From eb43fb585e8aacf9383280c5b48f9dce5627b524 Mon Sep 17 00:00:00 2001
From: clach <clach@users.noreply.github.com>
Date: Tue, 25 Jan 2022 23:46:48 -0800
Subject: [PATCH] [hdSt, hgi*] Exclude double precision transform overloads
 from the Storm volume shader when double precision values are not supported
 (e.g. using Metal). A new Hgi capability bit,
 HgiDeviceCapabilitiesBitsShaderDoublePrecision, is added, which codegen will
 use to create a shader define if needed.

(Internal change: 2211644)
---
 pxr/imaging/hdSt/codeGen.cpp           |  6 ++++++
 pxr/imaging/hdSt/shaders/volume.glslfx |  4 ++++
 pxr/imaging/hgi/enums.h                | 19 +++++++++++--------
 pxr/imaging/hgiGL/capabilities.cpp     |  2 ++
 pxr/imaging/hgiMetal/capabilities.mm   |  2 ++
 5 files changed, 25 insertions(+), 8 deletions(-)

diff --git a/pxr/imaging/hdSt/codeGen.cpp b/pxr/imaging/hdSt/codeGen.cpp
index 4e66c29f40..83fa9dbbd5 100644
--- a/pxr/imaging/hdSt/codeGen.cpp
+++ b/pxr/imaging/hdSt/codeGen.cpp
@@ -533,6 +533,9 @@ HdSt_CodeGen::Compile(HdStResourceRegistry*const registry)
         GetCapabilities()->IsSet(HgiDeviceCapabilitiesBitsShaderDrawParameters);
     const bool builtinBarycentricsEnabled = registry->GetHgi()->
         GetCapabilities()->IsSet(HgiDeviceCapabilitiesBitsBuiltinBarycentrics);
+    const bool doublePrecisionEnabled = registry->GetHgi()->
+        GetCapabilities()->IsSet(
+            HgiDeviceCapabilitiesBitsShaderDoublePrecision);
 
     if (bindlessBufferEnabled) {
         _genHeader << "#extension GL_NV_shader_buffer_load : require\n"
@@ -575,6 +578,9 @@ HdSt_CodeGen::Compile(HdStResourceRegistry*const registry)
     if (_materialTag == HdStMaterialTagTokens->masked) {
         _genFS << "#define HD_MATERIAL_TAG_MASKED 1\n";
     }
+    if (doublePrecisionEnabled) {
+        _genFS << "#define HD_SHADER_SUPPORTS_DOUBLE_PRECISION\n";
+    }
 
     // ------------------
     // Custom Buffer Bindings
diff --git a/pxr/imaging/hdSt/shaders/volume.glslfx b/pxr/imaging/hdSt/shaders/volume.glslfx
index 9b27dbbef8..6054b80f97 100644
--- a/pxr/imaging/hdSt/shaders/volume.glslfx
+++ b/pxr/imaging/hdSt/shaders/volume.glslfx
@@ -98,12 +98,14 @@ transformPoint(mat4 m, vec3 point)
     return result.xyz / result.w;
 }
 
+#ifdef HD_SHADER_SUPPORTS_DOUBLE_PRECISION   
 vec3
 transformPoint(dmat4 m, vec3 point)
 {
     const vec4 result = vec4(m * vec4(point, 1.0));
     return result.xyz / result.w;
 }
+#endif
 
 // Transform a direction by a 4x4 matrix
 vec3
@@ -113,12 +115,14 @@ transformDir(mat4 m, vec3 dir)
     return result.xyz;
 }
 
+#ifdef HD_SHADER_SUPPORTS_DOUBLE_PRECISION   
 vec3
 transformDir(dmat4 m, vec3 dir)
 {
     const vec4 result = vec4(m * vec4(dir, 0.0));
     return result.xyz;
 }
+#endif
 
