Merge branch 'dev' of /mnt/c/Users/geekerwan/AndroidStudioProjects/FoldCraftLauncher/MobileGL into gl43-claim

This commit is contained in:
2026-08-22 11:42:50 -04:00
6 changed files with 2218 additions and 182 deletions
@@ -10,10 +10,21 @@
#include <MG_Util/SelfTest/DriverBugProbes.h>
#include <algorithm>
#include <cstdlib>
#include <cstring>
#include <map>
#include <string>
#include <vector>
using namespace MobileGL;
using MobileGL::MG_Util::SelfTest::CollectGlesKnownDriverBugs;
using MobileGL::MG_Util::SelfTest::DriverBugVerdict;
using MobileGL::MG_Util::SelfTest::ProbeCrossStageImageQualifierMergeDropsWrites;
using MobileGL::MG_Util::SelfTest::ProbeGeometryStageSsboWriteAfterEmitDropped;
using MobileGL::MG_Util::SelfTest::ProbeImageLocationPerNameBudget;
using MobileGL::MG_Util::SelfTest::ProbeImageWriteReadCoherencyResidual;
using MobileGL::MG_Util::SelfTest::ProbeR32FMultisampleSwizzleCorruption;
namespace {
// A driver table with nothing resolved. Every probe has to treat this as "cannot tell",
@@ -21,6 +32,418 @@ namespace {
MG_External::GLESFunctionsTable EmptyFunctionTable() {
return MG_External::GLESFunctionsTable{};
}
// ===================== THE FAKE DRIVER =====================
//
// Same idea as the fake GLES table BackendLoaderTest drives the gl_InstanceID probe with:
// captureless lambdas over one file-scope state, with per-test knobs that turn each defect
// on and off. It is deliberately a MODEL of the defect rather than a canned answer - the
// fake reads the shader text the probe actually submitted and reproduces what the affected
// driver does with it, so a probe that stopped building the triggering shape would stop
// detecting, which is exactly what these tests are for.
//
// These tests call the Probe* functions directly rather than through
// CollectGlesKnownDriverBugs(): the collector goes through the once-per-process memos, and a
// memo latched by one test would decide the answer for every later one.
// The exact text an affected Adreno driver puts in the info log for this refusal.
const char* const kImageLocationLinkLog =
"Error: Image Image location or component exceeds max allowed.\nError: Linking failed.";
struct FakeDriver {
// ---- limits the probes gate on -------------------------------------
GLint maxColorTextureSamples = 4;
GLint maxImageUnits = 8;
GLint maxVertexImageUniforms = 8;
GLint maxFragmentImageUniforms = 8;
GLint maxGeometryImageUniforms = 3;
// The landed geometry probe reads this; zero keeps it inert so it cannot interfere.
GLint maxGeometrySsboBlocks = 0;
bool geometryImageLimitQueryRaisesError = false;
bool colorTextureSamplesQueryRaisesError = false;
// ---- defect knobs ---------------------------------------------------
// Probe 1: a swizzled-alpha, non-zero-sample .w fetch reads garbage from the second
// sampling program onward.
bool msaaSwizzledAlphaCorrupted = false;
// Probe 1's inconclusive path: EVERY sampled read is wrong, including the controls.
bool msaaEveryReadWrong = false;
// Probe 2: the link fails once the program declares more distinct image uniform NAMES
// than this.
int distinctImageNameBudget = 1000;
// Probe 3: a same-name coherent writeonly/readonly pair loses the writing stage's store.
bool sameNameImagePairDropsWrites = false;
// Probe 3's inconclusive path: the renamed control loses it too.
bool everyVertexImageWriteDropped = false;
// Probe 4: how many texels the in-invocation dependent read misses under the STRONGEST
// shape, how many it misses under the shape MobileGL emits today, and whether the
// two-draw control misses them too.
int coherencyStrongestShapeFailedTexels = 0;
int coherencyEmittedShapeFailedTexels = 0;
int coherencyControlFailedTexels = 0;
// ---- object bookkeeping ---------------------------------------------
GLenum pendingError = GL_NO_ERROR;
GLuint nextShaderId = 1;
GLuint nextProgramId = 1;
GLuint nextTextureId = 1;
GLuint nextFramebufferId = 1;
GLuint nextVertexArrayId = 1;
int aliveShaders = 0;
int alivePrograms = 0;
int aliveTextures = 0;
int aliveFramebuffers = 0;
int aliveVertexArrays = 0;
std::map<GLuint, std::string> shaderSources;
std::map<GLuint, std::vector<GLuint>> programShaders;
std::map<GLuint, bool> programLinked;
std::map<GLuint, std::string> programInfoLogs;
// texture id -> GL_TEXTURE_SWIZZLE_A
std::map<GLuint, GLenum> multisampleAlphaSwizzle;
GLuint boundMultisampleTexture = 0;
GLuint currentProgram = 0;
// How many programs that sample a multisample texture have been linked so far. The
// corruption starts at the second.
int sampledMultisampleProgramCount = 0;
// Set by glDrawArrays, consumed by glReadPixels.
GLfloat lastSampledValue = 1.0f;
int lastFailedTexelCount = 0;
};
FakeDriver g_fake;
void ResetFakeDriver() { g_fake = FakeDriver{}; }
const std::string& SourceOf(GLuint shader) {
static const std::string empty;
const auto it = g_fake.shaderSources.find(shader);
return it == g_fake.shaderSources.end() ? empty : it->second;
}
bool Contains(const std::string& haystack, const char* needle) {
return haystack.find(needle) != std::string::npos;
}
// Every `image2D <name>` the program declares, across all its stages.
std::vector<std::string> DeclaredImageNames(GLuint program) {
std::vector<std::string> names;
const auto attached = g_fake.programShaders.find(program);
if (attached == g_fake.programShaders.end()) return names;
for (const GLuint shader : attached->second) {
const std::string& source = SourceOf(shader);
std::size_t at = 0;
while ((at = source.find("image2D ", at)) != std::string::npos) {
at += std::strlen("image2D ");
const std::size_t end = source.find_first_of(";,)", at);
if (end == std::string::npos) break;
std::string name = source.substr(at, end - at);
while (!name.empty() && (name.back() == ' ' || name.back() == '\t')) name.pop_back();
if (std::find(names.begin(), names.end(), name) == names.end()) {
names.push_back(name);
}
at = end;
}
}
return names;
}
std::string StageSourceContaining(GLuint program, const char* needle) {
const auto attached = g_fake.programShaders.find(program);
if (attached == g_fake.programShaders.end()) return {};
for (const GLuint shader : attached->second) {
const std::string& source = SourceOf(shader);
if (Contains(source, needle)) return source;
}
return {};
}
// The uniform name in `... image2D <name>;` of the first declaration in `source`.
std::string FirstImageNameIn(const std::string& source) {
const std::size_t at = source.find("image2D ");
if (at == std::string::npos) return {};
const std::size_t start = at + std::strlen("image2D ");
const std::size_t end = source.find(';', start);
if (end == std::string::npos) return {};
return source.substr(start, end - start);
}
// Whatever the sampling vertex shader asked for: `texelFetch(mg_probeSampler, ivec2(0), N).C`.
void ParseSampledFetch(const std::string& source, int& sampleIndex, char& component) {
sampleIndex = -1;
component = '?';
const std::size_t at = source.find("texelFetch(mg_probeSampler, ivec2(0), ");
if (at == std::string::npos) return;
const std::size_t start = at + std::strlen("texelFetch(mg_probeSampler, ivec2(0), ");
sampleIndex = std::atoi(source.c_str() + start);
const std::size_t dot = source.find(").", start);
if (dot != std::string::npos && dot + 2 < source.size()) component = source[dot + 2];
}
MG_External::GLESFunctionsTable MakeFakeGLESFunctions() {
MG_External::GLESFunctionsTable funcs{};
funcs.glGetError = []() -> GLenum {
const GLenum error = g_fake.pendingError;
g_fake.pendingError = GL_NO_ERROR;
return error;
};
funcs.glGetIntegerv = [](GLenum pname, GLint* data) {
switch (pname) {
case GL_MAX_COLOR_TEXTURE_SAMPLES:
if (g_fake.colorTextureSamplesQueryRaisesError) {
g_fake.pendingError = GL_INVALID_ENUM;
} else {
*data = g_fake.maxColorTextureSamples;
}
break;
case GL_MAX_IMAGE_UNITS:
*data = g_fake.maxImageUnits;
break;
case GL_MAX_VERTEX_IMAGE_UNIFORMS:
*data = g_fake.maxVertexImageUniforms;
break;
case GL_MAX_FRAGMENT_IMAGE_UNIFORMS:
*data = g_fake.maxFragmentImageUniforms;
break;
case GL_MAX_GEOMETRY_IMAGE_UNIFORMS:
if (g_fake.geometryImageLimitQueryRaisesError) {
g_fake.pendingError = GL_INVALID_ENUM;
} else {
*data = g_fake.maxGeometryImageUniforms;
}
break;
case GL_MAX_GEOMETRY_SHADER_STORAGE_BLOCKS:
*data = g_fake.maxGeometrySsboBlocks;
break;
default:
break;
}
};
funcs.glGetIntegeri_v = [](GLenum, GLuint, GLint* data) { *data = 0; };
funcs.glGetFloatv = [](GLenum, GLfloat* data) {
data[0] = 0.0f;
data[1] = 0.0f;
data[2] = 0.0f;
data[3] = 0.0f;
};
funcs.glIsEnabled = [](GLenum) -> GLboolean { return GL_FALSE; };
funcs.glEnable = [](GLenum) {};
funcs.glDisable = [](GLenum) {};
funcs.glFinish = []() {};
funcs.glMemoryBarrier = [](GLbitfield) {};
funcs.glPixelStorei = [](GLenum, GLint) {};
funcs.glViewport = [](GLint, GLint, GLsizei, GLsizei) {};
funcs.glClear = [](GLbitfield) {};
funcs.glClearColor = [](GLfloat, GLfloat, GLfloat, GLfloat) {};
funcs.glActiveTexture = [](GLenum) {};
// ---- shaders and programs -------------------------------------------
funcs.glCreateShader = [](GLenum) -> GLuint {
++g_fake.aliveShaders;
return g_fake.nextShaderId++;
};
funcs.glShaderSource = [](GLuint shader, GLsizei count, const GLchar* const* strings,
const GLint*) {
std::string source;
for (GLsizei i = 0; i < count; ++i) {
if (strings[i] != nullptr) source += strings[i];
}
g_fake.shaderSources[shader] = std::move(source);
};
funcs.glCompileShader = [](GLuint) {};
funcs.glGetShaderiv = [](GLuint, GLenum pname, GLint* params) {
