows);
+  DCHECK_EQ(0, state->unpack_skip_images);
+
+  GLenum format = AdjustTexFormat(args.format);
+
+  GLsizei row_bytes = unpack_params.row_length *
+                      GLES2Util::ComputeImageGroupSize(format, args.type);
+  GLsizei alignment_diff = row_bytes % unpack_params.alignment;
+  if (alignment_diff != 0) {
+    row_bytes += unpack_params.alignment - alignment_diff;
+  }
+  DCHECK_EQ(0, row_bytes % unpack_params.alignment);
+  if (args.command_type == DoTexSubImageArguments::kTexSubImage3D) {
+    GLsizei image_height = args.height;
+    if (unpack_params.image_height != 0) {
+      image_height = unpack_params.image_height;
+    }
+    GLsizei image_bytes = row_bytes * image_height;
+    for (GLsizei image = 0; image < args.depth; ++image) {
+      GLsizei image_byte_offset = image * image_bytes;
+      for (GLsizei row = 0; row < args.height; ++row) {
+        GLsizei byte_offset = image_byte_offset + row * row_bytes;
+        const GLubyte* row_pixels =
+            reinterpret_cast<const GLubyte*>(args.pixels) + byte_offset;
+        glTexSubImage3D(args.target, args.level, args.xoffset,
+                        row + args.yoffset, image + args.zoffset, args.width, 1,
+                        1, format, args.type, row_pixels);
+      }
+    }
+  } else {
+    for (GLsizei row = 0; row < args.height; ++row) {
+      GLsizei byte_offset = row * row_bytes;
+      const GLubyte* row_pixels =
+          reinterpret_cast<const GLubyte*>(args.pixels) + byte_offset;
+      glTexSubImage2D(args.target, args.level, args.xoffset, row + args.yoffset,
+                      args.width, 1, format, args.type, row_pixels);
+    }
+  }
+
+  // Restore unpack state
+  glPixelStorei(GL_UNPACK_ALIGNMENT, unpack_params.alignment);
+  glPixelStorei(GL_UNPACK_ROW_LENGTH, unpack_params.row_length);
+}
+
 GLenum TextureManager::AdjustTexInternalFormat(GLenum format) const {
   if (feature_info_->gl_version_info().is_desktop_core_profile) {
     const Texture::CompatibilitySwizzle* swizzle =
         GetCompatibilitySwizzle(format);
     if (swizzle)
       return swizzle->dest_format;
   }
   return format;
 }
 
@@ skipping 466 matching lines @@
 uint32_t TextureManager::GetServiceIdGeneration() const {
   return current_service_id_generation_;
 }
 
 void TextureManager::IncrementServiceIdGeneration() {
   current_service_id_generation_++;
 }
 
 }  // namespace gles2
 }  // namespace gpu