Attempt to fix thanos effect on D3D11.

Telegram/SourceFiles/ui/effects/thanos_effect_renderer.cpp

@@ -20,7 +20,6 @@ https://github.com/telegramdesktop/tdesktop/blob/master/LEGAL
 namespace Ui {
 namespace {
 
-constexpr auto kParticleStride = int(24);
 constexpr auto kQuadVertexCount = int(6);
 constexpr auto kQuadVertexStride = int(2 * sizeof(float));
 constexpr auto kComputeWorkgroupSize = int(64);
@@ -157,12 +156,6 @@ void ThanosEffectRenderer::initialize(
 		sizeof(RenderUniforms));
 	_renderUniformBuffer->create();
 
-	_placeholderParticleBuffer = rhi->newBuffer(
-		QRhiBuffer::Immutable,
-		QRhiBuffer::VertexBuffer | QRhiBuffer::StorageBuffer,
-		kParticleStride);
-	_placeholderParticleBuffer->create();
-
 	_placeholderTexture = rhi->newTexture(
 		QRhiTexture::RGBA8,
 		QSize(1, 1));
@@ -176,6 +169,21 @@ void ThanosEffectRenderer::initialize(
 		QRhiSampler::ClampToEdge);
 	_placeholderSampler->create();
 
+	_placeholderStateTexture = rhi->newTexture(
+		QRhiTexture::RGBA32F,
+		QSize(1, 1),
+		1,
+		QRhiTexture::UsedWithLoadStore);
+	_placeholderStateTexture->create();
+
+	_placeholderStateSampler = rhi->newSampler(
+		QRhiSampler::Nearest,
+		QRhiSampler::Nearest,
+		QRhiSampler::None,
+		QRhiSampler::ClampToEdge,
+		QRhiSampler::ClampToEdge);
+	_placeholderStateSampler->create();
+
 	if (!createPipelines(rt)) {
 		LOG(("ThanosEffect: pipeline creation failed, disabling effect"));
 		_creationFailed = true;
@@ -203,14 +211,20 @@ bool ThanosEffectRenderer::createPipelines(QRhiRenderTarget *rt) {
 
 	_computeInitSrbLayout = _rhi->newShaderResourceBindings();
 	_computeInitSrbLayout->setBindings({
-		QRhiShaderResourceBinding::bufferLoadStore(
+		QRhiShaderResourceBinding::imageLoadStore(
 			0,
 			QRhiShaderResourceBinding::ComputeStage,
-			_placeholderParticleBuffer),
+			_placeholderStateTexture,
+			0),
 		QRhiShaderResourceBinding::uniformBuffer(
 			1,
 			QRhiShaderResourceBinding::ComputeStage,
 			_computeInitUniformBuffer),
+		QRhiShaderResourceBinding::imageLoadStore(
+			2,
+			QRhiShaderResourceBinding::ComputeStage,
+			_placeholderStateTexture,
+			0),
 	});
 	_computeInitSrbLayout->create();
 
@@ -224,14 +238,20 @@ bool ThanosEffectRenderer::createPipelines(QRhiRenderTarget *rt) {
 
 	_computeUpdateSrbLayout = _rhi->newShaderResourceBindings();
 	_computeUpdateSrbLayout->setBindings({
-		QRhiShaderResourceBinding::bufferLoadStore(
+		QRhiShaderResourceBinding::imageLoadStore(
 			0,
 			QRhiShaderResourceBinding::ComputeStage,
-			_placeholderParticleBuffer),
+			_placeholderStateTexture,
+			0),
 		QRhiShaderResourceBinding::uniformBuffer(
 			1,
 			QRhiShaderResourceBinding::ComputeStage,
 			_computeUpdateUniformBuffer),
+		QRhiShaderResourceBinding::imageLoadStore(
+			2,
+			QRhiShaderResourceBinding::ComputeStage,
+			_placeholderStateTexture,
+			0),
 	});
 	_computeUpdateSrbLayout->create();
 
@@ -255,6 +275,11 @@ bool ThanosEffectRenderer::createPipelines(QRhiRenderTarget *rt) {
 			QRhiShaderResourceBinding::FragmentStage,
 			_placeholderTexture,
 			_placeholderSampler),
+		QRhiShaderResourceBinding::sampledTexture(
+			2,
+			QRhiShaderResourceBinding::VertexStage,
+			_placeholderStateTexture,
+			_placeholderStateSampler),
 	});
 	_renderSrbLayout->create();
 