if (pname == GL_COMPILE_STATUS) *params = GL_TRUE;
};
funcs.glGetShaderInfoLog = [](GLuint, GLsizei bufSize, GLsizei*, GLchar* infoLog) {
if (bufSize > 0) infoLog[0] = '\0';
};
funcs.glDeleteShader = [](GLuint shader) {
if (shader != 0) --g_fake.aliveShaders;
};
funcs.glCreateProgram = []() -> GLuint {
++g_fake.alivePrograms;
return g_fake.nextProgramId++;
};
funcs.glAttachShader = [](GLuint program, GLuint shader) {
g_fake.programShaders[program].push_back(shader);
};
funcs.glLinkProgram = [](GLuint program) {
const std::vector<std::string> names = DeclaredImageNames(program);
const bool overBudget = static_cast<int>(names.size()) > g_fake.distinctImageNameBudget;
g_fake.programLinked[program] = !overBudget;
g_fake.programInfoLogs[program] = overBudget ? kImageLocationLinkLog : "";
if (!overBudget && !StageSourceContaining(program, "texelFetch(mg_probeSampler").empty()) {
++g_fake.sampledMultisampleProgramCount;
}
};
funcs.glGetProgramiv = [](GLuint program, GLenum pname, GLint* params) {
if (pname != GL_LINK_STATUS) return;
const auto it = g_fake.programLinked.find(program);
*params = (it == g_fake.programLinked.end() || it->second) ? GL_TRUE : GL_FALSE;
};
funcs.glGetProgramInfoLog = [](GLuint program, GLsizei bufSize, GLsizei*, GLchar* infoLog) {
if (bufSize <= 0) return;
const auto it = g_fake.programInfoLogs.find(program);
const std::string& log = it == g_fake.programInfoLogs.end() ? std::string() : it->second;
const GLsizei copied = static_cast<GLsizei>(
std::min<std::size_t>(log.size(), static_cast<std::size_t>(bufSize - 1)));
std::memcpy(infoLog, log.data(), static_cast<std::size_t>(copied));
infoLog[copied] = '\0';
};
funcs.glDeleteProgram = [](GLuint program) {
if (program != 0) --g_fake.alivePrograms;
};
funcs.glUseProgram = [](GLuint program) { g_fake.currentProgram = program; };
funcs.glGetUniformLocation = [](GLuint, const GLchar*) -> GLint { return 0; };
funcs.glUniform1i = [](GLint, GLint) {};
// ---- textures, framebuffers, vertex arrays ---------------------------
funcs.glGenTextures = [](GLsizei n, GLuint* textures) {
for (GLsizei i = 0; i < n; ++i) {
textures[i] = g_fake.nextTextureId++;
++g_fake.aliveTextures;
}
};
funcs.glBindTexture = [](GLenum target, GLuint texture) {
if (target == GL_TEXTURE_2D_MULTISAMPLE) g_fake.boundMultisampleTexture = texture;
};
funcs.glDeleteTextures = [](GLsizei n, const GLuint* textures) {
for (GLsizei i = 0; i < n; ++i) {
if (textures[i] != 0) --g_fake.aliveTextures;
}
};
funcs.glTexParameteri = [](GLenum target, GLenum pname, GLint param) {
if (target == GL_TEXTURE_2D_MULTISAMPLE && pname == GL_TEXTURE_SWIZZLE_A) {
g_fake.multisampleAlphaSwizzle[g_fake.boundMultisampleTexture] =
static_cast<GLenum>(param);
}
};
funcs.glTexImage2D = [](GLenum, GLint, GLint, GLsizei, GLsizei, GLint, GLenum, GLenum,
const void*) {};
funcs.glTexSubImage2D = [](GLenum, GLint, GLint, GLint, GLsizei, GLsizei, GLenum, GLenum,
const void*) {};
funcs.glTexStorage2D = [](GLenum, GLsizei, GLenum, GLsizei, GLsizei) {};
funcs.glTexStorage2DMultisample = [](GLenum, GLsizei, GLenum, GLsizei, GLsizei, GLboolean) {};
funcs.glGenFramebuffers = [](GLsizei n, GLuint* framebuffers) {
for (GLsizei i = 0; i < n; ++i) {
framebuffers[i] = g_fake.nextFramebufferId++;
++g_fake.aliveFramebuffers;
}
};
funcs.glBindFramebuffer = [](GLenum, GLuint) {};
funcs.glFramebufferTexture2D = [](GLenum, GLenum, GLenum, GLuint, GLint) {};
funcs.glCheckFramebufferStatus = [](GLenum) -> GLenum { return GL_FRAMEBUFFER_COMPLETE; };
funcs.glDeleteFramebuffers = [](GLsizei n, const GLuint* framebuffers) {
for (GLsizei i = 0; i < n; ++i) {
if (framebuffers[i] != 0) --g_fake.aliveFramebuffers;
}
};
funcs.glGenVertexArrays = [](GLsizei n, GLuint* arrays) {
for (GLsizei i = 0; i < n; ++i) {
arrays[i] = g_fake.nextVertexArrayId++;
++g_fake.aliveVertexArrays;
}
};
funcs.glBindVertexArray = [](GLuint) {};
funcs.glDeleteVertexArrays = [](GLsizei n, const GLuint* arrays) {
for (GLsizei i = 0; i < n; ++i) {
if (arrays[i] != 0) --g_fake.aliveVertexArrays;
}
};
funcs.glBindImageTexture = [](GLuint, GLuint, GLint, GLboolean, GLint, GLenum, GLenum) {};
// ---- the draw, where the defects live --------------------------------
funcs.glDrawArrays = [](GLenum, GLint, GLsizei) {
const GLuint program = g_fake.currentProgram;
const std::string sampling = StageSourceContaining(program, "texelFetch(mg_probeSampler");
if (!sampling.empty()) {
int sampleIndex = -1;
char component = '?';
ParseSampledFetch(sampling, sampleIndex, component);
const GLenum swizzle = g_fake.multisampleAlphaSwizzle.count(
g_fake.boundMultisampleTexture) != 0
? g_fake.multisampleAlphaSwizzle[g_fake.boundMultisampleTexture]
: GL_ALPHA;
// An R32F texel filled with (1, 0, 0, -) reads 1.0 through both the ALPHA and the
// RED swizzle sources, which is why one expected constant covers every shape.
g_fake.lastSampledValue = 1.0f;
if (g_fake.msaaEveryReadWrong) {
g_fake.lastSampledValue = 0.0f;
} else if (g_fake.msaaSwizzledAlphaCorrupted && swizzle == GL_RED && component == 'w' &&
sampleIndex != 0 && g_fake.sampledMultisampleProgramCount >= 2) {
// Uninitialised memory: a value that is neither the answer nor the clear.
g_fake.lastSampledValue = -1.34954e-17f;
}
return;
}
// Matched on the access qualifier alone, not on "coherent writeonly": the strongest
// coherency shape spells it "coherent volatile writeonly".
const std::string writeStage = StageSourceContaining(program, "writeonly");
const std::string readStage = StageSourceContaining(program, "readonly");
if (!writeStage.empty() && !readStage.empty() && Contains(readStage, "memoryBarrierImage")) {
// The coherency probe: one invocation stores and then reads back. `volatile` is
// what tells the strongest shape apart from the one MobileGL emits today, and
// giving them separate knobs is what lets a test pin the case where only the
// emitted shape is wrong - a fixable defect that must not be reported here.
g_fake.lastFailedTexelCount = Contains(readStage, "coherent volatile")
? g_fake.coherencyStrongestShapeFailedTexels
: g_fake.coherencyEmittedShapeFailedTexels;
return;
}
if (!writeStage.empty() && readStage.empty()) {
// The coherency control's store half; the load half decides the result.
g_fake.lastFailedTexelCount = 0;
return;
}
if (writeStage.empty() && !readStage.empty()) {
g_fake.lastFailedTexelCount = g_fake.coherencyControlFailedTexels;
return;
}
if (!writeStage.empty() && !readStage.empty()) {
// The qualifier-merge pair: the stores are lost when the two halves share a name.
const bool sharedName =
FirstImageNameIn(writeStage) == FirstImageNameIn(readStage) &&
!FirstImageNameIn(writeStage).empty();
const bool lost = g_fake.everyVertexImageWriteDropped ||
(g_fake.sameNameImagePairDropsWrites && sharedName);
g_fake.lastFailedTexelCount = lost ? 1 << 20 : 0;
return;
}
g_fake.lastFailedTexelCount = 0;
};
funcs.glReadPixels = [](GLint, GLint, GLsizei width, GLsizei height, GLenum format, GLenum type,
void* pixels) {
const std::size_t texels = static_cast<std::size_t>(width) * static_cast<std::size_t>(height);
if (format == GL_RED && type == GL_FLOAT) {
GLfloat* out = static_cast<GLfloat*>(pixels);
for (std::size_t i = 0; i < texels; ++i) out[i] = g_fake.lastSampledValue;
return;
}
GLubyte* out = static_cast<GLubyte*>(pixels);
const std::size_t failed =
std::min<std::size_t>(texels, static_cast<std::size_t>(g_fake.lastFailedTexelCount));
for (std::size_t i = 0; i < texels; ++i) {
const bool ok = i >= failed;
out[i * 4 + 0] = ok ? 0 : 255;
out[i * 4 + 1] = ok ? 255 : 0;
out[i * 4 + 2] = 0;
out[i * 4 + 3] = 255;
}
};
return funcs;
}
void ExpectProbeReleasedEverything() {
EXPECT_EQ(g_fake.aliveShaders, 0);
EXPECT_EQ(g_fake.alivePrograms, 0);
EXPECT_EQ(g_fake.aliveTextures, 0);
EXPECT_EQ(g_fake.aliveFramebuffers, 0);
EXPECT_EQ(g_fake.aliveVertexArrays, 0);
}
} // namespace
// The rule the whole section depends on: a probe that cannot run reports NO bug. If an
@@ -31,6 +454,10 @@ TEST(DriverBugProbes, AProbeThatCannotRunReportsNoBug) {
const MG_External::GLESFunctionsTable gl = EmptyFunctionTable();
EXPECT_FALSE(ProbeGeometryStageSsboWriteAfterEmitDropped(gl))
<< "a probe with no entry points to call must not claim the driver is affected";
EXPECT_FALSE(ProbeR32FMultisampleSwizzleCorruption(gl));
EXPECT_FALSE(ProbeImageLocationPerNameBudget(gl).detected);
EXPECT_FALSE(ProbeCrossStageImageQualifierMergeDropsWrites(gl));
EXPECT_FALSE(ProbeImageWriteReadCoherencyResidual(gl).detected);
}
// The section lists only bugs the device HAS, so a driver nothing could be probed on renders
@@ -52,3 +479,182 @@ TEST(DriverBugProbes, EveryFindingCarriesANameAndAnExplanation) {
finding.verdict == DriverBugVerdict::Unfixable);
}
}
// ===================== R32F MULTISAMPLE SWIZZLE =====================
TEST(DriverBugProbes, R32FMultisampleSwizzleIsCleanOnAConformingDriver) {
ResetFakeDriver();
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeR32FMultisampleSwizzleCorruption(gl));
ExpectProbeReleasedEverything();
}
TEST(DriverBugProbes, R32FMultisampleSwizzleIsDetectedFromTheSecondProgramOnward) {
ResetFakeDriver();
g_fake.msaaSwizzledAlphaCorrupted = true;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_TRUE(ProbeR32FMultisampleSwizzleCorruption(gl));
ExpectProbeReleasedEverything();
}
// The control rule, made executable: a driver on which even the default-swizzle, sample-zero and
// .x reads are wrong is broken in some larger way, and the probe may not name the alpha swizzle
// as the cause.
