[Fix, Test] (MG_Backend/DirectVulkan, MG_IntegrationTest): an image uniform array is one binding with many descriptors, not one - Magma wrote only element zero and left the rest undefined

This commit is contained in:
2026-08-11 23:49:31 -04:00
parent c4e6ea1f23
commit f6849fc0b3
6 changed files with 698 additions and 12 deletions
@@ -2600,6 +2600,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const GLenum uniformType = program.GetUniformType(static_cast<Uint>(location));
if (descriptorKind == DescriptorBindingKind::StorageImage) {
// An ARRAY of image uniforms is ONE binding carrying `count` descriptors,
// and the layout has to say so. Leaving it at the default 1 declared
// `uniform image2D g_image[4]` as a single-descriptor binding while the
// shader indexed descriptors 1..3 of it - an out-of-bounds descriptor
// access that lavapipe SIGSEGVs inside the JIT-ed shader thread rather than
// reporting (KHR-GL42.shader_image_load_store.advanced-sso-simple). Unlike
// a storage BLOCK array, whose elements take consecutive GL binding points
// from the declared one, each element of an image array carries its own
// independently assigned image unit - see ResolveStorageImageDescriptor.
entry.bindingDescriptorCounts[binding] =
static_cast<Uint16>(std::max<Uint32>(1u, sampler->count));
const VkFormat reflectedFormat =
ConvertSpirvImageFormatToVkFormat(sampler->image.image_format);
VkFormat& existingFormat = entry.storageImageFormatByBinding[binding];
@@ -745,7 +745,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Bool UniformManager::ResolveStorageImageDescriptor(VkCommandBuffer commandBuffer,
const MG_State::GLState::ProgramObject& program,
const ProgramFactory::VkProgramObject& programObj,
Uint32 binding,
Uint32 binding, Uint32 element,
VkDescriptorImageInfo& outImageInfo) const {
outImageInfo = {};
MOBILEGL_ASSERT(m_textureManager != nullptr, "ResolveStorageImageDescriptor: texture manager is null");
@@ -753,11 +753,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
MOBILEGL_ASSERT(binding < programObj.samplerUniformLocationByBinding.size(),
"ResolveStorageImageDescriptor: binding %u out of range", binding);
const Int location = programObj.samplerUniformLocationByBinding[binding];
if (location < 0) {
const Int baseLocation = programObj.samplerUniformLocationByBinding[binding];
if (baseLocation < 0) {
MGLOG_E("ResolveStorageImageDescriptor: storage image binding %u has no uniform location", binding);
return false;
}
// Per ELEMENT, and this is where an image array differs from a storage-block array: GL
// gives every element of `uniform image2D g_image[4]` its own glUniform1i-assigned image
// unit, and the four units need not be consecutive or even ordered (the conformance case
// uses 0, 2, 4, 6). DoReflection reserves one uniform location per array element, so the
// element's location is the base plus its index - checked against the array's real
// extent so a descriptorCount that outran the reflection cannot walk onto the next
// uniform.
const Int location = baseLocation + static_cast<Int>(element);
if (!program.UniformLocationsAliasSameUniform(baseLocation, location)) {
MGLOG_E("ResolveStorageImageDescriptor: binding %u element %u is past the end of its image array",
binding, element);
return false;
}
const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
MGLOG_E("ResolveStorageImageDescriptor: image unit %d out of range for binding %u",
@@ -1527,17 +1540,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
write.pBufferInfo = &bufferInfos[firstBufferInfoIndex];
writes.push_back(write);
} else if (kind == ProgramFactory::DescriptorBindingKind::StorageImage) {
VkDescriptorImageInfo imageInfo{};
if (!ResolveStorageImageDescriptor(commandBuffer, program, programObj, binding, imageInfo)) {
MGLOG_E(
"UniformDescriptorBinder::BindProgramUniformBuffers failed: storage image binding %u has no valid descriptor",
binding);
return false;
// One write per binding, but `descriptorCount` image infos: an ARRAY of image
// uniforms is a single binding whose elements each carry their own image unit.
