Add support for SW primitive statistics counting

For pre-GFX11, the primitive statistics counting is performed by HW via
the same mechanism of transform feedback. For GFX11+, since transform
feedback is done by SW emulation, the primitive statistics counting will
follow the same handling.

We add a new handler collectPrimitiveStats to deal with it. It is a
reduced version of SW transform feedback, only updating HW counters of
requested vertex streams. We don't merge it with SW transform feedback
because this will make the logic blurry. Indeed, some duplicated codes
are the trade-off.

For non-GS case, only the counter of stream 0 will be updated and
generated primitive count is passed by GE (we don't modify it). For GS
case, counters of all active vertex streams will be updated and we must
calculate generated primitive count first before doing such update. The
calculation is to count valid primitive mask bits in this NGG subgroup
and add them together.

The implementation of this PR is the foundation of
VK_EXT_primitives_generated_query.

lgc/builder/MiscBuilder.cpp

@@ -46,8 +46,10 @@ using namespace llvm;
 Instruction *BuilderImpl::CreateEmitVertex(unsigned streamId) {
   assert(m_shaderStage == ShaderStageGeometry);
 
-  // Mark this vertex stream as active if transform feedback is enabled or this is the rasterization stream.
-  if (m_pipelineState->enableXfb() || m_pipelineState->getRasterizerState().rasterStream == streamId)
+  // Mark this vertex stream as active if transform feedback is enabled, or primitive statistics counting is enabled,
+  // or this is the rasterization stream.
+  if (m_pipelineState->enableXfb() || m_pipelineState->enablePrimStats() ||
+      m_pipelineState->getRasterizerState().rasterStream == streamId)
     m_pipelineState->setVertexStreamActive(streamId);
 
   // Get GsWaveId
@@ -68,8 +70,10 @@ Instruction *BuilderImpl::CreateEmitVertex(unsigned streamId) {
 Instruction *BuilderImpl::CreateEndPrimitive(unsigned streamId) {
   assert(m_shaderStage == ShaderStageGeometry);
 
-  // Mark this vertex stream as active if transform feedback is enabled or this is the rasterization stream.
-  if (m_pipelineState->enableXfb() || m_pipelineState->getRasterizerState().rasterStream == streamId)
+  // Mark this vertex stream as active if transform feedback is enabled, or primitive statistics counting is enabled,
+  // or this is the rasterization stream.
+  if (m_pipelineState->enableXfb() || m_pipelineState->enablePrimStats() ||
+      m_pipelineState->getRasterizerState().rasterStream == streamId)
     m_pipelineState->setVertexStreamActive(streamId);
 
   // Get GsWaveId

lgc/include/lgc/state/PipelineState.h

@@ -396,7 +396,7 @@ class PipelineState final : public Pipeline {
   // Check if transform feedback is active
   bool enableXfb() const { return m_xfbStateMetadata.enableXfb; }
 
-  // Check if we need count primitives if XFB is disabled
+  // Check if we need primitive statistics counting
   bool enablePrimStats() const { return m_xfbStateMetadata.enablePrimStats; }
 
   // Get transform feedback strides