 // Compute time when a ray starting at pos with direction dir
 // exits the axis-aligned box with vertices lMin and lMax.
diff --git a/pxr/imaging/hgi/enums.h b/pxr/imaging/hgi/enums.h
index 0796c564fa..edf4bc9c82 100644
--- a/pxr/imaging/hgi/enums.h
+++ b/pxr/imaging/hgi/enums.h
@@ -55,18 +55,21 @@ using HgiBits = uint32_t;
 ///   The device supports multiple primitive, indirect drawing</li>
 /// <li>HgiDeviceCapabilitiesBitsBindlessTextures:
 ///   The device can access GPU textures using bindless handles</li>
+/// <li>HgiDeviceCapabilitiesBitsShaderDoublePrecision:
+///   The device supports double precision types in shaders</li>
 /// </ul>
 ///
 enum HgiDeviceCapabilitiesBits : HgiBits
 {
-    HgiDeviceCapabilitiesBitsPresentation         = 1 << 0,
-    HgiDeviceCapabilitiesBitsBindlessBuffers      = 1 << 1,
-    HgiDeviceCapabilitiesBitsConcurrentDispatch   = 1 << 2,
-    HgiDeviceCapabilitiesBitsUnifiedMemory        = 1 << 3,
-    HgiDeviceCapabilitiesBitsBuiltinBarycentrics  = 1 << 4,
-    HgiDeviceCapabilitiesBitsShaderDrawParameters = 1 << 5,
-    HgiDeviceCapabilitiesBitsMultiDrawIndirect    = 1 << 6,
-    HgiDeviceCapabilitiesBitsBindlessTextures     = 1 << 7
+    HgiDeviceCapabilitiesBitsPresentation          = 1 << 0,
+    HgiDeviceCapabilitiesBitsBindlessBuffers       = 1 << 1,
+    HgiDeviceCapabilitiesBitsConcurrentDispatch    = 1 << 2,
+    HgiDeviceCapabilitiesBitsUnifiedMemory         = 1 << 3,
+    HgiDeviceCapabilitiesBitsBuiltinBarycentrics   = 1 << 4,
+    HgiDeviceCapabilitiesBitsShaderDrawParameters  = 1 << 5,
+    HgiDeviceCapabilitiesBitsMultiDrawIndirect     = 1 << 6,
+    HgiDeviceCapabilitiesBitsBindlessTextures      = 1 << 7,
+    HgiDeviceCapabilitiesBitsShaderDoublePrecision = 1 << 8
 };
 
 using HgiDeviceCapabilities = HgiBits;
diff --git a/pxr/imaging/hgiGL/capabilities.cpp b/pxr/imaging/hgiGL/capabilities.cpp
index a9a8cb158a..5c413fb440 100644
--- a/pxr/imaging/hgiGL/capabilities.cpp
+++ b/pxr/imaging/hgiGL/capabilities.cpp
@@ -181,6 +181,8 @@ HgiGLCapabilities::_LoadCapabilities()
         builtinBarycentricsEnabled);
     _SetFlag(HgiDeviceCapabilitiesBitsShaderDrawParameters, 
         shaderDrawParametersEnabled);
+    _SetFlag(HgiDeviceCapabilitiesBitsShaderDoublePrecision, 
+        true);
 
     if (TfDebug::IsEnabled(HGI_DEBUG_DEVICE_CAPABILITIES)) {
         std::cout
diff --git a/pxr/imaging/hgiMetal/capabilities.mm b/pxr/imaging/hgiMetal/capabilities.mm
index 9dc10b10a6..b6537a6ead 100644
--- a/pxr/imaging/hgiMetal/capabilities.mm
+++ b/pxr/imaging/hgiMetal/capabilities.mm
@@ -51,6 +51,8 @@
 
     _SetFlag(HgiDeviceCapabilitiesBitsBuiltinBarycentrics, true);
 
+    _SetFlag(HgiDeviceCapabilitiesBitsShaderDoublePrecision, false);
+
 #if defined(ARCH_OS_MACOS)
     if (!unifiedMemory) {
         defaultStorageMode = MTLResourceStorageModeManaged;