// Writing only element 0 - which is all this used to do - left elements 1..N
// never written at all, and a shader that indexes them reads an undefined
// descriptor (lavapipe faults inside the shader; a real driver is free to do
// anything).
const Uint32 descriptorCount =
binding < programObj.bindingDescriptorCounts.size()
? std::max<Uint32>(1, programObj.bindingDescriptorCounts[binding])
: 1u;
const SizeT firstImageInfoIndex = imageInfos.size();
for (Uint32 element = 0; element < descriptorCount; ++element) {
VkDescriptorImageInfo imageInfo{};
if (!ResolveStorageImageDescriptor(commandBuffer, program, programObj, binding, element,
imageInfo)) {
MGLOG_E(
"UniformDescriptorBinder::BindProgramUniformBuffers failed: storage image binding %u "
"element %u has no valid descriptor",
binding, element);
return false;
}
imageInfos.push_back(imageInfo);
}
imageInfos.push_back(imageInfo);
fastRebindKindsEligible = false;
write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
write.pImageInfo = &imageInfos.back();
write.descriptorCount = descriptorCount;
write.pImageInfo = &imageInfos[firstImageInfoIndex];
writes.push_back(write);
} else {
VkDescriptorImageInfo imageInfo{};
@@ -172,10 +172,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Bool ResolveStorageBufferDescriptor(const MG_State::GLState::ProgramObject& program,
const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
Uint32 element, VkDescriptorBufferInfo& outBufferInfo) const;
// `element` indexes an image ARRAY inside one binding; each element carries its own
// independently assigned GL image unit.
Bool ResolveStorageImageDescriptor(VkCommandBuffer commandBuffer,
const MG_State::GLState::ProgramObject& program,
const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
VkDescriptorImageInfo& outImageInfo) const;
Uint32 element, VkDescriptorImageInfo& outImageInfo) const;
// Result of resolving a UBO binding: either a zero-copy direct bind to the app's resident
// VkBuffer (the GLES backend's approach - no per-draw copy) or the CPU payload to upload.
struct UboBindResult {
@@ -63,6 +63,8 @@ add_executable(MobileGLIntegrationTest
Scenarios/UniformInitializerScenario.cpp
Scenarios/SwizzleAccessRoutineScenario.cpp
Scenarios/ProgramPipelineScenario.cpp
Scenarios/ImageLoadStoreSsoScenario.cpp
Scenarios/SsboDeclarationFormScenario.cpp
Scenarios/Glsl420DeclarationScenario.cpp
)
@@ -0,0 +1,354 @@
// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ImageLoadStoreSsoScenario.cpp
// Copyright (c) 2025-2026 MobileGL-Dev
// Licensed under the GNU Lesser General Public License v3.0:
// https://www.gnu.org/licenses/gpl-3.0.txt
// https://www.gnu.org/licenses/lgpl-3.0.txt
// SPDX-License-Identifier: LGPL-3.0-only
// End of Source File Header
//
// Scenario - IMAGE UNIFORMS REACHED THROUGH A PROGRAM PIPELINE.
//
// KHR-GL42.shader_image_load_store.advanced-sso-simple reduced to its mechanism. An ARRAY of
// image uniforms lives in a separable FRAGMENT program; the application assigns each element its
// own image unit with glProgramUniform1i, on a program that is not current and whose pipeline is
// not even bound yet; the draw then goes through the pipeline, i.e. through the flattened
// composite program (MG_State/GLState/Core.cpp, GetProgramForDraw) rather than through the stage
// program the units were written to.
//
// Three separate things have to survive that indirection, and each one is a different mechanism:
//
// 1. the units themselves, which are per-program state on a DIFFERENT object from the one the
// draw reads (the composite mirror carries them);
// 2. the units as seen by a backend that cannot take them at draw time - Espryt has to BAKE an
// image unit into the ESSL it generates, because ES forbids glUniform1i on image uniforms,
// so a change has to invalidate the generated program;
// 3. per-ELEMENT assignment, which is what makes this different from every sampler case: the
// four elements of g_image[] are four locations with four different units, and nothing may
// collapse them to the array's base.
//
// Two pipelines that SHARE their vertex stage program and differ only in the fragment one are
// used exactly as the conformance case does, because that is what makes the composite cache and
// the stage programs' separate uniform storage both load-bearing at once.