@@ -267,13 +292,9 @@ bool ThanosEffectRenderer::createPipelines(QRhiRenderTarget *rt) {
 	QRhiVertexInputLayout inputLayout;
 	inputLayout.setBindings({
 		{ quint32(kQuadVertexStride) },
-		{ quint32(kParticleStride),
-			QRhiVertexInputBinding::PerInstance },
 	});
 	inputLayout.setAttributes({
 		{ 0, 0, QRhiVertexInputAttribute::Float2, 0 },
-		{ 1, 1, QRhiVertexInputAttribute::Float2, 0 },
-		{ 1, 2, QRhiVertexInputAttribute::Float, 16 },
 	});
 	_renderPipeline->setVertexInputLayout(inputLayout);
 
@@ -450,9 +471,8 @@ void ThanosEffectRenderer::render(
 
 			const QRhiCommandBuffer::VertexInput vbufs[] = {
 				{ _quadVertexBuffer, 0 },
-				{ item.particleBuffer, 0 },
 			};
-			cb->setVertexInput(0, 2, vbufs);
+			cb->setVertexInput(0, 1, vbufs);
 
 			const auto instanceCount =
 				item.particleCountX * item.particleCountY;
@@ -544,8 +564,6 @@ ThanosEffectRenderer::AnimatingItem ThanosEffectRenderer::createAnimatingItem(
 			uint32_t(1),
 			uint32_t(maxParticles / float64(result.particleCountY)));
 	}
-	const auto particleCount =
-		result.particleCountX * result.particleCountY;
 
 	auto *tex = _rhi->newTexture(
 		QRhiTexture::RGBA8,
@@ -572,16 +590,42 @@ ThanosEffectRenderer::AnimatingItem ThanosEffectRenderer::createAnimatingItem(
 	}
 	result.sampler = sampler;
 
-	auto *particleBuf = _rhi->newBuffer(
-		QRhiBuffer::Static,
-		QRhiBuffer::VertexBuffer | QRhiBuffer::StorageBuffer,
-		particleCount * kParticleStride);
-	if (!particleBuf->create()) {
-		delete particleBuf;
+	auto *stateTex = _rhi->newTexture(
+		QRhiTexture::RGBA32F,
+		QSize(int(result.particleCountX), int(result.particleCountY)),
+		1,
+		QRhiTexture::UsedWithLoadStore);
+	if (!stateTex->create()) {
+		delete stateTex;
 		destroyAnimatingItem(result);
 		return result;
 	}
-	result.particleBuffer = particleBuf;
+	result.particleStateTexture = stateTex;
+
+	auto *velocityTex = _rhi->newTexture(
+		QRhiTexture::RGBA32F,
+		QSize(int(result.particleCountX), int(result.particleCountY)),
+		1,
+		QRhiTexture::UsedWithLoadStore);
+	if (!velocityTex->create()) {
+		delete velocityTex;
+		destroyAnimatingItem(result);
+		return result;
+	}
+	result.particleVelocityTexture = velocityTex;
+
+	auto *stateSampler = _rhi->newSampler(
+		QRhiSampler::Nearest,
+		QRhiSampler::Nearest,
+		QRhiSampler::None,
+		QRhiSampler::ClampToEdge,
+		QRhiSampler::ClampToEdge);
+	if (!stateSampler->create()) {
+		delete stateSampler;
+		destroyAnimatingItem(result);
+		return result;
+	}
+	result.particleStateSampler = stateSampler;
 
 	auto *initUbo = _rhi->newBuffer(
 		QRhiBuffer::Dynamic,
@@ -607,14 +651,20 @@ ThanosEffectRenderer::AnimatingItem ThanosEffectRenderer::createAnimatingItem(
 
 	result.computeInitSrb = _rhi->newShaderResourceBindings();
 	result.computeInitSrb->setBindings({
-		QRhiShaderResourceBinding::bufferLoadStore(
+		QRhiShaderResourceBinding::imageLoadStore(
 			0,
 			QRhiShaderResourceBinding::ComputeStage,
-			particleBuf),
+			stateTex,
+			0),
 		QRhiShaderResourceBinding::uniformBuffer(
 			1,
 			QRhiShaderResourceBinding::ComputeStage,
 			initUbo),
+		QRhiShaderResourceBinding::imageLoadStore(
+			2,
+			QRhiShaderResourceBinding::ComputeStage,
+			velocityTex,
+			0),
 	});
 	if (!result.computeInitSrb->create()) {
 		destroyAnimatingItem(result);
@@ -623,14 +673,20 @@ ThanosEffectRenderer::AnimatingItem ThanosEffectRenderer::createAnimatingItem(
 
