Add initial support for NV_path_rendering extension to gpu command buffer

gpu/command_buffer/service/gles2_cmd_decoder.cc

Index: gpu/command_buffer/service/gles2_cmd_decoder.cc
diff --git a/gpu/command_buffer/service/gles2_cmd_decoder.cc b/gpu/command_buffer/service/gles2_cmd_decoder.cc
index 80ca20992977ab86eee0557860ea32b402cba336..d00237fd73c87877dbaf77de90c946a23163ae5c 100644
--- a/gpu/command_buffer/service/gles2_cmd_decoder.cc
+++ b/gpu/command_buffer/service/gles2_cmd_decoder.cc
@@ -960,6 +960,9 @@ class GLES2DecoderImpl : public GLES2Decoder,
 
   void DoDrawBuffersEXT(GLsizei count, const GLenum* bufs);
 
+  void DoMatrixLoadfCHROMIUM(GLenum matrixMode, const GLfloat* matrix);
+  void DoMatrixLoadIdentityCHROMIUM(GLenum matrixMode);
+
   // Creates a Program for the given program.
   Program* CreateProgram(
       GLuint client_id, GLuint service_id) {
@@ -10401,6 +10404,36 @@ void GLES2DecoderImpl::DoDrawBuffersEXT(
   }
 }
 
+void GLES2DecoderImpl::DoMatrixLoadfCHROMIUM(GLenum matrixMode,
+                                             const GLfloat* matrix) {
+  DCHECK(matrixMode == GL_PROJECTION_CHROMIUM
+         || matrixMode == GL_MODELVIEW_CHROMIUM);

[vmiura, 2014/04/15 18:02:43]

Instead of DCHECK, set GL error & return.

LOCAL_SET_GL_ERROR(
        GL_INVALID_ENUM,
        "DoMatrixLoadfCHROMIUM", "invalid enum");
    return;

[piman, 2014/04/15 22:30:22]

The check should already be done by the autogen'd code (see e.g.
GLES2DecoderImpl::HandleMatrixLoadfCHROMIUM in gles2_cmd_decoder_autogen.h)

However we need to raise GL_INVALID_OPERATION if the extension isn't present
(otherwise we will call a NULL pointer).
(same below)

[Kimmo Kinnunen, 2014/04/23 12:26:55]

Done.

[Kimmo Kinnunen, 2014/04/23 12:26:55]

DCHECK is for consistency with other code (which dchecks) and for the reader.

[vmiura, 2014/04/23 18:00:06]

Sorry, my error.

+
+  GLfloat* target_matrix = matrixMode == GL_PROJECTION_CHROMIUM ?
+    state_.projection_matrix : state_.modelview_matrix;
+  if (memcmp(target_matrix, matrix, sizeof(GLfloat) * 16)) {

[piman, 2014/04/15 22:30:22]

Is it worth doing this comparison? We're walking 64 bytes to possibly avoid a
function call?
(same below).

[Kimmo Kinnunen, 2014/04/23 12:26:55]

It avoids function call and a memcpy. The memcpy walks the same 64 bytes, so
they should be in cache and thus be a bit faster the second time? Due to
wrappers, GL function call is ~3 function calls, esp. on Android. Not that
either of these answer the question.

I just followed the existing code, which always checks for change before a call.
Your call, I can also remove the check?

[vmiura, 2014/04/23 18:00:06]

GL calls are heavy (multiple calls, pthread_get_local, etc.), but how often do
we have repeated matrices?  I don't mind keeping the check and tune when we have
some data.

[piman, 2014/04/24 19:30:07]

Right, you're really comparing:

(no test): 1 memcpy, 1 gl call

vs 

(test, hits): 1 memcmp (you only saved the gl call, memcmp is comparable in cost
to memcpy)
(test, misses): 1 memcmp + 1 memcpy + 1 gl call (additional cost of the memcmp)

So we're trading the cost of the extra memcmp on a miss for the cost of the gl
call that you save on a hit.
I suspect you'll miss way more often than you'll hit, that's why it doesn't seem
worth it (unless you think otherwise?)
If we have redundancy, it'd be better to catch it on the client side so that we
avoid the commands, taking space in the command buffer, etc. on top of avoiding
the GL call.

+    memcpy(target_matrix, matrix, sizeof(GLfloat) * 16);
+    glMatrixLoadfEXT(matrixMode, target_matrix);
+  }
+}
+
+void GLES2DecoderImpl::DoMatrixLoadIdentityCHROMIUM(GLenum matrixMode) {
+  DCHECK(matrixMode == GL_PROJECTION_CHROMIUM
+         || matrixMode == GL_MODELVIEW_CHROMIUM);

[vmiura, 2014/04/15 18:02:43]

Same, LOCAL_SET_GL_ERROR instead of DCHECK.

[Kimmo Kinnunen, 2014/04/23 12:26:55]

DCHECK is just a doc for the reader..

+  static GLfloat identity_matrix[16] = {

[piman, 2014/04/15 22:30:22]

nit: const, and use kIdentityMatrix naming convention.

[Kimmo Kinnunen, 2014/04/23 12:26:55]

Done.

+    1.0f, 0.0f, 0.0f, 0.0f,
+    0.0f, 1.0f, 0.0f, 0.0f,
+    0.0f, 0.0f, 1.0f, 0.0f,
+    0.0f, 0.0f, 0.0f, 1.0f
+  };
+  GLfloat* target_matrix = matrixMode == GL_PROJECTION_CHROMIUM ?
+    state_.projection_matrix : state_.modelview_matrix;
+  if (memcmp(target_matrix, identity_matrix, sizeof(identity_matrix))) {
+    memcpy(target_matrix, identity_matrix, sizeof(identity_matrix));
+    glMatrixLoadIdentityEXT(matrixMode);
+  }
+}
+
 bool GLES2DecoderImpl::ValidateAsyncTransfer(
     const char* function_name,
     TextureRef* texture_ref,