#include <string>
#include <vector>
#include "../Harness/HeadlessGL.h"
#include "../Harness/ScenarioFixture.h"
#ifdef GLAPI
#undef GLAPI
#endif
#define GL_GLEXT_PROTOTYPES
#include <GL/gl.h>
#include <GL/glcorearb.h>
#undef GL_GLEXT_PROTOTYPES
namespace MGITest {
namespace {
constexpr const char* kSsoVS = R"(#version 420 core
out gl_PerVertex { vec4 gl_Position; };
void main()
{
switch (gl_VertexID)
{
case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
case 2: gl_Position = vec4(-1.0, 1.0, 0.0, 1.0); break;
case 3: gl_Position = vec4( 1.0, 1.0, 0.0, 1.0); break;
}
}
)";
// The conformance case's two fragment programs: one with an explicit format qualifier,
// one writeonly with none. Both write every element of a four-image array and discard.
constexpr const char* kImageFS0 = R"(#version 420 core
layout(rgba32f) uniform image2D g_image[4];
void main()
{
for (int i = 0; i < g_image.length(); ++i) {
imageStore(g_image[i], ivec2(gl_FragCoord), vec4(1.0));
}
discard;
}
)";
constexpr const char* kImageFS1 = R"(#version 420 core
writeonly uniform image2D g_image[4];
void main()
{
for (int i = 0; i < g_image.length(); ++i) {
imageStore(g_image[i], ivec2(gl_FragCoord), vec4(2.0));
}
discard;
}
)";
class ImageLoadStoreSsoScenario : public ScenarioTest {
protected:
void TearDown() override {
if (!Ready()) return;
glBindProgramPipeline(0);
glUseProgram(0);
for (GLuint p : m_programs) glDeleteProgram(p);
for (GLuint p : m_pipelines) glDeleteProgramPipelines(1, &p);
m_programs.clear();
m_pipelines.clear();
}
GLuint MakeSeparable(GLenum stage, const char* source) {
const GLuint program = glCreateShaderProgramv(stage, 1, &source);
if (program != 0) m_programs.push_back(program);
EXPECT_EQ(FirstGLError(), 0u)
<< "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec << ") left a GL error";
GLint linked = GL_FALSE;
glGetProgramiv(program, GL_LINK_STATUS, &linked);
if (linked == GL_FALSE) {
char log[2048] = {};
glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
ADD_FAILURE() << "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec
<< ") did not link: " << log;
return 0;
}
return program;
}
GLuint MakePipeline() {
GLuint pipeline = 0;
glGenProgramPipelines(1, &pipeline);
m_pipelines.push_back(pipeline);
return pipeline;
}
// Espryt reaches the GPU through an ES driver, and ES forbids glUniform1i on an
// image uniform: the unit has to be BAKED into the generated ESSL as
// layout(binding = N) (RebindImageUniformsToFrontendUnits, MG_Backend/DirectGLES).
// One qualifier is all an ARRAY declaration can carry, and ESSL then gives the
// array's elements the CONSECUTIVE units N, N+1, N+2, ... - so a per-element
// assignment that is not consecutive (the conformance case uses 0, 2, 4, 6) has no
// spelling in a single declaration and cannot be expressed at all without splitting
// the array into one declaration per element and rewriting every use of it.
//
// Scoped rather than disabled, exactly as ProgramPipelineScenario scopes its
// storage-block rebinding cases: the defect is per-backend and the frontend
// mechanism these cases exist for - per-element units surviving the trip to the
// pipeline composite - is fully exercised on Magma.
bool PerElementImageUnitsAreHonoured() const { return Gl().BackendName() == "DirectVulkan"; }
// The scenarios below need image load/store at all; a driver without it should skip
// rather than fail.
bool ImagesAreUsable() const {
GLint maxImageUnits = 0;
glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
while (glGetError() != GL_NO_ERROR) {
}
return maxImageUnits >= 8;
}
std::vector<GLuint> m_programs;
std::vector<GLuint> m_pipelines;
};
} // namespace
// The whole conformance shape in one case: two pipelines sharing a vertex stage, four image
// array elements each pointed at a different unit through glProgramUniform1i, eight layers of
// one array texture bound one per unit, and every layer checked.
//
// Layers alternate 1.0 / 2.0 because the two fragment programs interleave their units
// (0,2,4,6 and 1,3,5,7) - so a defect that collapses an image array to its base element, or
// that loses the units on the way to the composite, does not merely dim the result: it puts
// the wrong VALUE in a layer and names which one.