 	result.computeUpdateSrb = _rhi->newShaderResourceBindings();
 	result.computeUpdateSrb->setBindings({
-		QRhiShaderResourceBinding::bufferLoadStore(
+		QRhiShaderResourceBinding::imageLoadStore(
 			0,
 			QRhiShaderResourceBinding::ComputeStage,
-			particleBuf),
+			stateTex,
+			0),
 		QRhiShaderResourceBinding::uniformBuffer(
 			1,
 			QRhiShaderResourceBinding::ComputeStage,
 			updateUbo),
+		QRhiShaderResourceBinding::imageLoadStore(
+			2,
+			QRhiShaderResourceBinding::ComputeStage,
+			velocityTex,
+			0),
 	});
 	if (!result.computeUpdateSrb->create()) {
 		destroyAnimatingItem(result);
@@ -659,6 +715,11 @@ ThanosEffectRenderer::AnimatingItem ThanosEffectRenderer::createAnimatingItem(
 			QRhiShaderResourceBinding::FragmentStage,
 			tex,
 			sampler),
+		QRhiShaderResourceBinding::sampledTexture(
+			2,
+			QRhiShaderResourceBinding::VertexStage,
+			stateTex,
+			stateSampler),
 	});
 	if (!result.renderSrb->create()) {
 		destroyAnimatingItem(result);
@@ -681,7 +742,9 @@ void ThanosEffectRenderer::destroyAnimatingItem(AnimatingItem &item) {
 	deferDelete(item.renderUniformBuffer);
 	deferDelete(item.computeUpdateUniformBuffer);
 	deferDelete(item.computeInitUniformBuffer);
-	deferDelete(item.particleBuffer);
+	deferDelete(item.particleStateSampler);
+	deferDelete(item.particleVelocityTexture);
+	deferDelete(item.particleStateTexture);
 	deferDelete(item.sampler);
 	deferDelete(item.texture);
 	item = {};
@@ -710,12 +773,14 @@ void ThanosEffectRenderer::releaseResources() {
 	delete _computeInitSrbLayout;
 	_computeInitSrbLayout = nullptr;
 
+	delete _placeholderStateSampler;
+	_placeholderStateSampler = nullptr;
+	delete _placeholderStateTexture;
+	_placeholderStateTexture = nullptr;
 	delete _placeholderSampler;
 	_placeholderSampler = nullptr;
 	delete _placeholderTexture;
 	_placeholderTexture = nullptr;
-	delete _placeholderParticleBuffer;
-	_placeholderParticleBuffer = nullptr;
 
 	delete _renderUniformBuffer;
 	_renderUniformBuffer = nullptr;

Telegram/SourceFiles/ui/effects/thanos_effect_renderer.h

@@ -66,7 +66,9 @@ private:
 	struct AnimatingItem {
 		QRhiTexture *texture = nullptr;
 		QRhiSampler *sampler = nullptr;
-		QRhiBuffer *particleBuffer = nullptr;
+		QRhiTexture *particleStateTexture = nullptr;
+		QRhiTexture *particleVelocityTexture = nullptr;
+		QRhiSampler *particleStateSampler = nullptr;
 		QRhiBuffer *computeInitUniformBuffer = nullptr;
 		QRhiBuffer *computeUpdateUniformBuffer = nullptr;
 		QRhiBuffer *renderUniformBuffer = nullptr;
@@ -95,9 +97,10 @@ private:
 	QRhiBuffer *_computeUpdateUniformBuffer = nullptr;
 	QRhiBuffer *_renderUniformBuffer = nullptr;
 
-	QRhiBuffer *_placeholderParticleBuffer = nullptr;
 	QRhiTexture *_placeholderTexture = nullptr;
 	QRhiSampler *_placeholderSampler = nullptr;
+	QRhiTexture *_placeholderStateTexture = nullptr;
+	QRhiSampler *_placeholderStateSampler = nullptr;
 
 	QRhiShaderResourceBindings *_computeInitSrbLayout = nullptr;
 	QRhiShaderResourceBindings *_computeUpdateSrbLayout = nullptr;

Telegram/cmake/qrhi_shaders.cmake

@@ -44,7 +44,13 @@ foreach(_src ${_shader_sources})
     if("${_ext}" STREQUAL ".comp")
         set(_glsl_ver "310es,430")
     else()
-        set(_glsl_ver "100es,120,150")
+        file(READ ${_src} _src_contents)
+        string(FIND "${_src_contents}" "texelFetch" _has_texelfetch)
+        if(NOT _has_texelfetch EQUAL -1)
+            set(_glsl_ver "300es,150")
+        else()
+            set(_glsl_ver "100es,120,150")
+        endif()
     endif()
 
     add_custom_command(

Telegram/shaders/thanos.vert

@@ -2,12 +2,11 @@
 
 layout(location = 0) in vec2 inQuadPos;
 