TEST(DriverBugProbes, R32FMultisampleSwizzleReportsNothingWhenTheControlsAreWrongToo) {
ResetFakeDriver();
g_fake.msaaSwizzledAlphaCorrupted = true;
g_fake.msaaEveryReadWrong = true;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeR32FMultisampleSwizzleCorruption(gl))
<< "with every read wrong the probe has no evidence that the alpha swizzle is the variable";
}
TEST(DriverBugProbes, R32FMultisampleSwizzleNeedsMoreThanOneSample) {
ResetFakeDriver();
g_fake.msaaSwizzledAlphaCorrupted = true;
g_fake.maxColorTextureSamples = 1;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeR32FMultisampleSwizzleCorruption(gl));
}
// ===================== IMAGE LOCATION PER NAME =====================
TEST(DriverBugProbes, ImageLocationBudgetIsCleanWhenNamesDoNotCost) {
ResetFakeDriver();
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
const auto measurement = ProbeImageLocationPerNameBudget(gl);
EXPECT_FALSE(measurement.detected);
ExpectProbeReleasedEverything();
}
TEST(DriverBugProbes, ImageLocationBudgetIsDetectedWhenOnlyTheSharedNamesLink) {
ResetFakeDriver();
// Four image uniforms per stage: twelve distinct names in the subject, four in the control.
g_fake.distinctImageNameBudget = 5;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
const auto measurement = ProbeImageLocationPerNameBudget(gl);
EXPECT_TRUE(measurement.detected);
EXPECT_EQ(measurement.perStageImageUniforms, g_fake.maxGeometryImageUniforms + 1);
EXPECT_EQ(measurement.subjectDistinctNames, measurement.perStageImageUniforms * 3);
EXPECT_EQ(measurement.controlDistinctNames, measurement.perStageImageUniforms);
EXPECT_NE(measurement.driverMessage.find("exceeds max allowed"), String::npos)
<< "the report quotes the driver rather than paraphrasing it";
ExpectProbeReleasedEverything();
}
// The control rule again: when the shared-name program is refused too, the shape is simply too
// big for this driver and the refusal is honest.
TEST(DriverBugProbes, ImageLocationBudgetReportsNothingWhenTheControlAlsoFails) {
ResetFakeDriver();
g_fake.distinctImageNameBudget = 2;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageLocationPerNameBudget(gl).detected);
}
TEST(DriverBugProbes, ImageLocationBudgetNeedsAGeometryStageThatCanHoldImages) {
ResetFakeDriver();
g_fake.distinctImageNameBudget = 5;
g_fake.maxGeometryImageUniforms = 0;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageLocationPerNameBudget(gl).detected);
}
TEST(DriverBugProbes, ImageLocationBudgetStaysSilentOnAContextWithoutTheGeometryLimit) {
ResetFakeDriver();
g_fake.distinctImageNameBudget = 5;
g_fake.geometryImageLimitQueryRaisesError = true;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageLocationPerNameBudget(gl).detected)
<< "a pre-ES-3.2 context has no geometry stage to build the shape out of";
}
// ===================== CROSS-STAGE QUALIFIER MERGE =====================
TEST(DriverBugProbes, QualifierMergeIsCleanWhenTheDriverKeepsTheStore) {
ResetFakeDriver();
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeCrossStageImageQualifierMergeDropsWrites(gl));
ExpectProbeReleasedEverything();
}
TEST(DriverBugProbes, QualifierMergeIsDetectedWhenOnlyTheSharedNameLosesTheStore) {
ResetFakeDriver();
g_fake.sameNameImagePairDropsWrites = true;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_TRUE(ProbeCrossStageImageQualifierMergeDropsWrites(gl));
ExpectProbeReleasedEverything();
}
// A driver that loses the RENAMED store too cannot write images from the vertex stage at all -
// a different and much larger claim, which this probe may not make.
TEST(DriverBugProbes, QualifierMergeReportsNothingWhenTheRenamedControlAlsoFails) {
ResetFakeDriver();
g_fake.sameNameImagePairDropsWrites = true;
g_fake.everyVertexImageWriteDropped = true;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeCrossStageImageQualifierMergeDropsWrites(gl));
}
TEST(DriverBugProbes, QualifierMergeNeedsVertexStageImageUniforms) {
ResetFakeDriver();
g_fake.sameNameImagePairDropsWrites = true;
g_fake.maxVertexImageUniforms = 0;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeCrossStageImageQualifierMergeDropsWrites(gl));
}
// ===================== IMAGE COHERENCY RESIDUAL =====================
TEST(DriverBugProbes, ImageCoherencyIsCleanWhenTheDependentReadObservesTheStore) {
ResetFakeDriver();
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
const auto measurement = ProbeImageWriteReadCoherencyResidual(gl);
EXPECT_FALSE(measurement.detected);
ExpectProbeReleasedEverything();
}
TEST(DriverBugProbes, ImageCoherencyResidualIsDetectedAndQuantified) {
ResetFakeDriver();
g_fake.coherencyStrongestShapeFailedTexels = 376;
g_fake.coherencyEmittedShapeFailedTexels = 418;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
const auto measurement = ProbeImageWriteReadCoherencyResidual(gl);
EXPECT_TRUE(measurement.detected);
EXPECT_EQ(measurement.mismatchedTexels, 376);
EXPECT_EQ(measurement.emittedShapeMismatchedTexels, 418)
<< "the row reports what applications get, not only what is theoretically reachable";
EXPECT_GT(measurement.totalTexels, 418) << "the report needs a denominator to quote a rate";
ExpectProbeReleasedEverything();
}
// The reason the subject is the STRONGEST shape and not the one MobileGL emits. Mesa llvmpipe
// misses every texel with `coherent` + memoryBarrierImage() and none once the pair is also
// `volatile` - a defect MobileGL could fix by emitting a different shape, which is not what
// UNFIXABLE means and does not belong in this section.
TEST(DriverBugProbes, ImageCoherencyReportsNothingWhenAStrongerShapeWouldFixIt) {
ResetFakeDriver();
g_fake.coherencyStrongestShapeFailedTexels = 0;
g_fake.coherencyEmittedShapeFailedTexels = 4096;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageWriteReadCoherencyResidual(gl).detected)
<< "a driver the volatile shape satisfies has a fixable defect, not an unfixable one";
}
// The control rule once more: a driver whose glFinish-separated two-draw dependency is ALSO
// dirty has a bigger defect than an in-invocation ordering residual, and this probe must not
// dress that up as one.
TEST(DriverBugProbes, ImageCoherencyReportsNothingWhenTheFinishSeparatedControlIsDirtyToo) {
ResetFakeDriver();
g_fake.coherencyStrongestShapeFailedTexels = 376;
g_fake.coherencyControlFailedTexels = 4096;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageWriteReadCoherencyResidual(gl).detected);
}
TEST(DriverBugProbes, ImageCoherencyNeedsBothHalvesOfTheSplitPairInOneStage) {
ResetFakeDriver();
g_fake.coherencyStrongestShapeFailedTexels = 376;
g_fake.maxFragmentImageUniforms = 1;
const MG_External::GLESFunctionsTable gl = MakeFakeGLESFunctions();
EXPECT_FALSE(ProbeImageWriteReadCoherencyResidual(gl).detected);
}
File diff suppressed because it is too large Load Diff
+118 -3
View File
@@ -33,9 +33,7 @@ namespace MobileGL::MG_Util::SelfTest {
//
// ADDING A SIBLING IS ONE FUNCTION: write an `Optional<DriverBugFinding> ProbeXxx(gl)`
// that returns nullopt when the driver is not affected, and add it to the table in
// CollectGlesKnownDriverBugs(). Known siblings still to be probed: the R32F-MSAA swizzle
// corruption, the image-location-per-name link limit, the cross-stage qualifier merge, and
// the coherency residual.
// CollectGlesKnownDriverBugs().
// What MobileGL can do about a bug this device HAS. There is deliberately no "not
// affected" member: a driver that passes the probe produces no finding at all, so the
@@ -78,6 +76,123 @@ namespace MobileGL::MG_Util::SelfTest {
// ProbeGeometryStageSsboWriteAfterEmitDropped(), evaluated at most once per process.
Bool GeometryStageSsboWriteAfterEmitDropped(const MG_External::GLESFunctionsTable& gl);
// Samples one R32F GL_TEXTURE_2D_MULTISAMPLE texel through a swizzled alpha channel, twice,
// with a separately linked program each time. Returns true only when the swizzled read goes
// wrong while every control read stays right.
//
// Adreno 830 returns uninitialised memory - a different value every run - for
// texelFetch(sampler2DMS, ..., sampleIndex != 0).w on an R32F multisample texture whose
// GL_TEXTURE_SWIZZLE_A is not the default, from the SECOND such program in the context
// onward. The first program reads correctly, which is why the probe links two.