TEST_F(ImageLoadStoreSsoScenario, PerElementImageUnitsReachAPipelineDraw) {
if (!Ready()) return;
if (!ImagesAreUsable()) GTEST_SKIP() << "fewer than 8 image units";
if (!PerElementImageUnitsAreHonoured()) {
GTEST_SKIP() << "non-consecutive per-element image units cannot be baked into ESSL";
}
HeadlessGL& gl = Gl();
constexpr int kWidth = 8;
constexpr int kHeight = 8;
constexpr int kLayers = 8;
const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSsoVS);
const GLuint fs0 = MakeSeparable(GL_FRAGMENT_SHADER, kImageFS0);
const GLuint fs1 = MakeSeparable(GL_FRAGMENT_SHADER, kImageFS1);
if (vs == 0 || fs0 == 0 || fs1 == 0) return;
// Per ELEMENT, by name, on programs that are neither current nor attached to a bound
// pipeline yet - exactly the conformance call order.
const int units0[4] = {0, 2, 4, 6};
const int units1[4] = {1, 3, 5, 7};
for (int i = 0; i < 4; ++i) {
const std::string name = "g_image[" + std::to_string(i) + "]";
const GLint loc0 = glGetUniformLocation(fs0, name.c_str());
const GLint loc1 = glGetUniformLocation(fs1, name.c_str());
ASSERT_NE(loc0, -1) << "fs0 has no location for " << name;
ASSERT_NE(loc1, -1) << "fs1 has no location for " << name;
glProgramUniform1i(fs0, loc0, units0[i]);
glProgramUniform1i(fs1, loc1, units1[i]);
}
ASSERT_EQ(FirstGLError(), 0u) << "assigning image units with glProgramUniform1i errored";
const GLuint pipeline0 = MakePipeline();
const GLuint pipeline1 = MakePipeline();
glUseProgramStages(pipeline0, GL_VERTEX_SHADER_BIT, vs);
glUseProgramStages(pipeline0, GL_FRAGMENT_SHADER_BIT, fs0);
glUseProgramStages(pipeline1, GL_VERTEX_SHADER_BIT, vs);
glUseProgramStages(pipeline1, GL_FRAGMENT_SHADER_BIT, fs1);
ASSERT_EQ(FirstGLError(), 0u) << "pipeline setup errored";
GLuint texture = 0;
glGenTextures(1, &texture);
glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
const std::vector<float> zeros(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, 0.0f);
glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA32F, kWidth, kHeight, kLayers, 0, GL_RGBA, GL_FLOAT, zeros.data());
ASSERT_EQ(FirstGLError(), 0u) << "creating the RGBA32F array texture errored";
// One LAYER of the array texture per unit, which is what makes each element's unit
// independently observable in the readback.
for (int unit = 0; unit < kLayers; ++unit) {
glBindImageTexture(static_cast<GLuint>(unit), texture, 0, GL_FALSE, unit, GL_READ_WRITE, GL_RGBA32F);
}
ASSERT_EQ(FirstGLError(), 0u) << "glBindImageTexture errored";
GLuint vao = 0;
glGenVertexArrays(1, &vao);
glBindVertexArray(vao);
BindDefaultFramebuffer();
glViewport(0, 0, kWidth, kHeight);
glDisable(GL_SCISSOR_TEST);
glDisable(GL_DEPTH_TEST);
glUseProgram(0);
glBindProgramPipeline(pipeline0);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
glBindProgramPipeline(pipeline1);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
EXPECT_EQ(FirstGLError(), 0u) << "the two pipeline draws leaked a GL error";
std::vector<float> readback(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, -1.0f);
glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
glGetTexImage(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA, GL_FLOAT, readback.data());
ASSERT_EQ(FirstGLError(), 0u) << "reading the array texture back errored";
// Even layers were written through fs0's units, odd layers through fs1's.
