gpu: Perf test to check costs of texture upload at different stages.

gpu/perftests/texture_upload_perftest.cc

+// Copyright 2015 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 <map>
+#include <vector>
+
+#include "base/memory/ref_counted.h"
+#include "base/memory/scoped_ptr.h"
+#include "base/time/time.h"
+#include "base/timer/elapsed_timer.h"
+#include "testing/gtest/include/gtest/gtest.h"
+#include "testing/perf/perf_test.h"
+#include "ui/gfx/geometry/size.h"
+#include "ui/gl/gl_bindings.h"
+#include "ui/gl/gl_context.h"
+#include "ui/gl/gl_surface.h"
+#include "ui/gl/scoped_make_current.h"
+
+namespace gfx {
+namespace {
+
+const int kUploadPerfWarmupRuns = 10;
+const int kUploadPerfTestRuns = 100;
+
+#define SHADER(Src) #Src
+
+// clang-format off
+const char kVertexShader[] =
+SHADER(
+  attribute vec2 a_position;
+  varying vec2 v_texCoord;
+  void main() {
+    gl_Position = vec4(a_position.x, a_position.y, 0.0, 1.0);
+    v_texCoord = vec2((a_position.x + 1) * 0.5, (a_position.y + 1) * 0.5);
+  }
+);
+const char kFragmentShader[] =
+SHADER(
+  uniform sampler2D a_texture;
+  varying vec2 v_texCoord;
+  void main() {
+    gl_FragColor = texture2D(a_texture, v_texCoord.xy);
+  }
+);
+// clang-format on
+
+void CheckNoGlError() {
+  CHECK_EQ(static_cast<GLenum>(GL_NO_ERROR), glGetError());
+}
+
+// Utility function to compile a shader from a string.
+GLuint LoadShader(const GLenum type, const char* const src) {
+  GLuint shader = 0;
+  shader = glCreateShader(type);
+  CHECK_NE(0u, shader);
+  glShaderSource(shader, 1, &src, NULL);
+  glCompileShader(shader);
+
+  GLint compiled = 0;
+  glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+  if (compiled == 0) {
+    GLint len = 0;
+    glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &len);
+    if (len > 1) {
+      scoped_ptr<char> error_log(new char[len]);
+      glGetShaderInfoLog(shader, len, NULL, error_log.get());
+      LOG(ERROR) << "Error compiling shader: " << error_log.get();
+    }
+  }
+  CHECK_NE(0, compiled);
+  return shader;
+}
+
+void GenerateTextureData(const gfx::Size& size,
+                         const int seed,
+                         std::vector<uint8>* const pixels) {
+  pixels->resize(size.GetArea() * 4);
+  for (int y = 0; y < size.height(); ++y) {
+    for (int x = 0; x < size.width(); ++x) {
+      const size_t offset = (y * size.width() + x) * 4;
+      pixels->at(offset) = (y + seed) % 256;
+      pixels->at(offset + 1) = (x + seed) % 256;
+      pixels->at(offset + 2) = (y + x + seed) % 256;
+      pixels->at(offset + 3) = 255;
+    }
+  }
+}
+
+// PerfTest to check costs of texture upload at different stages
+// on different platforms.
+class TextureUploadPerfTest : public testing::Test {
+ public:
+  TextureUploadPerfTest() : size_(512, 512) {}
+
+  // Overridden from testing::Test
+  void SetUp() override {
+    // Initialize an offscreen surface and a gl context.
+    GLSurface::InitializeOneOff();
+    surface_ = GLSurface::CreateOffscreenGLSurface(size_);
+    gl_context_ =
+        GLContext::CreateGLContext(NULL,  // share_group
+                                   surface_.get(), gfx::PreferIntegratedGpu);
+
+    // Prepare a simple program and a vertex buffer that will be
+    // used to draw a quad on the offscreen surface.
+    ui::ScopedMakeCurrent smc(gl_context_.get(), surface_.get());
+    vertex_shader_ = LoadShader(GL_VERTEX_SHADER, kVertexShader);
+    fragment_shader_ = LoadShader(GL_FRAGMENT_SHADER, kFragmentShader);
+    program_object_ = glCreateProgram();
+    CHECK_NE(0u, program_object_);
+
+    glAttachShader(program_object_, vertex_shader_);
+    glAttachShader(program_object_, fragment_shader_);
+    glBindAttribLocation(program_object_, 0, "a_position");
+    glLinkProgram(program_object_);
+
+    GLint linked = -1;
+    glGetProgramiv(program_object_, GL_LINK_STATUS, &linked);
+    CHECK_NE(0, linked);
+
+    sampler_location_ = glGetUniformLocation(program_object_, "a_texture");
+    CHECK_NE(-1, sampler_location_);
+
+    glGenBuffersARB(1, &vertex_buffer_);
+    CHECK_NE(0u, vertex_buffer_);
+    glBindBuffer(GL_ARRAY_BUFFER, vertex_buffer_);
+    static GLfloat positions[] = {
+        -1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f,
+    };
+    glBufferData(GL_ARRAY_BUFFER, sizeof(positions), positions, GL_STATIC_DRAW);
+    CheckNoGlError();
+  }
+
+  void TearDown() override {
+    ui::ScopedMakeCurrent smc(gl_context_.get(), surface_.get());
+    if (program_object_ != 0) {
+      glDeleteProgram(program_object_);
+    }
+    if (vertex_shader_ != 0) {
+      glDeleteShader(vertex_shader_);
+    }
+    if (fragment_shader_ != 0) {
+      glDeleteShader(fragment_shader_);
+    }
+    if (vertex_buffer_ != 0) {
+      glDeleteShader(vertex_buffer_);
+    }
+
+    gl_context_ = nullptr;
+    surface_ = nullptr;
+  }
+
+ protected:
+  struct Measurement {
+    std::string name;
+    float wall_time;

[reveman, 2015/01/27 22:58:31]

Please use base::TimeDelta.