//
// THREE CONTROLS, each identical to the subject but for one variable, and all three must
// read correctly for a wrong subject to count: (1) the same fetch with
// GL_TEXTURE_SWIZZLE_A left at its default, (2) the same fetch at sample index 0, and
// (3) the same swizzled texture read through .x instead of .w. Without them a driver that
// simply cannot render R32F, or cannot sample multisample textures at all, would be
// reported as having this very specific corruption.
//
// Returns false when the driver cannot host the shape (no multisample R32F colour target,
// fewer than two samples, a missing entry point, an incomplete framebuffer): an
// inconclusive probe must never be reported as a bug. Restores every piece of GL state it
// touches.
Bool ProbeR32FMultisampleSwizzleCorruption(const MG_External::GLESFunctionsTable& gl);
// ProbeR32FMultisampleSwizzleCorruption(), evaluated at most once per process.
Bool R32FMultisampleSwizzleCorrupted(const MG_External::GLESFunctionsTable& gl);
// What the image-location budget probe measured. `detected` is the only field the verdict
// depends on; the rest exist so the report can say what the shape was instead of asserting
// a number that was true on one device in one campaign.
struct ImageLocationBudgetMeasurement {
Bool detected = false;
// Image uniforms declared per stage in both the subject and the control - one more than
// GL_MAX_GEOMETRY_IMAGE_UNIFORMS, which is the smallest of the three stages' budgets.
Int perStageImageUniforms = 0;
// Distinct uniform NAMES in the subject (per-stage-unique) and in the control (shared).
Int subjectDistinctNames = 0;
Int controlDistinctNames = 0;
// The first line of the driver's info log for the failing link, so the report quotes the
// driver rather than paraphrasing it.
String driverMessage;
};
// Links the same three-stage (vertex, geometry, fragment) program twice: once with every
// stage naming its image uniforms uniquely, once with all three stages sharing one set of
// names. Both declare the same number of image uniforms per stage, on the same bindings,
// with the same qualifier and the same stores - the names are the only difference.
//
// Adreno 830 charges its image-location budget per distinct NAME, so the shared-name program
// links while the per-stage-named one is rejected with "Image location or component exceeds
// max allowed", even though nothing about the image USAGE changed. That is what makes the
// shared-name link the control: it proves the driver can host this exact amount of image
// work and that only the naming moved the answer.
//
// `detected` is false unless the subject fails AND the control links. Both failing means the
// shape is simply too large for the driver (an honest refusal); both linking means the
// driver does not have this bug.
ImageLocationBudgetMeasurement ProbeImageLocationPerNameBudget(const MG_External::GLESFunctionsTable& gl);
// ProbeImageLocationPerNameBudget(), evaluated at most once per process.
const ImageLocationBudgetMeasurement& ImageLocationPerNameBudget(const MG_External::GLESFunctionsTable& gl);
// Draws one quad whose vertex stage stores to a `coherent writeonly` image and whose
// fragment stage reads the same image declared `coherent readonly` under the SAME name, then
// checks every fragment saw the store. Returns true only when the same-name program loses
// the store while the different-name control keeps it.
//
// Adreno 830 merges the two declarations into one uniform and silently discards the writing
// stage's stores. The control is the identical pair of shaders with the two halves renamed -
// exactly what MobileGL's image-uniform repair emits - which keeps every store. Without it
// the probe would be indistinguishable from "this driver cannot store to images from the
// vertex stage", which is a different and much larger claim.
//
// Returns false when the driver advertises no vertex-stage image uniforms, when an entry
// point is missing, or when the setup fails.
Bool ProbeCrossStageImageQualifierMergeDropsWrites(const MG_External::GLESFunctionsTable& gl);
// ProbeCrossStageImageQualifierMergeDropsWrites(), evaluated at most once per process.
Bool CrossStageImageQualifierMergeDropsWrites(const MG_External::GLESFunctionsTable& gl);
// What the image coherency probe measured. The residual is reported rather than hard-coded:
// it is a rate, it differs between devices, and a report that quotes a number measured
// somewhere else is worse than no number at all.
struct ImageCoherencyResidualMeasurement {
Bool detected = false;
// Texels the STRONGEST in-shader shape missed - that is what makes the defect unfixable.
Int mismatchedTexels = 0;
// Texels the shape MobileGL emits today missed, on the same driver in the same run. It
// is what applications actually get, and it is not always the same number.
Int emittedShapeMismatchedTexels = 0;
Int totalTexels = 0;
};
// Counts the texels whose dependent imageLoad() did not observe the imageStore() that
// precedes it in the same fragment invocation.
//
// THE SUBJECT IS THE STRONGEST SHAPE THE LANGUAGE OFFERS - a `coherent volatile`
// readonly/writeonly pair on one binding with BOTH memoryBarrierImage() and memoryBarrier()
// between the store and the read - and that choice is the whole reason the row can say
// "unfixable". Probing only the shape MobileGL emits today (`coherent` plus
// memoryBarrierImage()) reports a bug on drivers where simply adding `volatile` makes the
// read correct, which is a defect MobileGL could fix rather than one it cannot: measured on
// Mesa llvmpipe, the emitted shape misses every texel while the `volatile` shape misses
// none. Only a driver that fails even the strongest shape has no in-shader substitute left.
//
// The control is the same dependency split across TWO draws with a glMemoryBarrier and a
// glFinish between them. It separates "this driver cannot make image writes visible at all"
// (control also dirty - a far worse defect, and the probe declines to call it this one) from
// the finding, which is about ordering inside one invocation.
//
// `detected` is false unless the strongest shape is dirty AND the control is clean. The
// shape MobileGL emits is measured either way, so the report can say what applications get.
ImageCoherencyResidualMeasurement ProbeImageWriteReadCoherencyResidual(
const MG_External::GLESFunctionsTable& gl);
// ProbeImageWriteReadCoherencyResidual(), evaluated at most once per process.
const ImageCoherencyResidualMeasurement& ImageWriteReadCoherencyResidual(
const MG_External::GLESFunctionsTable& gl);
// Every known driver bug this GLES driver actually has. Bugs it does not have are absent,
// so an unaffected device renders an empty section rather than a wall of "not affected".
Vector<DriverBugFinding> CollectGlesKnownDriverBugs(const MG_External::GLESFunctionsTable& gl);
+300 -162
View File
@@ -37,17 +37,19 @@
namespace MobileGL::MG_Util::SelfTest {
namespace {
// Display ranks for PostCheck::displayRank: within one backend section, FAIL
// rows render first, then WARN, PASS, INFO, then the device-driver identity
// rows render first, then WARN, then PASS, then the device-driver identity
// strings, and always last (regardless of status) the strings MobileGL itself
// reports to applications. Rows are stable-sorted, so relative order within a
// rank is preserved. Purely cosmetic: the verdict computation is unaffected.
//
// There is no rank between PASS and the identity blocks because there are no INFO
// capability rows any more - see the taxonomy on ReportBuilder below.
enum DisplayRank : Int {
RankFail = 0,
RankWarn = 1,
RankPass = 2,
RankInfo = 3,
RankDriverReported = 4,
RankMobileGLReported = 5,
RankDriverReported = 3,
RankMobileGLReported = 4,
};
// Both backends' fp64 rows end the same way, and the sentence they end with depends on
@@ -68,6 +70,28 @@ namespace MobileGL::MG_Util::SelfTest {
"to advertise it anyway";
}
// ===================== THE ROW VERDICT TAXONOMY =====================
//
// EVERY CAPABILITY ROW IS PASS, WARN OR FAIL. INFO IS FOR IDENTITY ONLY - renderer
// names, version strings, driver strings - and there is deliberately no way to emit an
// INFO capability row from here: the only INFO emitters are the two identity helpers at
// the bottom of this struct. A row that says "not supported; no impact today" tells a
// reader nothing about whether their application will work, which is the one question
// the screen exists to answer.
//
// PASS - the backend supports the capability directly.
// WARN - the backend does NOT support it directly, but a MobileGL quirk substitutes
// and the application still sees correct behaviour. The detail names the
// substitute and whatever it costs.
// FAIL - unsupported, with no substitute: an application that uses it gets wrong
// output, a failed draw, or nothing at all. The detail says what breaks.
//
// FAIL comes in two flavours, and the difference is about the BACKEND, not the row.
// Fail() is for a capability the backend cannot start without, and it drives the
// backend summary to UNSUPPORTED. FailOptional() is for a capability that is just as
// unusable but that the backend runs fine without, so the summary stays DEGRADED - a
// device with no dual-source blend still plays Minecraft, and reporting the whole
// backend as unusable because of it would be a lie in the other direction.
struct ReportBuilder {
BackendPostReport report;
Bool fatalFailed = false;
@@ -77,20 +101,27 @@ namespace MobileGL::MG_Util::SelfTest {
report.checks.push_back({Move(name), "PASS", Move(detail), RankPass});
}
// FAIL on a capability the backend cannot run without: the backend summary becomes
// UNSUPPORTED.
void Fail(String name, String detail) {
fatalFailed = true;
report.checks.push_back({Move(name), "FAIL", Move(detail), RankFail});
}
// FAIL on a capability with no substitute that the backend can nonetheless run
// without. The row is as red as any other FAIL - an application using it does not
// work - but the backend summary degrades rather than declaring the whole backend
// unusable.
void FailOptional(String name, String detail) {
warnUnmet = true;
report.checks.push_back({Move(name), "FAIL", Move(detail), RankFail});
}
void Warn(String name, String detail) {
warnUnmet = true;
report.checks.push_back({Move(name), "WARN", Move(detail), RankWarn});
}
void Info(String name, String detail) {
report.checks.push_back({Move(name), "INFO", Move(detail), RankInfo});
}
// A "Backend driver reported ..." identity string straight from the device
// driver; rendered after the regular rows.
void DriverReported(String name, String detail) {
@@ -158,17 +189,19 @@ namespace MobileGL::MG_Util::SelfTest {
// applications DO, not just what they can do: with the extension advertised, Iris
// and Sodium batch their pipeline compiles and poll GL_COMPLETION_STATUS_KHR.
//
// PASS when it is on (the intended configuration once the default flips), INFO when
// it is off - "off" is a supported configuration, not a degradation, so it must not
// colour the verdict. Either way the row names MOBILEGL_ASYNC_SHADER_COMPILE, so a
// user reading a POST page can tell which side of the switch they are on and how to
// change it.
// PASS when it is on (the intended configuration once the default flips), WARN when it
// is off: the capability is not advertised, and what stands in for it - compiling on
// the calling thread - produces exactly the same programs, just without the overlap.