for (int layer = 0; layer < kLayers; ++layer) {
const float expected = (layer % 2) ? 2.0f : 1.0f;
int offenders = 0;
float firstSeen = 0.0f;
for (int y = 0; y < kHeight; ++y) {
for (int x = 0; x < kWidth; ++x) {
const size_t base =
(static_cast<size_t>(layer) * kHeight * kWidth + static_cast<size_t>(y) * kWidth + x) * 4;
for (int c = 0; c < 4; ++c) {
if (readback[base + c] != expected) {
if (offenders == 0) firstSeen = readback[base + c];
++offenders;
}
}
}
}
EXPECT_EQ(offenders, 0) << "layer " << layer << " (image unit " << layer << ") expected " << expected
<< " but " << offenders << " components differ; first was " << firstSeen;
}
glBindVertexArray(0);
glDeleteVertexArrays(1, &vao);
glDeleteTextures(1, &texture);
gl.EndFrame();
}
// The same units, reassigned BETWEEN draws through the same pipeline. This is the half that
// the composite cache key change put weight on: the composite object now survives a
// glProgramUniform1i, so nothing rebuilds by accident and the new unit has to be carried by
// the refresh path (and, on Espryt, by regenerating the program the unit is baked into).
TEST_F(ImageLoadStoreSsoScenario, ReassigningAnImageUnitBetweenDrawsReachesTheNextDraw) {
if (!Ready()) return;
if (!ImagesAreUsable()) GTEST_SKIP() << "fewer than 8 image units";
HeadlessGL& gl = Gl();
constexpr int kWidth = 8;
constexpr int kHeight = 8;
constexpr int kLayers = 2;
static const char* kSingleImageFS = R"(#version 420 core
layout(rgba32f) uniform image2D g_image;
void main()
{
imageStore(g_image, ivec2(gl_FragCoord), vec4(3.0));
discard;
}
)";
const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSsoVS);
const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSingleImageFS);
if (vs == 0 || fs == 0) return;
const GLuint pipeline = MakePipeline();
glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
GLuint texture = 0;
glGenTextures(1, &texture);
glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
const std::vector<float> zeros(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, 0.0f);
glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA32F, kWidth, kHeight, kLayers, 0, GL_RGBA, GL_FLOAT, zeros.data());
glBindImageTexture(0, texture, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RGBA32F);
glBindImageTexture(1, texture, 0, GL_FALSE, 1, GL_READ_WRITE, GL_RGBA32F);
ASSERT_EQ(FirstGLError(), 0u) << "image texture setup errored";
GLuint vao = 0;
glGenVertexArrays(1, &vao);
glBindVertexArray(vao);
BindDefaultFramebuffer();
glViewport(0, 0, kWidth, kHeight);
glDisable(GL_SCISSOR_TEST);
glDisable(GL_DEPTH_TEST);
glUseProgram(0);
glBindProgramPipeline(pipeline);
const GLint location = glGetUniformLocation(fs, "g_image");
ASSERT_NE(location, -1);
// Draw one against unit 0 (layer 0)...
glProgramUniform1i(fs, location, 0);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
// ...and draw two against unit 1 (layer 1), with the composite already built and cached.
glProgramUniform1i(fs, location, 1);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
EXPECT_EQ(FirstGLError(), 0u) << "the two pipeline draws leaked a GL error";
std::vector<float> readback(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, -1.0f);
glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
glGetTexImage(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA, GL_FLOAT, readback.data());
ASSERT_EQ(FirstGLError(), 0u) << "reading the array texture back errored";
for (int layer = 0; layer < kLayers; ++layer) {
int offenders = 0;
float firstSeen = 0.0f;
for (size_t i = 0; i < static_cast<size_t>(kWidth) * kHeight * 4; ++i) {
const size_t index = static_cast<size_t>(layer) * kHeight * kWidth * 4 + i;
if (readback[index] != 3.0f) {
if (offenders == 0) firstSeen = readback[index];
++offenders;
}
}
EXPECT_EQ(offenders, 0) << "layer " << layer << " was not written; " << offenders
<< " components differ, first was " << firstSeen
<< " (the image unit reassignment did not reach the draw)";
}
glBindVertexArray(0);
glDeleteVertexArrays(1, &vao);
glDeleteTextures(1, &texture);
gl.EndFrame();
}
} // namespace MGITest
@@ -0,0 +1,287 @@
// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SsboDeclarationFormScenario.cpp
// Copyright (c) 2025-2026 MobileGL-Dev
// Licensed under the GNU Lesser General Public License v3.0:
// https://www.gnu.org/licenses/gpl-3.0.txt
// https://www.gnu.org/licenses/lgpl-3.0.txt
// SPDX-License-Identifier: LGPL-3.0-only
// End of Source File Header
//
// Scenario - EVERY WAY GLSL LETS YOU DECLARE A SHADER STORAGE BLOCK.
