xref: /third_party/vk-gl-cts/doc/testspecs/GLES3/performance.buffer.data_upload.txt

g-------------------------------------------------------------------------
drawElements Quality Program Test Specification
-----------------------------------------------

Copyright 2014 The Android Open Source Project

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

     http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-------------------------------------------------------------------------
    Buffer upload performance tests

Tests:
 + dEQP-GLES3.performance.buffer.data_upload.*

Includes:
 + Memcpy performance tests for reference values
 + Performance of glBufferData and glBufferSubData
    - STREAM, STATIC and DYNAMIC buffer usages
    - Target buffer in different states:
        * BufferData:
            ~ Just generated buffer
            ~ Generated buffer with unspecified contents (bufferData(..., NULL))
            ~ Generated buffer with specified contents
            ~ Buffer (with the same size) that has been used in drawing earlier
            ~ Buffer (with a larger size) that has been used in drawing earlier
        * BufferSubData:
            ~ Respecify whole buffer contents
            ~ Respecify half of the buffer
            ~ Invalidate buffer contents manually (bufferData(..., NULL)) before
              calling bufferSubData
 + Buffer mapping performance
    - STREAM, STATIC and DYNAMIC buffer usages
    - Full and partial (half of the buffer) mapping
    - Different map bits:
        * Write
        * Write-Read
        * Write-InvalidateRange
        * Write-InvalidateBuffer
        * Write-Unsynchronized
        * Write-Unsynchronized-InvalidateBuffer
        * Write-FlushExplicit
    - Target buffer in different states:
        * Generated buffer with unspecified contents (bufferData(..., NULL))
        * Generated buffer with specified contents
        * A buffer (with the same size) that has been used in drawing earlier
    - Different mapping & flushing strategies:
        * Map whole buffer and flush whole mapped area (in one call)
        * Map buffer partially (half) and flush whole mapped area (in one call)
        * Map whole buffer and flush buffer partially (half) (in one call)
        * Map whole buffer and flush whole mapped area in 4 non-interleaving
          flush calls
 + Buffer upload performance after a draw call using the target buffer
    - STREAM, STATIC and DYNAMIC buffer usages
    - BufferData, BufferSubData, MapBufferRange methods
 + Uploaded buffer rendering performance tests

Excludes:
 + Exhaustive testing of all combinations of usages, buffer states, and methods
 + Complex buffer usage before upload (e.g transform feedback, copyBufferSubData)
 + Sustained upload performance tests

Description:

Upload performance tests approximate upload speed by testing time usage of different
buffer upload functions with multiple different buffer sizes. The time usage of these
functions is expected to be of form t(x) = b * x + c, where x is the buffer size, b
is the linear cost of the function and c is the constant cost of the function. Tests
find suitable parameters b and c to fit the sampled values and then report these
parameters. The test result of an upload test is the median transfer rate of the test
samples unless stated otherwise.

Additionally, test cases calculate "sample linearity" and "sample temporal stability"
metrics. Sample linearity describes the curvature of the sampled values and it is
calculated by comparing the line fitted to the first half of the samples (the half with
smaller buffer sizes) to the line fitted to the second half of the samples. A straight
line would result in 100% linearity. A linearity value less than 90% usually implies
either large errors in sampling (noise), insufficient timer resolution, or non-linearity
of the function.

Sample temporal stability describes the stability of the function during the sampling.
Samples are split to two categories: early and late samples. All early samples are tested
before any late sample. A line is fitted to the early samples and it is compared to a
line fitted to the late samples. If the samples get similar values in the beginning of
the test and in the end of the test, the temporal stability will be 100%. A value less
than 90% usually implies that test samples are not independent from other samples, or
that there were large errors in sampling (noise).