// Either way the row names MOBILEGL_ASYNC_SHADER_COMPILE, so a user reading a POST page
// can tell which side of the switch they are on and how to change it.
void AppendAsyncShaderCompileRow(ReportBuilder& builder) {
constexpr const char* rowName = "Asynchronous shader compilation";
if (!MG_Util::Async::AsyncShaderCompileEnabled()) {
builder.Info(rowName,
"off; glCompileShader and glLinkProgram run on the calling thread and "
"GL_KHR_parallel_shader_compile is not advertised (set environment variable "
builder.Warn(rowName,
"off; GL_KHR_parallel_shader_compile is not advertised and "
"glCompileShader/glLinkProgram run on the calling thread instead. The "
"programs are identical - only the overlap is lost, so a shaderpack load "
"takes as long as its compiles do (set environment variable "
"MOBILEGL_ASYNC_SHADER_COMPILE=1 to enable it)");
return;
}
@@ -301,22 +334,30 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("Polygon mode",
"glPolygonMode GL_LINE/GL_POINT available via GL_NV/ANGLE_polygon_mode");
} else {
builder.Warn("Polygon mode",
"no GL_NV/ANGLE_polygon_mode; glPolygonMode GL_LINE/GL_POINT falls back to GL_FILL");
builder.FailOptional("Polygon mode",
"no GL_NV/ANGLE_polygon_mode; glPolygonMode GL_LINE/GL_POINT silently "
"falls back to GL_FILL. There is no substitute - wireframe and point "
"rasterization would have to be rebuilt out of line/point primitives - "
"so an application asking for either gets solid triangles instead");
}
if (caps.SupportsIndexedColorMask) {
builder.Pass("Indexed color mask",
"per-draw-buffer glColorMaski available (ES 3.2 core or draw_buffers_indexed)");
} else {
builder.Warn("Indexed color mask",
"no indexed glColorMaski; per-draw-buffer color masks fall back to draw buffer 0");
builder.FailOptional("Indexed color mask",
"no indexed glColorMaski; every per-draw-buffer colour mask collapses "
"onto draw buffer 0's, so an MRT pass that masks its attachments "
"differently writes the wrong channels to all but one of them, with "
"nothing to substitute");
}
if (caps.SupportsDualSourceBlend) {
builder.Pass("Dual-source blend",
"GL_SRC1_* dual-source blend factors available via GL_EXT_blend_func_extended");
} else {
builder.Warn("Dual-source blend",
"no GL_EXT_blend_func_extended; GL_SRC1_* dual-source blend factors hard-fail at draw");
builder.FailOptional("Dual-source blend",
"no GL_EXT_blend_func_extended; a draw using a GL_SRC1_* blend factor "
"hard-fails, and a second fragment output cannot be produced any other "
"way");
}
if (es31) {
@@ -328,10 +369,13 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("Vertex shader storage blocks",
format("GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS = {}", maxVertexSsboBlocks));
} else {
builder.Warn("Vertex shader storage blocks",
format("GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS = {}; the Flywheel/Create indirect draw "
"machinery cannot read indirect command buffers from the vertex stage",
maxVertexSsboBlocks));
builder.FailOptional(
"Vertex shader storage blocks",
format("GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS = {}; the vertex stage cannot read a "
"storage buffer at all, and there is nothing to read one with instead - the "
"Flywheel/Create indirect draw machinery, which fetches its per-instance data "
"from a vertex-stage SSBO, cannot run",
maxVertexSsboBlocks));
}
if (caps.MaxShaderStorageBufferBindings >= 8) {
@@ -350,14 +394,15 @@ namespace MobileGL::MG_Util::SelfTest {
if (caps.SupportsPersistentMapping) {
builder.Pass("GL_EXT_buffer_storage", "supported (persistent buffer mapping)");
} else {
builder.Info("GL_EXT_buffer_storage",
"not supported; no impact today: the frontend fully emulates persistent "
"mapping regardless of this extension");
builder.Warn("GL_EXT_buffer_storage",
"not supported; the frontend emulates persistent mapping with its own "
"shadow storage instead, so glBufferStorage and a GL_MAP_PERSISTENT_BIT "
"mapping behave correctly - at the cost of the shadow copy");
}
if (caps.SupportsBaseInstance) {
builder.Pass("GL_EXT_base_instance", "supported (native baseInstance draws)");
} else {
builder.Info("GL_EXT_base_instance",
builder.Warn("GL_EXT_base_instance",
"not supported; direct baseInstance draws are emulated by shifting the "
"instanced arrays' attribute offsets, and gl_BaseInstance by a uniform. "
"The one gap is an INDIRECT draw whose command carries a non-zero "
@@ -366,15 +411,17 @@ namespace MobileGL::MG_Util::SelfTest {
// Both multi-draw rows gate on the capability flags, not the entry-point pointers:
// eglGetProcAddress may hand back a non-NULL stub for these on drivers without the
// extension (NVIDIA ES does, and its glMultiDrawElementsBaseVertexEXT stub silently
// drops every draw), so the pointers prove nothing. Absence is INFO in both cases
// because MobileGL falls back to an equivalent per-draw loop.
// drops every draw), so the pointers prove nothing. Absence is WARN in both cases:
// MobileGL falls back to an equivalent per-draw loop, so the output is identical and
// only the command count changes.
if (caps.SupportsMultiDrawIndirect) {
builder.Pass("Multi-draw indirect",
"glMultiDrawArrays/ElementsIndirectEXT available via GL_EXT_multi_draw_indirect");
} else {
builder.Info("Multi-draw indirect",
"GL_EXT_multi_draw_indirect not supported; no impact today: multi-draw "
"indirect is decomposed into per-command indirect draws regardless");
builder.Warn("Multi-draw indirect",
"GL_EXT_multi_draw_indirect not supported; MobileGL decomposes a multi-draw "
"indirect batch into per-command indirect draws, which renders the same "
"thing for one driver call per command instead of one per batch");
}
if (caps.SupportsMultiDrawElementsBaseVertex) {
builder.Pass("Multi-draw base vertex",
@@ -382,7 +429,7 @@ namespace MobileGL::MG_Util::SelfTest {
"with GL_EXT_multi_draw_arrays); glMultiDrawElementsBaseVertex batches into one "
"driver call");
} else {
builder.Info("Multi-draw base vertex",
builder.Warn("Multi-draw base vertex",
"glMultiDrawElementsBaseVertexEXT not supported (needs EXT/OES_"
"draw_elements_base_vertex plus GL_EXT_multi_draw_arrays); the batch "
"takes the next emulation tier instead, with identical output - see "
@@ -407,9 +454,12 @@ namespace MobileGL::MG_Util::SelfTest {
"available (ES 3.1 core); the opt-in \"compute\" multi-draw tier can flatten a "
"whole batch into one draw");
} else {
builder.Info("Compute shaders",
"not available (pre-ES 3.1); no impact on the default multi-draw tiers, which "
"never use compute");
builder.FailOptional("Compute shaders",
"not available (pre-ES 3.1); MobileGL advertises "
"GL_ARB_compute_shader on an OpenGL 4.x context and there is no way to "
"run a glDispatchCompute without the ES counterpart, so a program with "
"a compute shader cannot be built at all. The default multi-draw tiers "
"never use compute, so nothing else is lost");
}
{
// The same resolution the backend runs, over the capabilities probed here.
@@ -420,45 +470,58 @@ namespace MobileGL::MG_Util::SelfTest {
// was consulted.
using MG_Backend::DirectGLES::MultiDrawImpl::ResolveTier;
String resolution;
ResolveTier(caps, glesFuncs, MG_Config::Features.EsprytMultiDrawMode, &resolution);
builder.Info("Multi-draw elements tier",
"glMultiDrawElements(BaseVertex) emulation: " + resolution +
"; override with MOBILEGL_ESPRYT_MULTIDRAW_MODE");
const MG_Config::GLESMultiDrawMode tier =
ResolveTier(caps, glesFuncs, MG_Config::Features.EsprytMultiDrawMode, &resolution);
const String detail = "glMultiDrawElements(BaseVertex) emulation: " + resolution +
"; override with MOBILEGL_ESPRYT_MULTIDRAW_MODE";
// PASS only on the tier that hands the whole batch to the driver in one call.
// Every other tier is a MobileGL substitute: the output is identical, the
// command count is not.
if (tier == MG_Config::GLESMultiDrawMode::Ext) {
builder.Pass("Multi-draw elements tier", detail);
} else {
builder.Warn("Multi-draw elements tier",
detail + " - the batch is replayed rather than handed over whole, "
"which renders the same thing for more driver calls");
}
}
if (caps.SupportsTextureBorderClamp) {
builder.Pass("Texture border clamp",
"supported (GL_TEXTURE_BORDER_COLOR reaches the driver, so "
"GL_CLAMP_TO_BORDER samples the colour the application set)");
} else {
builder.Warn("Texture border clamp",
"not supported (pre-ES 3.2 without GL_EXT/OES_texture_border_clamp); "
"GL_TEXTURE_BORDER_COLOR is not synced to the driver at all, so anything "
"sampling outside a GL_CLAMP_TO_BORDER texture reads the driver's default "
"border instead of the requested colour");
builder.FailOptional(
"Texture border clamp",
"not supported (pre-ES 3.2 without GL_EXT/OES_texture_border_clamp); "
"GL_TEXTURE_BORDER_COLOR is not synced to the driver at all, so anything "
"sampling outside a GL_CLAMP_TO_BORDER texture reads the driver's default "
"border instead of the requested colour, and no wrap mode substitutes for it");
}
if (caps.SupportsTextureCubeMapArray) {
builder.Pass("Texture cube map array",
"supported (GL_TEXTURE_CUBE_MAP_ARRAY textures get real storage and can be "
"attached to a framebuffer)");
} else {
builder.Warn("Texture cube map array",
"not supported (pre-ES 3.2 without GL_EXT/OES_texture_cube_map_array); a cube "
"map array texture gets no driver storage at all, so sampling one reads nothing "
"and rendering to one does not reach the screen");
builder.FailOptional(
"Texture cube map array",
"not supported (pre-ES 3.2 without GL_EXT/OES_texture_cube_map_array); a cube "
"map array texture gets no driver storage at all, so sampling one reads nothing "
"and rendering to one does not reach the screen. Nothing substitutes: the "
"shaders that declare a samplerCubeArray do not compile either");
}
// WARN, not FAIL, and the choice is deliberate. The consequence is severe - buffer
// textures are CORE in OpenGL 3.1 and MobileGL advertises a 4.x context, so an
// application may use one without asking, and nothing degrades gracefully: the
// texture gets no driver storage, and every shader declaring a samplerBuffer fails
// to compile outright, because SPIRV-Cross emits `#extension GL_EXT_texture_buffer :
// require` for it below ESSL 320, so the program never links and every draw using it
// silently draws nothing. That is how Minecraft 26.3, whose cloud layer is built
// entirely from gl_VertexID plus texelFetch on a GL_R8I buffer texture, loses its
// clouds. But FAIL means "this backend cannot run on this driver", and that is not
// true: such a device runs everything that does not touch a buffer texture. It is
// also exactly the shape of the "Texture cube map array" row above, which loses its
// shaders to the same SPIRV-Cross `: require` mechanism and is a WARN - two adjacent
// rows with one consequence must not carry two severities.