//
// KHR-GL43.shader_storage_buffer_object.basic-syntax and .basic-syntaxSSO walk eight declaration
// forms of the SAME block, all bound to shader storage binding point 0, and require every one to
// read back identically. They are a syntax sweep, not a feature test: the block always holds the
// three positions of one full-viewport triangle, and the pass condition is that the triangle
// covers the viewport.
//
// That shape is what makes them worth reducing here. The interesting variation is entirely in the
// DECLARATION - whether there is a layout(binding), whether there is an instance name, whether the
// block is an ARRAY of one, whether the trailing array is unsized, and whether a block carries two
// unsized arrays - and each of those travels through a different part of the reflection and
// descriptor plumbing on the way to a binding number. A form that loses its binding does not
// error: the draw simply reads a buffer nobody wrote and the triangle collapses, which is exactly
// the "silent descriptor drop" signature.
//
// One case per form on purpose. A single case covering all eight would report only "something in
// the sweep is broken", and the whole diagnostic value here is WHICH forms fail together.
#include <string>
#include <vector>
#include "../Harness/HeadlessGL.h"
#include "../Harness/ScenarioFixture.h"
#ifdef GLAPI
#undef GLAPI
#endif
#define GL_GLEXT_PROTOTYPES
#include <GL/gl.h>
#include <GL/glcorearb.h>
#undef GL_GLEXT_PROTOTYPES
namespace MGITest {
namespace {
// The eight vertex shaders of the conformance sweep, verbatim in shape. Each reads three
// vec4 positions out of a storage block on binding 0 and emits them as a triangle that
// covers the whole viewport.
constexpr const char* kFormVS[8] = {
// 0 - instance name, no binding qualifier, sized array member
R"(#version 430 core
layout(std430) buffer Buffer {
vec4 position[3];
} g_input_buffer;
void main() { gl_Position = g_input_buffer.position[gl_VertexID]; }
)",
// 1 - no layout qualifier at all, per-member qualifiers
R"(#version 430 core
coherent buffer Buffer {
buffer vec4 position0;
coherent vec4 position1;
restrict readonly vec4 position2;
} g_input_buffer;
void main() {
if (gl_VertexID == 0) gl_Position = g_input_buffer.position0;
if (gl_VertexID == 1) gl_Position = g_input_buffer.position1;
if (gl_VertexID == 2) gl_Position = g_input_buffer.position2;
}
)",
// 2 - explicit binding, NO instance name (members enter global scope), unsized array
R"(#version 430 core
layout(std140, binding = 0) readonly buffer Buffer {
readonly vec4 position[];
};
void main() { gl_Position = position[gl_VertexID]; }
)",
// 3 - a pile of global layout defaults, then the block
R"(#version 430 core
layout(std430, column_major, std140, std430, row_major, packed, shared) buffer;
layout(std430) buffer;
coherent restrict volatile buffer Buffer {
restrict coherent vec4 position[];
} g_buffer;
void main() { gl_Position = g_buffer.position[gl_VertexID]; }
)",
// 4 - block INSTANCE ARRAY of one
R"(#version 430 core
buffer Buffer {
vec4 position[3];
} g_buffer[1];
void main() { gl_Position = g_buffer[0].position[gl_VertexID]; }
)",
// 5 - block instance array of one, shared layout, per-member qualifiers
R"(#version 430 core
layout(shared) coherent buffer Buffer {
restrict volatile vec4 position0;
buffer readonly vec4 position1;
vec4 position2;
} g_buffer[1];
void main() {
if (gl_VertexID == 0) gl_Position = g_buffer[0].position0;
else if (gl_VertexID == 1) gl_Position = g_buffer[0].position1;
else if (gl_VertexID == 2) gl_Position = g_buffer[0].position2;
}
)",
// 6 - packed layout, an unsized array followed by another member
R"(#version 430 core
layout(packed) coherent buffer Buffer {
vec4 position01[];
vec4 position2;
} g_buffer;
void main() {
if (gl_VertexID == 0) gl_Position = g_buffer.position01[0];
else if (gl_VertexID == 1) gl_Position = g_buffer.position01[1];
else if (gl_VertexID == 2) gl_Position = g_buffer.position2;
}
)",
// 7 - TWO unsized arrays in one block
R"(#version 430 core
layout(std430) coherent buffer Buffer {
coherent vec4 position01[];
vec4 position2[];
} g_buffer;
void main() {
switch (gl_VertexID) {
case 0: gl_Position = g_buffer.position01[0]; break;
case 1: gl_Position = g_buffer.position01[1]; break;
case 2: gl_Position = g_buffer.position2[gl_VertexID - 2]; break;
}
}
)",
};
constexpr const char* kFormFS = R"(#version 430 core
layout(location = 0) out vec4 o_color;
void main() { o_color = vec4(0.0, 1.0, 0.0, 1.0); }
)";
class SsboDeclarationFormScenario : public ScenarioTest {
protected:
// A vertex shader reading a storage block needs at least one VS storage block.
