cassdlcm/mediapipe
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

F16 texture support in GPU multiclass TensorsToSegmentation

Browse Source 
PiperOrigin-RevId: 523512145
This commit is contained in:
MediaPipe Team 2023-04-11 14:56:36 -07:00 committed by Copybara-Service
parent 738aea7a06
commit e2b13523f1
1 changed files with 32 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
mediapipe/tasks/cc/vision/image_segmenter/calculators/segmentation_postprocessor_gl.cc
View File

@ -469,16 +469,37 @@ SegmentationPostprocessorGl::GetSegmentationResultGpu(const Shape& input_shape,
bool is_softmax = bool is_softmax =
options_.segmenter_options().activation() == SegmenterOptions::SOFTMAX; options_.segmenter_options().activation() == SegmenterOptions::SOFTMAX;
// To make logic easier for now, we use F32 only if we have all three of the
// following features available for it:
// (1) color rendering
// (2) linear filtering
// (3) blending
// Otherwise, we just try for F16. See b/277656755 for more information.
// TODO: In the future, separate these 3 different restrictions.
// TODO: Also, we should extend this logic to non-web platforms.
static bool can_use_f32 =
helper_.GetGlContext().HasGlExtension("EXT_color_buffer_float") &&
helper_.GetGlContext().HasGlExtension("OES_texture_float_linear") &&
helper_.GetGlContext().HasGlExtension("EXT_float_blend");
static bool can_use_f16_backup =
helper_.GetGlContext().HasGlExtension("EXT_color_buffer_half_float");
RET_CHECK(can_use_f32 || can_use_f16_backup)
<< "Segmentation postprocessing error: GPU does not fully support "
<< "4-channel float32 or float16 formats.";
const GpuBufferFormat activation_output_format = const GpuBufferFormat activation_output_format =
GpuBufferFormat::kRGBAFloat128; can_use_f32 ? GpuBufferFormat::kRGBAFloat128
const GpuBufferFormat chunk_output_format = GpuBufferFormat::kRGBAFloat128; : GpuBufferFormat::kRGBAHalf64;
const GpuBufferFormat chunk_output_format =
can_use_f32 ? GpuBufferFormat::kRGBAFloat128
: GpuBufferFormat::kRGBAHalf64;
// Uint8 pipeline and conversions are lacking, so for now we just use F32 // Uint8 pipeline and conversions are lacking, so for now we just use F32
// textures even for category masks. // textures even for category masks.
// TODO: Also, some platforms (like certain iOS devices) do not const GpuBufferFormat final_output_format =
// allow for rendering to RGBAF32 textures, so we should switch to using can_use_f32 ? GpuBufferFormat::kGrayFloat32
// F16 textures in those instances. : GpuBufferFormat::kGrayHalf16;
const GpuBufferFormat final_output_format = GpuBufferFormat::kGrayFloat32;
const Tensor::OpenGlTexture2dView read_view = const Tensor::OpenGlTexture2dView read_view =
tensor.GetOpenGlTexture2dReadView(); tensor.GetOpenGlTexture2dReadView();
@ -702,9 +723,11 @@ SegmentationPostprocessorGl::GetSegmentationResultGpu(const Shape& input_shape,
// GLSL uses IEEE 754 single-precision floating-point for encoding its // GLSL uses IEEE 754 single-precision floating-point for encoding its
// floats (at least for number representation, although not necessarily // floats (at least for number representation, although not necessarily
// for operations). So we can clear to a reasonable minimum float value // for operations). So we can clear to a reasonable minimum float value
// accordingly. // accordingly. Min f32 value is -(2 - 2^(-23))*2^127, while min f16
const float kFloatMin32 = -3.402823466e+38; // value is -(2 - 2^(-10))*2^15 = 65504. We use minima sufficiently
glClearColor(kFloatMin32, -1.0, 0.0, 1.0); // close to these for our clear.
const float kFloatMin = can_use_f32 ? -3.402823466e+38 : -65500.0;
glClearColor(kFloatMin, -1.0, 0.0, 1.0);
helper_.BindFramebuffer(max_texture); helper_.BindFramebuffer(max_texture);
glClear(GL_COLOR_BUFFER_BIT); glClear(GL_COLOR_BUFFER_BIT);
// Set our clear color back to a "normal" default. // Set our clear color back to a "normal" default.