// FAIL, and specifically FailOptional. The consequence is severe - buffer textures
// are CORE in OpenGL 3.1 and MobileGL advertises a 4.x context, so an application
// may use one without asking, and nothing degrades gracefully: the texture gets no
// driver storage, and every shader declaring a samplerBuffer fails to compile
// outright, because SPIRV-Cross emits `#extension GL_EXT_texture_buffer : require`
// for it below ESSL 320, so the program never links and every draw using it silently
// draws nothing. That is how Minecraft 26.3, whose cloud layer is built entirely
// from gl_VertexID plus texelFetch on a GL_R8I buffer texture, loses its clouds.
// There is no substitute, which is what makes the row FAIL; the backend still RUNS
// everything that does not touch a buffer texture, which is what keeps the failure
// out of the backend summary. It is exactly the shape of the "Texture cube map
// array" row above, which loses its shaders to the same SPIRV-Cross `: require`
// mechanism - two adjacent rows with one consequence must carry one severity.
// The limit is stated on every tier because it is the one number an application can
// read, and on the None tier it is knowingly a fiction (see below).
{
@@ -495,16 +558,17 @@ namespace MobileGL::MG_Util::SelfTest {
break;
case Tier::None:
default:
builder.Warn("Buffer textures",
format("not supported (pre-ES 3.2 without GL_EXT/OES_texture_buffer); "
"glTexBuffer does not exist, so a buffer texture gets no storage, "
"and any shader declaring a samplerBuffer fails to compile and "
"leaves its program unlinked - every draw using it is a silent "
"no-op. MobileGL still reports GL_MAX_TEXTURE_BUFFER_SIZE = {}: "
"the value is a floor it cannot honour, kept because an OpenGL "
"4.x context may not answer 0 and GL has no way to say that a "
"core feature is missing",
advertisedLimit));
builder.FailOptional(
"Buffer textures",
format("not supported (pre-ES 3.2 without GL_EXT/OES_texture_buffer); "
"glTexBuffer does not exist, so a buffer texture gets no storage, "
"and any shader declaring a samplerBuffer fails to compile and "
"leaves its program unlinked - every draw using it is a silent "
"no-op. MobileGL still reports GL_MAX_TEXTURE_BUFFER_SIZE = {}: "
"the value is a floor it cannot honour, kept because an OpenGL "
"4.x context may not answer 0 and GL has no way to say that a "
"core feature is missing",
advertisedLimit));
break;
}
}
@@ -513,7 +577,10 @@ namespace MobileGL::MG_Util::SelfTest {
// either answer, and the rows exist so the two halves of the loss are named at
// startup instead of discovered as a shader that will not compile or an
// unexplained GL_INVALID_OPERATION at draw setup.
builder.Pass("fp64", AppendFp64AdvertisementNote(
// WARN, not PASS: ESSL has no 64-bit float type, so this backend does not support
// fp64 directly at all. What it has is a complete substitute - the shaders build and
// run - which is exactly what WARN means.
builder.Warn("fp64", AppendFp64AdvertisementNote(
"demoted to fp32 - ESSL has no 64-bit float type, so every double / "
"dvec / dmat in a shader is narrowed to 32 bits before transpilation "
"(DemoteFloat64Pass). Such shaders COMPILE AND RUN, at single "
@@ -531,10 +598,11 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("Tessellation patch parameters",
"glPatchParameteri present (GL_PATCH_VERTICES reaches the driver)");
} else {
builder.Warn("Tessellation patch parameters",
"glPatchParameteri missing (pre-ES 3.2 without GL_EXT_tessellation_shader); "
"GL_PATCH_VERTICES stays at the driver default of 3 and a patch draw of any "
"other size renders nothing");
builder.FailOptional("Tessellation patch parameters",
"glPatchParameteri missing (pre-ES 3.2 without "
"GL_EXT_tessellation_shader); GL_PATCH_VERTICES stays at the driver "
"default of 3 and a patch draw of any other size renders nothing - "
"the patch size cannot be communicated any other way");
}
if (glesFuncs.glGenTransformFeedbacks != nullptr && glesFuncs.glBindTransformFeedback != nullptr &&
glesFuncs.glPauseTransformFeedback != nullptr && glesFuncs.glResumeTransformFeedback != nullptr) {
@@ -550,7 +618,9 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("GL_EXT_texture_norm16", "supported");
} else {
builder.Warn("GL_EXT_texture_norm16",
"not supported; 16-bit normalized texture formats need emulation");
"not supported; MobileGL substitutes a wider format for every 16-bit "
"normalized texture, so the texels are still readable at their declared "
"precision at the cost of the extra storage");
}
if (caps.SupportsRenderSnorm) {
builder.Pass("GL_EXT_render_snorm",
@@ -583,23 +653,31 @@ namespace MobileGL::MG_Util::SelfTest {
"render targets (Iris reports GL_FRAMEBUFFER_UNSUPPORTED and refuses to load)");
}
// INFO, never WARN: this is the HOST driver's ability to compile its own ESSL on
// its own threads, and MobileGL's asynchronous compilation does not depend on it
// in the slightest - the pool parallelises GLSL -> SPIR-V -> ESSL translation,
// which is where a shaderpack load actually spends its time, and it does that on
// a driver that has never heard of the extension. The row exists so that the day
// the driver-side half is overlapped too, the POST already says which devices can.
builder.Info("Driver GL_KHR_parallel_shader_compile",
caps.SupportsParallelShaderCompile
? "supported; the device driver can also compile the translated ESSL off-thread"
: "not supported; the device driver compiles the translated ESSL on the calling "
"thread (MobileGL's own compile pool is unaffected)");
// WARN and never FAIL when it is absent: this is the HOST driver's ability to
// compile its own ESSL on its own threads, and MobileGL's own compile pool stands in
// for all of it that matters - the pool parallelises GLSL -> SPIR-V -> ESSL
// translation, which is where a shaderpack load actually spends its time, and it
// does that on a driver that has never heard of the extension. The row exists so
// that the day the driver-side half is overlapped too, the POST already says which
// devices can.
if (caps.SupportsParallelShaderCompile) {
builder.Pass("Driver GL_KHR_parallel_shader_compile",
"supported; the device driver can also compile the translated ESSL off-thread");
} else {
builder.Warn("Driver GL_KHR_parallel_shader_compile",
"not supported; the device driver compiles the translated ESSL on the calling "
"thread. MobileGL's own compile pool substitutes for the expensive half of the "
"work (GLSL -> SPIR-V -> ESSL) and is unaffected, so loads still overlap");
}
builder.Info("Indirect gl_InstanceID semantics",
caps.IndirectDrawInstanceIdIncludesBaseInstance
? "includes baseInstance (ANGLE-style; MobileGL's shader rewrite keeps gl_InstanceID "
"zero-based)"
: "conforming (zero-based)");
if (caps.IndirectDrawInstanceIdIncludesBaseInstance) {
builder.Warn("Indirect gl_InstanceID semantics",
"includes baseInstance (ANGLE-style), which is not what GL promises; "
"MobileGL's shader rewrite subtracts it back out so gl_InstanceID stays "
"zero-based and instanced indirect draws index their arrays correctly");
} else {
builder.Pass("Indirect gl_InstanceID semantics", "conforming (zero-based)");
}
builder.DriverReported("Backend driver reported GL_VENDOR", caps.GLESVendorString);
builder.DriverReported("Backend driver reported GL_RENDERER", caps.GLESRendererString);
@@ -614,17 +692,21 @@ namespace MobileGL::MG_Util::SelfTest {
const MG_External::GLESFunctionsTable& glesFuncs) {
const String disabledNote = TimerQueryDisabledNote();
if (!caps.SupportsDisjointTimerQuery) {
builder.Warn("Timer queries",
"GL_EXT_disjoint_timer_query not supported; timer queries unavailable; "
"Minecraft F3 GPU% will not show" +
disabledNote);
builder.FailOptional("Timer queries",
"GL_EXT_disjoint_timer_query not supported; there is no way to time "
"GPU work from the client, so glBeginQuery(GL_TIME_ELAPSED) has "
"nothing to stand in for it and Minecraft's F3 GPU% will not show" +
disabledNote);
return;
}
// Every emit carries the extension-presence fact the old standalone
// GL_EXT_disjoint_timer_query row showed, plus the probe outcome.
const String extensionPresent = "GL_EXT_disjoint_timer_query extension present";
// FailOptional: a driver that advertises the extension and then cannot serve a
// query is broken in a way nothing substitutes for, but timing GPU work is not
// something the backend needs in order to run.
const auto fail = [&](const String& detail) {
builder.Fail("Timer queries", extensionPresent + "; but " + detail + disabledNote);
builder.FailOptional("Timer queries", extensionPresent + "; but " + detail + disabledNote);
};
if (!glesFuncs.glGenQueries || !glesFuncs.glDeleteQueries || !glesFuncs.glBeginQuery ||
@@ -758,8 +840,11 @@ namespace MobileGL::MG_Util::SelfTest {
const String pathNote = native ? "GL_NV_shader_noperspective_interpolation present (native path)"
: "GL_NV_shader_noperspective_interpolation absent (gl_Position.w / "
"gl_FragCoord.w emulation path)";
// FailOptional: a shaderpack that declares a noperspective varying renders it wrong
// and nothing stands in for the interpolation, but everything that does not use one
// is unaffected, so the backend still runs.