-layout(location = 1) in vec2 inOffset;
-layout(location = 2) in float inLifetime;
-
 layout(location = 0) out vec2 v_texcoord;
 layout(location = 1) out float v_alpha;
 
+layout(binding = 2) uniform sampler2D particleStateTex;
+
 layout(std140, binding = 0) uniform Params {
 	vec4 rect;
 	vec2 size;
@@ -20,6 +19,13 @@ void main() {
 	uint pX = particleId % particleResolution.x;
 	uint pY = particleId / particleResolution.x;
 
+	vec4 state = texelFetch(
+		particleStateTex,
+		ivec2(int(pX), int(pY)),
+		0);
+	vec2 inOffset = state.xy;
+	float inLifetime = state.z;
+
 	vec2 particleSize = size / vec2(particleResolution);
 
 	vec2 topLeft = vec2(float(pX) * particleSize.x, float(pY) * particleSize.y);

Telegram/shaders/thanos_init.comp

@@ -2,16 +2,8 @@
 
 layout(local_size_x = 64) in;
 
-struct Particle {
-	vec2 offset;
-	vec2 velocity;
-	float lifetime;
-	float _pad;
-};
-
-layout(std430, binding = 0) buffer Particles {
-	Particle particles[];
-};
+layout(rgba32f, binding = 0) uniform image2D particleStateImage;
+layout(rgba32f, binding = 2) uniform image2D particleVelocityImage;
 
 layout(std140, binding = 1) uniform Params {
 	uint particleCountX;
@@ -45,11 +37,12 @@ void main() {
 	float direction = hashFloat(s) * (3.14159265 * 2.0);
 	float speed = (0.1 + hashFloat(s + 1u) * 0.1) * 320.0;
 
-	Particle p;
-	p.offset = vec2(0.0, 0.0);
-	p.velocity = vec2(cos(direction) * speed, sin(direction) * speed);
-	p.lifetime = 1.5 + hashFloat(s + 2u) * 1.5;
-	p._pad = 0.0;
+	vec2 velocity = vec2(cos(direction) * speed, sin(direction) * speed);
+	float lifetime = 1.5 + hashFloat(s + 2u) * 1.5;
 
-	particles[gid] = p;
+	ivec2 coord = ivec2(
+		int(gid % particleCountX),
+		int(gid / particleCountX));
+	imageStore(particleStateImage, coord, vec4(0.0, 0.0, lifetime, 0.0));
+	imageStore(particleVelocityImage, coord, vec4(velocity, 0.0, 0.0));
 }

Telegram/shaders/thanos_update.comp

@@ -2,16 +2,8 @@
 
 layout(local_size_x = 64) in;
 
-struct Particle {
-	vec2 offset;
-	vec2 velocity;
-	float lifetime;
-	float _pad;
-};
-
-layout(std430, binding = 0) buffer Particles {
-	Particle particles[];
-};
+layout(rgba32f, binding = 0) uniform image2D particleStateImage;
+layout(rgba32f, binding = 2) uniform image2D particleVelocityImage;
 
 layout(std140, binding = 1) uniform Params {
 	uint particleCountX;
@@ -43,11 +35,20 @@ void main() {
 	float particleXFraction = float(particleX) / float(particleCountX);
 	float particleFraction = easeInWindow(effectFraction, particleXFraction);
 
-	Particle p = particles[gid];
+	ivec2 coord = ivec2(
+		int(particleX),
+		int(gid / particleCountX));
 
-	p.offset += p.velocity * timeStep * particleFraction;
-	p.velocity.y += 80.0 * timeStep * particleFraction;
-	p.lifetime = max(0.0, p.lifetime - 0.6 * timeStep * particleFraction);
+	vec4 state = imageLoad(particleStateImage, coord);
+	vec2 offset = state.xy;
+	float lifetime = state.z;
 
-	particles[gid] = p;
+	vec2 velocity = imageLoad(particleVelocityImage, coord).xy;
+
+	offset += velocity * timeStep * particleFraction;
+	velocity.y += 80.0 * timeStep * particleFraction;
+	lifetime = max(0.0, lifetime - 0.6 * timeStep * particleFraction);
+
+	imageStore(particleStateImage, coord, vec4(offset, lifetime, 0.0));
+	imageStore(particleVelocityImage, coord, vec4(velocity, 0.0, 0.0));
 }