[Daniele Castagna, 2015/01/28 17:28:39]

Done.

+  };
+  // Upload and draw on the offscren surface.
+  // Return a list of pair. Each pair describe a gl operation and the wall
+  // time elapsed in milliseconds.
+  std::vector<Measurement> UploadAndDraw(const std::vector<uint8>& pixels,
+                                         const GLenum format,
+                                         const GLenum type) {
+    std::vector<Measurement> measurements;
+    ui::ScopedMakeCurrent smc(gl_context_.get(), surface_.get());
+
+    base::ElapsedTimer total_timer;
+    GLuint texture_id = 0;
+
+    base::ElapsedTimer tex_timer;
+    glActiveTexture(GL_TEXTURE0);
+    glGenTextures(1, &texture_id);
+    glBindTexture(GL_TEXTURE_2D, texture_id);
+
+    glTexImage2D(GL_TEXTURE_2D, 0, format, size_.width(), size_.height(), 0,
+                 format, type, &pixels[0]);
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+    CheckNoGlError();
+    measurements.emplace_back(
+        Measurement{"teximage2d", tex_timer.Elapsed().InMillisecondsF()});

[reveman, 2015/01/27 22:58:31]

Please check what part of c++11 is allowed in chromium code.

[Daniele Castagna, 2015/01/28 17:28:38]

Removed everything that is blacklisted. Non-Static Class Member Initializers are
allowed.

+
+    base::ElapsedTimer draw_timer;
+    glUseProgram(program_object_);
+    glUniform1i(sampler_location_, 0);
+
+    glBindBuffer(GL_ARRAY_BUFFER, vertex_buffer_);
+    glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 0, 0);
+    glEnableVertexAttribArray(0);
+
+    glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+    measurements.emplace_back(
+        Measurement{"drawarrays", draw_timer.Elapsed().InMillisecondsF()});
+
+    base::ElapsedTimer finish_timer;
+    glFinish();
+    CheckNoGlError();
+    measurements.emplace_back(
+        Measurement{"finish", finish_timer.Elapsed().InMillisecondsF()});
+    measurements.emplace_back(
+        Measurement{"total", total_timer.Elapsed().InMillisecondsF()});
+
+    glDeleteTextures(1, &texture_id);
+
+    std::vector<uint8> pixels_rendered(size_.GetArea() * 4);
+    glReadPixels(0, 0, size_.width(), size_.height(), GL_RGBA, type,
+                 &pixels_rendered[0]);
+    CheckNoGlError();
+
+    // TODO(dcastagna): don't assume the format of the texture and do
+    // the appropriate format conversion.
+    ASSERT_EQ(GL_RGBA, format);
+    EXPECT_EQ(pixels, pixels_rendered);
+    return measurements;
+  }
+
+  const gfx::Size size_;  // for the offscreen surface and the texture
+  scoped_refptr<GLContext> gl_context_;
+  scoped_refptr<GLSurface> surface_;
+
+  GLuint vertex_shader_ = 0;
+  GLuint fragment_shader_ = 0;
+  GLuint program_object_ = 0;
+  GLint sampler_location_ = -1;
+  GLuint vertex_buffer_ = 0;
+};
+
+// Perf test that generates, uploads and draws a texture on a surface repeatedly
+// and prints out aggregated measurements for all the runs.
+TEST_F(TextureUploadPerfTest, glTexImage2d) {
+  std::vector<uint8> pixels;
+  std::map<std::string, Measurement> measurements;  // indexed by name

[reveman, 2015/01/27 22:58:31]

Vector maybe?

[Daniele Castagna, 2015/01/28 17:28:38]

Done.

+  for (int i = 0; i < kUploadPerfWarmupRuns + kUploadPerfTestRuns; ++i) {
+    GenerateTextureData(size_, i + 1, &pixels);
+    auto run = UploadAndDraw(pixels, GL_RGBA, GL_UNSIGNED_BYTE);
+    if (i >= kUploadPerfWarmupRuns) {
+      for (const Measurement& m : run) {
+        Measurement& e = measurements[m.name];
+        e.name = m.name;
+        e.wall_time += m.wall_time;
+      }
+    }
+  }
+
+  for (const auto& e : measurements) {
+    const Measurement& m = e.second;
+    perf_test::PrintResult(m.name, "", "", m.wall_time / kUploadPerfTestRuns,
+                           "ms", true);
+  }
+}
+
+}  // namespace
+}  // namespace cc