Skip to content

A
angle
Commit 7b77dc5b by Charlie Lao Committed by Commit Bot
Options

Vulkan: minimize-gpu-work: Skip data copy when possible

When --minimize-gpu-work is specified while replaying app traces, the goal is to avoid any GPU work when possible and focus on driver cpu logic overhead. Data copy can be lengthy and each driver optimize it differently for some real world usage scenario. This should be looked along with normal app trace playback performance. When --minimize-gpu-work is specified, we want to leave this out of picture. Previously I have fixed TexImage2D by overwriting pixel pointer with null. But there is a hole here when PBO is used. This CL fix the case that when data is sourced from PBO, we ensure to skip data copy as well. This CL also noops TexSubImage call instead of doing 1x1 copy. Again depends on driver implementation, some may use CPU others use GPU which will have different overhead. We can easily write a test to cover these performance optimizations. By skipping the subImage call here we will have less noise to deal with for CPU overhead investigation. Bug: b/184766477 Change-Id: I84a5d26d2f25f8f0a6c5c9da72737906d6356a53 Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2847100 Commit-Queue: Charlie Lao <cclao@google.com> Reviewed-by: 's avatarCody Northrop <cnorthrop@google.com> Reviewed-by: 's avatarJamie Madill <jmadill@chromium.org>
parent 16a919b1
Showing with 39 additions and 0 deletions
src/tests/perf_tests/TracePerfTest.cpp
......@@ -24,6 +24,19 @@
#include <functional>
#include <sstream>
// When --minimize-gpu-work is specified, we want to reduce GPU work to minimum and lift up the CPU
// overhead to surface so that we can see how much CPU overhead each driver has for each app trace.
// On some driver(s) the bufferSubData/texSubImage calls end up dominating the frame time when the
// actual GPU work is minimized. Even reducing the texSubImage calls to only update 1x1 area is not
// enough. The driver may be implementing copy on write by cloning the entire texture to another
// memory storage for texSubImage call. While this information is also important for performance,
// they should be evaluated separately in real app usage scenario, or write stand alone tests for
// these. For the purpose of CPU overhead and avoid data copy to dominate the trace, I am using this
// flag to noop the texSubImage and bufferSubData call when --minimize-gpu-work is specified. Feel
// free to disable this when you have other needs. Or it can be turned to another run time option
// when desired.
#define NOOP_SUBDATA_SUBIMAGE_FOR_MINIMIZE_GPU_WORK
using namespace angle;
using namespace egl_platform;
......@@ -285,7 +298,9 @@ void KHRONOS_APIENTRY BufferSubDataMinimizedProc(GLenum target,
GLsizeiptr size,
const void *data)
{
#if !defined(NOOP_SUBDATA_SUBIMAGE_FOR_MINIMIZE_GPU_WORK)
glBufferSubData(target, offset, 1, data);
#endif
}
void KHRONOS_APIENTRY TexImage2DMinimizedProc(GLenum target,
......@@ -298,7 +313,17 @@ void KHRONOS_APIENTRY TexImage2DMinimizedProc(GLenum target,
GLenum type,
const void *pixels)
{
GLint unpackBuffer = 0;
glGetIntegerv(GL_PIXEL_UNPACK_BUFFER_BINDING, &unpackBuffer);
if (unpackBuffer)
{
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
}
glTexImage2D(target, level, internalformat, width, height, border, format, type, nullptr);
if (unpackBuffer)
{
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, unpackBuffer);
}
}
void KHRONOS_APIENTRY TexSubImage2DMinimizedProc(GLenum target,
......@@ -311,7 +336,9 @@ void KHRONOS_APIENTRY TexSubImage2DMinimizedProc(GLenum target,
GLenum type,
const void *pixels)
{
#if !defined(NOOP_SUBDATA_SUBIMAGE_FOR_MINIMIZE_GPU_WORK)
glTexSubImage2D(target, level, xoffset, yoffset, 1, 1, format, type, pixels);
#endif
}
void KHRONOS_APIENTRY TexImage3DMinimizedProc(GLenum target,
......@@ -325,8 +352,18 @@ void KHRONOS_APIENTRY TexImage3DMinimizedProc(GLenum target,
GLenum type,
const void *pixels)
{
GLint unpackBuffer = 0;
glGetIntegerv(GL_PIXEL_UNPACK_BUFFER_BINDING, &unpackBuffer);
if (unpackBuffer)
{
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
}
glTexImage3D(target, level, internalformat, width, height, depth, border, format, type,
nullptr);
if (unpackBuffer)
{
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, unpackBuffer);
}
}
void KHRONOS_APIENTRY TexSubImage3DMinimizedProc(GLenum target,
......@@ -341,7 +378,9 @@ void KHRONOS_APIENTRY TexSubImage3DMinimizedProc(GLenum target,
GLenum type,
const void *pixels)
{
#if !defined(NOOP_SUBDATA_SUBIMAGE_FOR_MINIMIZE_GPU_WORK)
glTexSubImage3D(target, level, xoffset, yoffset, zoffset, 1, 1, 1, format, type, pixels);
#endif
}
void KHRONOS_APIENTRY GenerateMipmapMinimizedProc(GLenum target)
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment