rhi: Make it possible to discard depth/stencil even when using a texture

Also implement this for OpenGL ES since it can be relevant with tiled
architectures wrt performance.

Task-number: QTBUG-122669
Change-Id: I90dcfe4f5f9edbb8dfb51189d46b89ef2c7a7c06
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Andy Nichols <andy.nichols@qt.io>

7 files changed, 63 insertions, 5 deletions

diff --git a/src/gui/rhi/qrhi.cpp b/src/gui/rhi/qrhi.cpp
index 4cba1beec3..4940b5d85e 100644
--- a/src/gui/rhi/qrhi.cpp
+++ b/src/gui/rhi/qrhi.cpp
@@ -2542,6 +2542,24 @@ QRhiColorAttachment::QRhiColorAttachment(QRhiRenderBuffer *renderBuffer)
         QRhiTextureRenderTarget *rt = rhi->newTextureRenderTarget({ colorAtt, depthStencil });
     \endcode
 
+    \note when multisample resolving is enabled, the multisample data may not be
+    written out at all. This means that the multisample texture in a color
+    attachment must not be used afterwards with shaders for sampling (or other
+    puroses) whenever a resolve texture is set, since the multisample color
+    buffer is merely an intermediate storage then that gets no data written back
+    on some GPU architectures at all.
+
+    \note When using setDepthTexture(), not setDepthStencilBuffer(), and the
+    depth (stencil) data is not of interest afterwards, set the
+    DoNotStoreDepthStencilContents flag on the QRhiTextureRenderTarget. This
+    allows indicating to the underlying 3D API that the depth/stencil data can
+    be discarded, leading potentially to better performance with tiled GPU
+    architectures. When the depth-stencil buffer is a QRhiRenderBuffer (and also
+    for the multisample color texture, see previous note) this is implicit, but
+    with a depth (stencil) QRhiTexture the intention needs to be declared
+    explicitly. By default QRhi assumes that the data is of interest (e.g., the
+    depth texture is sampled in a shader afterwards).
+
     \note This is a RHI API with limited compatibility guarantees, see \l QRhi
     for details.
 
@@ -5053,6 +5071,12 @@ QRhiResource::Type QRhiSwapChainRenderTarget::resourceType() const
     (QRhiTextureRenderTargetDescription::depthTexture() is set) because
     depth/stencil renderbuffers may not have any physical backing and data may
     not be written out in the first place.
+
+    \value DoNotStoreDepthStencilContents Indicates that the contents of the
+    depth texture does not need to be written out. Relevant only when a
+    QRhiTexture, not QRhiRenderBuffer, is used as the depth-stencil buffer,
+    because for QRhiRenderBuffer this is implicit. This enum value is introduced
+    in Qt 6.8.
  */
 
 /*!
diff --git a/src/gui/rhi/qrhi.h b/src/gui/rhi/qrhi.h
index 30d2b3c909..a9836bf892 100644
--- a/src/gui/rhi/qrhi.h
+++ b/src/gui/rhi/qrhi.h
@@ -1181,7 +1181,8 @@ class Q_GUI_EXPORT QRhiTextureRenderTarget : public QRhiRenderTarget
 public:
     enum Flag {
         PreserveColorContents = 1 << 0,
-        PreserveDepthStencilContents = 1 << 1
+        PreserveDepthStencilContents = 1 << 1,
+        DoNotStoreDepthStencilContents = 1 << 2
     };
     Q_DECLARE_FLAGS(Flags, Flag)
 
diff --git a/src/gui/rhi/qrhigles2.cpp b/src/gui/rhi/qrhigles2.cpp
index f6b3056fa3..8e98bcd736 100644
--- a/src/gui/rhi/qrhigles2.cpp
+++ b/src/gui/rhi/qrhigles2.cpp
@@ -3679,6 +3679,13 @@ void QRhiGles2::executeCommandBuffer(QRhiCommandBuffer *cb)
             if (caps.compute)
                 f->glMemoryBarrier(cmd.args.barrier.barriers);
             break;
+        case QGles2CommandBuffer::Command::InvalidateFramebuffer:
+            if (caps.gles && caps.ctxMajor >= 3) {
+                f->glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER,
+                                           cmd.args.invalidateFramebuffer.attCount,
+                                           cmd.args.invalidateFramebuffer.att);
+            }
+            break;
         default:
             break;
         }
@@ -4506,7 +4513,10 @@ void QRhiGles2::endPass(QRhiCommandBuffer *cb, QRhiResourceUpdateBatch *resource
                 const bool hasZ = resolveTexD->m_flags.testFlag(QRhiTexture::ThreeDimensional)
                     || resolveTexD->m_flags.testFlag(QRhiTexture::TextureArray);
                 cmd.args.blitFromRenderbuffer.dstLayer = hasZ ? colorAtt.resolveLayer() : 0;
-            } else if (!caps.glesMultisampleRenderToTexture) {
+            } else if (caps.glesMultisampleRenderToTexture) {
+                // Nothing to do, resolving into colorAtt.resolveTexture() is automatic,
+                // colorAtt.texture() is in fact not used for anything.
+            } else {
                 Q_ASSERT(colorAtt.texture());
                 QGles2Texture *texD = QRHI_RES(QGles2Texture, colorAtt.texture());
                 if (texD->pixelSize() != size) {
@@ -4542,6 +4552,21 @@ void QRhiGles2::endPass(QRhiCommandBuffer *cb, QRhiResourceUpdateBatch *resource
                 }
             }
         }
+
+        const bool mayDiscardDepthStencil = rtTex->m_desc.depthStencilBuffer()
+            || (rtTex->m_desc.depthTexture() && rtTex->m_flags.testFlag(QRhiTextureRenderTarget::DoNotStoreDepthStencilContents));
+        if (mayDiscardDepthStencil) {
+            QGles2CommandBuffer::Command &cmd(cbD->commands.get());
+            cmd.cmd = QGles2CommandBuffer::Command::InvalidateFramebuffer;
+            if (caps.needsDepthStencilCombinedAttach) {
+                cmd.args.invalidateFramebuffer.attCount = 1;
+                cmd.args.invalidateFramebuffer.att[0] = GL_DEPTH_STENCIL_ATTACHMENT;
+            } else {
+                cmd.args.invalidateFramebuffer.attCount = 2;
+                cmd.args.invalidateFramebuffer.att[0] = GL_DEPTH_ATTACHMENT;
+                cmd.args.invalidateFramebuffer.att[1] = GL_STENCIL_ATTACHMENT;
+            }
+        }
     }
 