const auto fail = [&](const String& detail) {
builder.Fail("noperspective interpolation", pathNote + "; " + detail);
builder.FailOptional("noperspective interpolation", pathNote + "; " + detail);
};
if (!g.glCreateShader || !g.glShaderSource || !g.glCompileShader || !g.glGetShaderiv ||
@@ -1258,8 +1343,10 @@ namespace MobileGL::MG_Util::SelfTest {
const String timestampFacts =
format("timestampValidBits = {} on the graphics queue family; timestampPeriod = {} ns per tick",
timestampValidBits, timestampPeriod);
// FailOptional, for the same reason as the GLES row: the backend does not need to
// time GPU work in order to run.
const auto fail = [&](const String& detail) {
builder.Fail("Timer queries", timestampFacts + "; but " + detail + disabledNote);
builder.FailOptional("Timer queries", timestampFacts + "; but " + detail + disabledNote);
};
const auto vkCreateDeviceFn =
reinterpret_cast<PFN_vkCreateDevice>(getInstanceProcAddr(instance, "vkCreateDevice"));
@@ -1479,7 +1566,13 @@ namespace MobileGL::MG_Util::SelfTest {
Bool subgroupPropertiesAvailable,
const VkPhysicalDeviceSubgroupProperties& subgroupProperties) {
constexpr const char* RowName = "Subgroup first-reduction witness";
const auto fail = [&](String detail) { builder.Fail(RowName, Move(detail)); };
// FailOptional, not Fail. The witness reports whether the NATIVE subgroup
// first-reduction works; when it does not, the renderer takes its non-subgroup
// iteration path and draws the same image. Both an Adreno 830 and Mesa lavapipe
// fail this row's topology check today while running the DirectVulkan backend
// perfectly well, so a fatal verdict here would have the screen announce that a
// backend the user is looking at through that very backend cannot run.
const auto fail = [&](String detail) { builder.FailOptional(RowName, Move(detail)); };
if (!subgroupPropertiesAvailable) {
fail("vkGetPhysicalDeviceProperties2 could not provide raw Vulkan subgroup properties");
@@ -1506,7 +1599,13 @@ namespace MobileGL::MG_Util::SelfTest {
const IterationRPWitnessEligibilityResult eligibility = EvaluateIterationRPWitnessEligibility(limits);
if (eligibility.eligibility == IterationRPWitnessEligibility::SkipUnsupportedNativeFeatureSet) {
builder.Info(RowName, eligibility.detail);
// WARN, not FAIL: there is nothing to witness on a device with no native
// subgroup contract, and the renderer takes its non-subgroup iteration path,
// which produces the same image.
builder.Warn(RowName,
eligibility.detail +
"; the renderer takes its non-subgroup iteration path instead, which "
"renders the same thing without the first-reduction shortcut");
return;
}
if (eligibility.eligibility == IterationRPWitnessEligibility::FailInadequateLimits) {
@@ -2205,20 +2304,23 @@ namespace MobileGL::MG_Util::SelfTest {
if (features.multiDrawIndirect == VK_TRUE) {
builder.Pass("multiDrawIndirect", "indirect multi-draw batches run as single native commands");
} else {
builder.Info("multiDrawIndirect",
"unsupported; multi-draw batches fall back to one draw per command (tier "
"\"indirect\" of the multi-draw dispatch is unavailable)");
builder.Warn("multiDrawIndirect",
"unsupported; MobileGL unrolls a multi-draw batch into one draw per command "
"(tier \"indirect\" of the multi-draw dispatch is unavailable), which renders "
"the same thing for more commands");
}
if (features.drawIndirectFirstInstance == VK_TRUE) {
builder.Pass("drawIndirectFirstInstance", "indirect commands may carry a non-zero firstInstance");
} else {
builder.Warn("drawIndirectFirstInstance",
"unsupported; indirect commands with a non-zero baseInstance cannot run natively");
builder.FailOptional("drawIndirectFirstInstance",
"unsupported; an indirect command carrying a non-zero baseInstance "
"cannot run, and the offset cannot be folded into the command from the "
"CPU because the command is on the GPU");
}
// Multi-draw dispatch tiers (ext -> indirect -> unroll). INFO on the missing
// pieces: every tier has a fallback, nothing is lost, only batched into more
// commands. The renderer resolves the same chain at device creation, clamped
// by MOBILEGL_MAGMA_MULTIDRAW_MODE.
// Multi-draw dispatch tiers (ext -> indirect -> unroll). WARN on the missing
// pieces: every tier has a fallback that renders the same thing, only batched
// into more commands. The renderer resolves the same chain at device creation,
// clamped by MOBILEGL_MAGMA_MULTIDRAW_MODE.
{
Bool multiDrawExtUsable = false;
if (HasVkExtension(deviceExtensions, VK_EXT_MULTI_DRAW_EXTENSION_NAME) &&
@@ -2235,8 +2337,9 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("VK_EXT_multi_draw",
"supported; a glMultiDraw* batch runs as one vkCmdDrawMulti(Indexed)EXT");
} else {
builder.Info("VK_EXT_multi_draw",
"unsupported; glMultiDraw* batches use the indirect or unrolled tier");
builder.Warn("VK_EXT_multi_draw",
"unsupported; glMultiDraw* batches take the indirect or unrolled tier "
"instead, with identical output");
}
const char* resolvedTier = multiDrawExtUsable ? "ext"
: features.multiDrawIndirect == VK_TRUE ? "indirect"
@@ -2249,32 +2352,46 @@ namespace MobileGL::MG_Util::SelfTest {
: multiDrawMode == MG_Config::MultiDrawMode::Indirect ? "indirect"
: "unroll");
}
builder.Info("Multi-draw dispatch tier", tierDetail);
// PASS only on the tier that hands the whole batch to the driver in one command.
if (multiDrawExtUsable) {
builder.Pass("Multi-draw dispatch tier", tierDetail);
} else {
builder.Warn("Multi-draw dispatch tier",
tierDetail + "; the batch is replayed rather than handed over whole, which "
"renders the same thing for more commands");
}
}
if (features.vertexPipelineStoresAndAtomics == VK_TRUE) {
builder.Pass("vertexPipelineStoresAndAtomics",
"supported by driver (not currently enabled by the DirectVulkan backend)");
} else {
builder.Warn("vertexPipelineStoresAndAtomics",
"unsupported; shaders that write storage buffers from the vertex stage will not work");
builder.FailOptional("vertexPipelineStoresAndAtomics",
"unsupported; a shader that writes a storage buffer or runs an atomic "
"from the vertex stage cannot build a pipeline, and the write cannot be "
"moved to another stage without changing what the shader does");
}
if (features.fillModeNonSolid == VK_TRUE) {
builder.Pass("fillModeNonSolid", "glPolygonMode GL_LINE/GL_POINT rasterization supported");
} else {
builder.Warn("fillModeNonSolid",
"unsupported; glPolygonMode GL_LINE/GL_POINT falls back to GL_FILL (no wireframe/point "
"rasterization)");
builder.FailOptional("fillModeNonSolid",
"unsupported; glPolygonMode GL_LINE/GL_POINT silently falls back to "
"GL_FILL, and wireframe/point rasterization cannot be rebuilt out of "
"the triangle pipeline");
}
if (features.independentBlend == VK_TRUE) {
builder.Pass("independentBlend", "per-draw-buffer glColorMaski and indexed blend state supported");
} else {
builder.Warn("independentBlend",
"unsupported; per-draw-buffer glColorMaski falls back to draw buffer 0 for all attachments");
builder.FailOptional("independentBlend",
"unsupported; every attachment takes draw buffer 0's colour mask and "
"blend state, so an MRT pass that configures them separately writes the "
"wrong channels to all but one attachment");
}
if (features.dualSrcBlend == VK_TRUE) {
builder.Pass("dualSrcBlend", "GL_SRC1_* dual-source blend factors supported");
} else {
builder.Warn("dualSrcBlend", "unsupported; GL_SRC1_* dual-source blend factors hard-fail at draw");
builder.FailOptional("dualSrcBlend",
"unsupported; a draw using a GL_SRC1_* blend factor hard-fails, and a "
"second fragment output cannot be produced any other way");
}
// The Magma counterpart of the GLES "Buffer textures" row, so the two sections can be
// read side by side. Vulkan has no optional-feature bit here: a uniform texel buffer is
@@ -2314,10 +2431,11 @@ namespace MobileGL::MG_Util::SelfTest {
"back on its own; a format that refuses the flag is detected at image "
"creation and declines per-slice attachment)");
} else {
builder.Warn("2D-array-compatible 3D images",
"VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT unavailable for colour attachments; "
"glFramebufferTextureLayer on a GL_TEXTURE_3D texture is declined for every "
"slice past the first");
builder.FailOptional("2D-array-compatible 3D images",
"VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT unavailable for colour "
"attachments; glFramebufferTextureLayer on a GL_TEXTURE_3D texture "
"is declined for every slice past the first, and a 3D slice cannot "
"be rendered into any other way");
}
}
if (features.imageCubeArray == VK_TRUE) {
@@ -2325,9 +2443,10 @@ namespace MobileGL::MG_Util::SelfTest {
"GL_TEXTURE_CUBE_MAP_ARRAY textures get a Vulkan image and can be sampled and "
"attached to a framebuffer per layer");
} else {
builder.Warn("imageCubeArray",
"unsupported; a GL_TEXTURE_CUBE_MAP_ARRAY texture gets no image at all, so sampling "
"one reads nothing and glFramebufferTextureLayer on one is declined");
builder.FailOptional("imageCubeArray",
"unsupported; a GL_TEXTURE_CUBE_MAP_ARRAY texture gets no image at all, "
"so sampling one reads nothing and glFramebufferTextureLayer on one is "
"declined - there is no substitute image type");
}
// MobileGL follows the device here: shaderFloat64 decides whether a module keeps its
// 64-bit floats or has them narrowed before pipeline creation (DemoteFloat64Pass). Adreno
@@ -2342,7 +2461,10 @@ namespace MobileGL::MG_Util::SelfTest {
"FETCH here, so such a program is narrowed whole exactly as it would be on a "
"device without the feature"));
} else {
builder.Pass("fp64", AppendFp64AdvertisementNote(
// WARN rather than PASS: the device does not support fp64 at all here, and what
// stands in for it is a MobileGL pass that narrows the shader. It runs, at single
// precision - the definition of a substitute.