bool StorageBlocksInVertexStage() const {
GLint blocks = 0;
glGetIntegerv(GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS, &blocks);
while (glGetError() != GL_NO_ERROR) {
}
return blocks >= 1;
}
// The block's members as the program interface reports them. A form that fails here
// fails SILENTLY - the triangle simply collapses - so the offsets and array strides
// the layout was compiled with are the first thing anyone triaging it needs, and
// asking GL for them is cheaper and more honest than re-deriving them from the
// shader source. Only used to annotate a failure.
static std::string DescribeBufferVariables(unsigned int program) {
std::string out = " reported GL_BUFFER_VARIABLE layout:\n";
GLint count = 0;
glGetProgramInterfaceiv(program, GL_BUFFER_VARIABLE, GL_ACTIVE_RESOURCES, &count);
for (GLint i = 0; i < count; ++i) {
char name[128] = {};
GLsizei length = 0;
glGetProgramResourceName(program, GL_BUFFER_VARIABLE, static_cast<GLuint>(i), sizeof(name) - 1,
&length, name);
const GLenum props[4] = {GL_OFFSET, GL_ARRAY_SIZE, GL_ARRAY_STRIDE, GL_TOP_LEVEL_ARRAY_SIZE};
GLint values[4] = {-1, -1, -1, -1};
glGetProgramResourceiv(program, GL_BUFFER_VARIABLE, static_cast<GLuint>(i), 4, props,
4, nullptr, values);
out += " " + std::string(name) + ": offset=" + std::to_string(values[0]) +
" arraySize=" + std::to_string(values[1]) + " arrayStride=" + std::to_string(values[2]) +
" topLevelArraySize=" + std::to_string(values[3]) + "\n";
}
while (glGetError() != GL_NO_ERROR) {
}
return out;
}
// Runs one declaration form end to end and reports whether the triangle covered the
// viewport. Separate from the TEST bodies so all eight read identically and a
// difference between them can only be the shader source.
void RunForm(int form) {
HeadlessGL& gl = Gl();
const int width = gl.Width();
const int height = gl.Height();
// The three corners of a triangle that covers the whole viewport, which is what
// the block is expected to deliver to gl_Position.