     cbD->recordingPass = QGles2CommandBuffer::NoPass;
diff --git a/src/gui/rhi/qrhigles2_p.h b/src/gui/rhi/qrhigles2_p.h
index 4407cfea36..04ef596267 100644
--- a/src/gui/rhi/qrhigles2_p.h
+++ b/src/gui/rhi/qrhigles2_p.h
@@ -337,7 +337,8 @@ struct QGles2CommandBuffer : public QRhiCommandBuffer
             BindComputePipeline,
             Dispatch,
             BarriersForPass,
-            Barrier
+            Barrier,
+            InvalidateFramebuffer
         };
         Cmd cmd;
 
@@ -536,6 +537,10 @@ struct QGles2CommandBuffer : public QRhiCommandBuffer
             struct {
                 GLbitfield barriers;
             } barrier;
+            struct {
+                int attCount;
+                GLenum att[3];
+            } invalidateFramebuffer;
         } args;
     };
 
diff --git a/src/gui/rhi/qrhimetal.mm b/src/gui/rhi/qrhimetal.mm
index 7530186f11..6302a958c9 100644
--- a/src/gui/rhi/qrhimetal.mm
+++ b/src/gui/rhi/qrhimetal.mm
@@ -4226,7 +4226,7 @@ bool QMetalTextureRenderTarget::create()
             QMetalTexture *depthTexD = QRHI_RES(QMetalTexture, m_desc.depthTexture());
             d->fb.dsTex = depthTexD->d->tex;
             d->fb.hasStencil = rhiD->isStencilSupportingFormat(depthTexD->format());
-            d->fb.depthNeedsStore = true;
+            d->fb.depthNeedsStore = !m_flags.testFlag(DoNotStoreDepthStencilContents);
             if (d->colorAttCount == 0) {
                 d->pixelSize = depthTexD->pixelSize();
                 d->sampleCount = depthTexD->samples;
diff --git a/src/gui/rhi/qrhivulkan.cpp b/src/gui/rhi/qrhivulkan.cpp
index 2775530d6a..c009d22d04 100644
--- a/src/gui/rhi/qrhivulkan.cpp
+++ b/src/gui/rhi/qrhivulkan.cpp
@@ -1439,6 +1439,7 @@ bool QRhiVulkan::createOffscreenRenderPass(QVkRenderPassDescriptor *rpD,
                                            const QRhiColorAttachment *lastColorAttachment,
                                            bool preserveColor,
                                            bool preserveDs,
+                                           bool storeDs,
                                            QRhiRenderBuffer *depthStencilBuffer,
                                            QRhiTexture *depthTexture)
 {
@@ -1486,7 +1487,7 @@ bool QRhiVulkan::createOffscreenRenderPass(QVkRenderPassDescriptor *rpD,
         const VkSampleCountFlagBits samples = depthTexture ? QRHI_RES(QVkTexture, depthTexture)->samples
                                                            : QRHI_RES(QVkRenderBuffer, depthStencilBuffer)->samples;
         const VkAttachmentLoadOp loadOp = preserveDs ? VK_ATTACHMENT_LOAD_OP_LOAD : VK_ATTACHMENT_LOAD_OP_CLEAR;
-        const VkAttachmentStoreOp storeOp = depthTexture ? VK_ATTACHMENT_STORE_OP_STORE : VK_ATTACHMENT_STORE_OP_DONT_CARE;
+        const VkAttachmentStoreOp storeOp = storeDs ? VK_ATTACHMENT_STORE_OP_STORE : VK_ATTACHMENT_STORE_OP_DONT_CARE;
         VkAttachmentDescription attDesc = {};
         attDesc.format = dsFormat;
         attDesc.samples = samples;
@@ -6859,6 +6860,7 @@ QRhiRenderPassDescriptor *QVkTextureRenderTarget::newCompatibleRenderPassDescrip
                                          m_desc.cendColorAttachments(),
                                          m_flags.testFlag(QRhiTextureRenderTarget::PreserveColorContents),
                                          m_flags.testFlag(QRhiTextureRenderTarget::PreserveDepthStencilContents),
+                                         m_desc.depthTexture() && !m_flags.testFlag(DoNotStoreDepthStencilContents),
                                          m_desc.depthStencilBuffer(),
                                          m_desc.depthTexture()))
     {
diff --git a/src/gui/rhi/qrhivulkan_p.h b/src/gui/rhi/qrhivulkan_p.h
index a364634676..bdd65e177b 100644
--- a/src/gui/rhi/qrhivulkan_p.h
+++ b/src/gui/rhi/qrhivulkan_p.h
@@ -773,6 +773,7 @@ public:
                                    const QRhiColorAttachment *lastColorAttachment,
                                    bool preserveColor,
                                    bool preserveDs,
+                                   bool storeDs,
                                    QRhiRenderBuffer *depthStencilBuffer,
                                    QRhiTexture *depthTexture);
     bool ensurePipelineCache(const void *initialData = nullptr, size_t initialDataSize = 0);