builder.Warn("fp64", AppendFp64AdvertisementNote(
"demoted to fp32 (device shaderFloat64 = unsupported) - every double / dvec "
"/ dmat in a shader is narrowed to 32 bits before pipeline creation, so such "
"shaders BUILD AND RUN at single precision instead of failing to create a "
@@ -2375,8 +2497,10 @@ namespace MobileGL::MG_Util::SelfTest {
if (shaderDrawParameters) {
builder.Pass("shaderDrawParameters", "gl_DrawID/gl_BaseVertex/gl_BaseInstance shaders supported");
} else {
builder.Warn("shaderDrawParameters",
"unavailable; shaders using gl_DrawID/gl_BaseInstance will not work");
builder.FailOptional("shaderDrawParameters",
"unavailable; a shader reading gl_DrawID, gl_BaseVertex or "
"gl_BaseInstance has no SPIR-V builtin to read them from, so such "
"shaders do not work and nothing supplies the values instead");
}
summary.shaderDrawParametersSupported = shaderDrawParameters;
@@ -2413,10 +2537,12 @@ namespace MobileGL::MG_Util::SelfTest {
"supported; flat varyings take GL's last vertex and transform feedback records "
"strip/fan triangles in GL's vertex order");
} else {
builder.Warn("provokingVertexLast",
"unsupported; flat-shaded varyings take a primitive's first vertex instead of GL's "
"last, and transform feedback records TRIANGLE_STRIP/TRIANGLE_FAN triangles rotated "
"(e.g. 0,1,2 / 1,3,2 instead of 0,1,2 / 2,1,3)");
builder.FailOptional("provokingVertexLast",
"unsupported; flat-shaded varyings take a primitive's first vertex "
"instead of GL's last, and transform feedback records "
"TRIANGLE_STRIP/TRIANGLE_FAN triangles rotated (e.g. 0,1,2 / 1,3,2 "
"instead of 0,1,2 / 2,1,3). Rewriting the convention would mean "
"reordering every index buffer, which MobileGL does not do");
}
if (provokingVertexLast && !transformFeedbackPreservesProvokingVertex) {
builder.Warn("transformFeedbackPreservesProvokingVertex",
@@ -2445,9 +2571,10 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("primitiveTopologyListRestart",
"primitive restart supported on list topologies (GL_PRIMITIVE_RESTART)");
} else {
builder.Warn("primitiveTopologyListRestart",
"unsupported; primitive restart works on strip/fan topologies only, list-topology restart "
"hard-fails at draw");
builder.FailOptional("primitiveTopologyListRestart",
"unsupported; primitive restart works on strip/fan topologies only, and "
"a list-topology draw with GL_PRIMITIVE_RESTART enabled hard-fails - "
"splitting the index stream on the CPU is not done");
}
// Core 1.0 features the backend turns GL stages into pipeline stages with.
@@ -2457,9 +2584,10 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("tessellationShader",
"supported (GL_PATCHES draws run the tessellation control/evaluation stages)");
} else {
builder.Warn("tessellationShader",
"unsupported; a program with a tessellation control/evaluation shader cannot build a "
"pipeline, so GL_PATCHES draws render nothing");
builder.FailOptional("tessellationShader",
"unsupported; a program with a tessellation control/evaluation shader "
"cannot build a pipeline, so GL_PATCHES draws render nothing and there "
"is no stage to run the tessellation on instead");
}
Bool vertexAttributeInstanceRateDivisor = false;
@@ -2477,10 +2605,11 @@ namespace MobileGL::MG_Util::SelfTest {
builder.Pass("vertexAttributeInstanceRateDivisor",
"supported (glVertexAttribDivisor advances an attribute every N instances)");
} else {
builder.Warn("vertexAttributeInstanceRateDivisor",
"unsupported; Vulkan's instance input rate can only advance once per instance, so "
"every non-zero glVertexAttribDivisor behaves as 1 and instanced attributes meant to "
"change every N instances change every one");
builder.FailOptional("vertexAttributeInstanceRateDivisor",
"unsupported; Vulkan's instance input rate can only advance once per "
"instance, so every non-zero glVertexAttribDivisor behaves as 1 and "
"instanced attributes meant to change every N instances change every "
"one - silently wrong geometry, with no substitute fetch rate");
}
VkPhysicalDeviceSubgroupProperties subgroupProperties{};
@@ -2504,11 +2633,18 @@ namespace MobileGL::MG_Util::SelfTest {
format("basic subgroup operations in compute, subgroup size {}",
subgroupProperties.subgroupSize));
} else {
builder.Warn("Compute shader subgroup",
"basic subgroup operations are not usable from compute shaders");
builder.FailOptional("Compute shader subgroup",
"basic subgroup operations are not usable from compute shaders, so "
"MobileGL withholds GL_KHR_shader_subgroup and the subgroup "
"iteration-render-pass path cannot run; there is no scalar rewrite "
"that stands in for a subgroup reduction");
}
} else {
builder.Warn("Compute shader subgroup", "subgroup properties could not be queried");
builder.FailOptional("Compute shader subgroup",
"subgroup properties could not be queried (no "
"vkGetPhysicalDeviceProperties2, or a pre-1.1 device), so MobileGL "
"withholds GL_KHR_shader_subgroup and the subgroup paths are "
"unavailable whatever the hardware can actually do");
}
ProbeVulkanIterationRPWitness(builder, getInstanceProcAddr, instance, physicalDevice, computeQueueFamilyIndex,
@@ -2528,9 +2664,9 @@ namespace MobileGL::MG_Util::SelfTest {
if (indexTypeUint8) {
builder.Pass("Index type uint8", "supported (native GL_UNSIGNED_BYTE index buffers)");
} else {
builder.Warn("Index type uint8",
"not supported; GL_UNSIGNED_BYTE index buffers cannot be drawn (the backend "
"has no conversion fallback and asserts on uint8 index draws)");
builder.FailOptional("Index type uint8",
"not supported; a GL_UNSIGNED_BYTE index buffer cannot be drawn - the "
"backend has no widening conversion and asserts on uint8 index draws");
}
builder.DriverReported("Backend driver reported device", String(properties.deviceName));
builder.DriverReported("Backend driver reported driver version", driverVersionString + " (vendor-encoded)");
@@ -2548,12 +2684,14 @@ namespace MobileGL::MG_Util::SelfTest {
ProbeVulkanTimerQuery(builder, getInstanceProcAddr, instance, physicalDevice,
graphicsQueueFamilyIndex, graphicsQueueTimestampValidBits, timestampPeriod);
} else {
builder.Warn("Timer queries",
format("timestampValidBits = 0 on the graphics queue family; timestampPeriod = {} ns "
"per tick; timestamps unsupported on the graphics queue; timer queries "
"unavailable",
timestampPeriod) +
TimerQueryDisabledNote());
builder.FailOptional(
"Timer queries",
format("timestampValidBits = 0 on the graphics queue family; timestampPeriod = {} ns "
"per tick; the graphics queue cannot write a timestamp at all, so there is "
"nothing to time GPU work with and glBeginQuery(GL_TIME_ELAPSED) has no "
"substitute",
timestampPeriod) +
TimerQueryDisabledNote());
}
if (vkGetPhysicalDeviceFormatPropertiesFn != nullptr) {
MG_External::VulkanCapabilities formatProbeCapabilities{};
+20 -5
View File
@@ -14,21 +14,36 @@
namespace MobileGL::MG_Util::SelfTest {
// One row of a backend power-on self-test (POST) report.
//
// EVERY CAPABILITY ROW IS PASS, WARN OR FAIL; INFO IS FOR IDENTITY ONLY.
// PASS - the backend supports the capability directly.
// WARN - not directly, but a MobileGL quirk substitutes and the application still sees
// correct behaviour; the detail names the substitute and what it costs.
// FAIL - unsupported with no substitute; an application that uses it gets wrong output, a
// failed draw, or nothing.
// INFO - identity only: renderer name, API version, driver strings, and the strings
// MobileGL itself reports to applications. Never a capability answer.
// A FAIL row does not by itself mean the backend cannot run - see BackendPostReport::verdict.
struct PostCheck {
String name;
String status; // "PASS" | "WARN" | "FAIL" | "INFO"
String detail;
// Display ordering rank within a backend section (lower renders first): FAIL,
// WARN, PASS, INFO, then the device-driver identity strings, then the strings
// WARN, PASS, then the device-driver identity strings, then the strings
// MobileGL itself reports to applications. Rows are stable-sorted by this rank
// before the report is returned; it is not serialized to JSON.
Int displayRank = 0;
};
// Verdict for one backend's device driver.
// - UNSUPPORTED: a fatal check failed; the backend cannot run on this driver.
// - DEGRADED: every fatal check passed but at least one soft expectation is unmet.
// - OK: all expectations met.
// Verdict for one backend's device driver, derived from the rows.
// - UNSUPPORTED: a REQUIRED capability failed; the backend cannot run on this driver.
// - DEGRADED: every required capability is present, but at least one row is WARN or is a
// FAIL on an optional capability - the backend runs, and something an application might
// ask for is substituted or missing.
// - OK: every row passed.
// So a section can carry FAIL rows and still be DEGRADED rather than UNSUPPORTED: a device
// with no dual-source blend still runs. Which capabilities are required is decided at the
// row (ReportBuilder::Fail vs ReportBuilder::FailOptional in DriverPost.cpp).
// available is false when no probeable driver exists at all (library missing, display
// uninitializable, zero Vulkan physical devices, ...).
struct BackendPostReport {
@@ -239,13 +239,34 @@ public final class PostActivity extends Activity {
nativeLoaded = true;
}
/**
* Writes the whole report to logcat. The report is the only machine-readable form of the
* POST, and logcat drops everything past roughly 4000 bytes of a single entry - which is
* less than one backend section, so a one-call log silently truncated the report to about
* the first dozen rows. Each chunk is prefixed with its index so a reader can reassemble
* them in order (concatenate the payloads after the "] " separator).
*/
private static void logReport(String json) {
if (json == null) {
Log.i(TAG, "<null report>");
return;
}
final int chunkSize = 3000;
final int chunks = (json.length() + chunkSize - 1) / chunkSize;
for (int index = 0; index < chunks; ++index) {
final int start = index * chunkSize;
final int end = Math.min(start + chunkSize, json.length());
Log.i(TAG, "[" + (index + 1) + "/" + chunks + "] " + json.substring(start, end));
}
}
private static void runDriverPost() {
String json = null;
Throwable failure = null;
try {
ensureNativeLoaded();
json = nativeRunDriverPost();
Log.i(TAG, json == null ? "<null report>" : json);
logReport(json);
} catch (Throwable error) {
Log.e(TAG, "Driver POST failed", error);
failure = error;