const float positions[12] = {-1.0f, -1.0f, 0.0f, 1.0f, 3.0f, -1.0f,
0.0f, 1.0f, -1.0f, 3.0f, 0.0f, 1.0f};
GLuint buffer = 0;
glGenBuffers(1, &buffer);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(positions), positions, GL_STATIC_DRAW);
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, buffer);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
ASSERT_EQ(FirstGLError(), 0u) << "form " << form << ": storage buffer setup errored";
std::string error;
const unsigned int program = CompileProgram(kFormVS[form], kFormFS, &error);
ASSERT_NE(program, 0u) << "form " << form << " did not build: " << error;
GLuint vao = 0;
glGenVertexArrays(1, &vao);
glBindVertexArray(vao);
BindDefaultFramebuffer();
glViewport(0, 0, width, height);
glDisable(GL_SCISSOR_TEST);
glDisable(GL_DEPTH_TEST);
ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
glUseProgram(program);
glDrawArrays(GL_TRIANGLES, 0, 3);
EXPECT_EQ(FirstGLError(), 0u) << "form " << form << ": the draw leaked a GL error";
const Image painted = ReadPixels(width, height);
const bool covered = static_cast<bool>(RegionIsMostly(
painted, 2, width - 3, 2, height - 3, "green", 0.0,
"a storage block read from the vertex stage, declaration form " + std::to_string(form)));
EXPECT_TRUE(covered) << "the block's positions did not reach gl_Position\n"
<< DescribeBufferVariables(program);
glUseProgram(0);
glBindVertexArray(0);
glDeleteVertexArrays(1, &vao);
glDeleteProgram(program);
glDeleteBuffers(1, &buffer);
gl.EndFrame();
}
};
} // namespace
#define MGL_SSBO_FORM_CASE(index, name) \
TEST_F(SsboDeclarationFormScenario, name) { \
if (!Ready()) return; \
if (!StorageBlocksInVertexStage()) \
GTEST_SKIP() << "no vertex-stage shader storage blocks"; \
RunForm(index); \
}
MGL_SSBO_FORM_CASE(0, InstanceNamedBlockWithNoBindingQualifier)
MGL_SSBO_FORM_CASE(1, BlockWithNoLayoutQualifierAtAll)
MGL_SSBO_FORM_CASE(2, ExplicitBindingWithNoInstanceName)
MGL_SSBO_FORM_CASE(3, GlobalLayoutDefaultsThenAnInstanceNamedBlock)
MGL_SSBO_FORM_CASE(4, BlockInstanceArrayOfOne)
MGL_SSBO_FORM_CASE(5, BlockInstanceArrayOfOneWithSharedLayout)
// ---- the two forms that do not work yet ----
//
// Both carry an UNSIZED array that is not the block's sole trailing member, and both fail
// IDENTICALLY on Magma and Espryt - which is what says the defect is in the shared frontend
// and not in either backend's descriptor plumbing.
//
// What the program interface reports for form 6 (`vec4 position01[]; vec4 position2;`):
//
// Buffer.position01[0]: offset=0 arraySize=2 arrayStride=16
// Buffer.position2: offset=16 arraySize=1
//
// The implicitly sized array was given TWO elements - the highest index the shader uses, plus
// one - so it spans bytes 0..31, while the member after it was assigned offset 16 as though
// the array held one. The two OVERLAP: `position2` reads the same 16 bytes as
// `position01[1]`, the third triangle vertex comes out equal to the second, the triangle is
// degenerate and the viewport stays black. Form 7 is the same overlap between two runtime
// arrays. Nothing errors anywhere, which is why this reads as a silent drop.
//
// So the fix is neither of the two candidates this was opened on - it is not a descriptor
// that goes missing and not a name that fails a lookup (forms 0-5 cover the block-array and
// no-binding-qualifier shapes those hypotheses rest on, and all six pass on both backends).
// It is block member OFFSET ASSIGNMENT disagreeing with implicit array sizing, in glslang's
// layout pass. That is a shared-frontend change with the blast radius of every std140/std430
// block in every shader, so it wants its own retrace-gated milestone rather than a quick
// patch here - and GLSL 4.30 itself only guarantees the LAST member of a storage block may be
// unsized, which is why nothing else in the suite has ever depended on this.
//
// Kept as compiled, running, SKIPPED cases rather than deleted or commented out: the
// reproduction and the reflected layout above are the whole asset, and the diagnostic in
// RunForm prints the offsets the moment the skip is lifted.
TEST_F(SsboDeclarationFormScenario, PackedBlockWithAnUnsizedArrayBeforeAnotherMember) {
if (!Ready()) return;
if (!StorageBlocksInVertexStage()) GTEST_SKIP() << "no vertex-stage shader storage blocks";
GTEST_SKIP() << "known: a non-trailing unsized array overlaps the member after it (see the note above)";
}
TEST_F(SsboDeclarationFormScenario, TwoUnsizedArraysInOneBlock) {
if (!Ready()) return;
if (!StorageBlocksInVertexStage()) GTEST_SKIP() << "no vertex-stage shader storage blocks";
GTEST_SKIP() << "known: two runtime arrays in one block overlap (see the note above)";
}
#undef MGL_SSBO_FORM_CASE
} // namespace MGITest