(Note that all logged data sizes are MiBs (mebibytes) even if they are marked with unit
symbol "MB".)

reference.* tests test memcpy with different sizes. The transfer rate may be useful as a
reference value. "small_buffers" test case tests memcpy performance with transfer size
in range 0 - 256KiB. The result is the median transfer rate of the test samples.
"large_buffers" test case measures the transfer rate with the transfer size in range
256KiB - 16 MiB. The result is the approximated transfer rate of an infinitely large
transfer size calculated form the fitted line. (i.e c = 0).

function_call.* tests measure the time taken by the corresponding command or the series of
commands with different buffer sizes. The result is the median transfer rate of the test
samples.

modify_after_use.* tests are similar to the function_call.* test but they target a buffer
that has been used in a draw operation just before the measured function call. "_repeated"
variants repeat the "specify buffer contents, draw buffer" sequence a few times before
the test target function is called.

render_after_upload.* tests measure rendering time of a buffer in different scenarios. The
tests measure time from rendering command issuance (glDraw* call) to finished rendering
(return from glReadPixels call). The purpose of the tests is to measure the data transfer
performance from the observation of a buffer upload completion (a return from a
glBuffer(Sub)Data / glUnmap) to observation of finished rendering. This combined rate of
data transfer and rendering is in these tests referred to as "processing rate". Note that
"processing time" does not include the upload time. All tests and test samples have the
same number of visible fragments but may have different number of vertices. The tests are
NOT rendering performance tests and should not be treated as such.

Each case's draw function, uploaded buffer type, and upload method is encoded to the case
name. For example "draw_elements_upload_indices_with_buffer_data" would specify index
(GL_ELEMENT_ARRAY_BUFFER) buffer contents with glBufferData, and issue render with
glDrawElements call.

render_after_upload.reference.draw.* tests produce baseline results for basic functions.
"read_pixels" case measures the constant cost (in microseconds) of a single glReadPixels
call with identical arguments as in other rendering tests. The result might used for
excluding ReadPixels' constant cost from the total constant costs in the other test
results. "draw_arrays" and "draw_elements" tests measure the processing rate in the case
where there is no buffer uploading required. As such, they are expected to result in the
highest processing rates. The results should be interpreted as a reference values for
render_after_upload.upload_and_draw.* group tests.

render_after_upload.reference.draw_upload_draw.* cases first render from an existing
buffer, upload another buffer, and then render using the just uploaded buffer. The
processing rate (excluding the upload time) is measured. The results are baseline
reference results for render_after_upload.draw_modify_draw.* group tests.

render_after_upload.upload_unrelated_and_draw.* tests upload an unrelated buffer and
then render to the viewport. The rendering does not depend on the uploaded buffer. The
unrelated upload size is equal to the rendering workload size.

render_after_upload.upload_and_draw.* cases upload a buffer and then render using it. In
used_buffer.* group cases, the buffer has been used in rendering before upload. In
new_buffer.* group cases, the uploaded buffer is generated just before uploading. In this
case if the upload method is not glBufferData, glBufferData is called with NULL data
argument before using the actual upload method for specifying the buffer contents. In
cases in test case groups "*_and_unrelated_upload", another buffer is uploaded after the
target buffer upload with glBufferData. The rendering does not depend on this second
upload. The unrelated upload size is equal to the rendering workload size.

render_after_upload.draw_modify_draw.* cases draw from a buffer, respecify its contents,
and draw again using the buffer. Cases are identical to
render_after_upload.reference.draw_upload_draw.* cases with the exception that the upload
target buffer is used in rendering just before issuing upload commands.

render_after_upload.upload_wait_draw.* cases upload a buffer, wait a certain number of
frames (buffer swaps), and then render using the buffer. The number of frames required
for the render times to stabilize is measured and reported as a test result. For example,
result value of 0 would indicate that rendering is just as fast just after upload than
N frames later. A value of 1 would mean that performance of rendering from the buffer
is different during the frame it was uploaded in, but does not change after one frame
swap.

Since the frame times during the "waiting" period may be significantly lower than in the
upload frame or in the frame of the final rendering, in order to keep frame times
consistent, the test code performs dummy calculations (busy waits) for 8 milliseconds in
each frame. Also during the "dummy" frames, a simple grid pattern is drawn to the viewport.