Implement GL_ARB_shader_texture_image_samples. Also add in gl_MaxSamples and the float imageAtomicExchange.

git-svn-id: https://cvs.khronos.org/svn/repos/ogl/trunk/ecosystem/public/sdk/tools/glslang@27721 e7fa87d3-cd2b-0410-9028-fcbf551c1848

StandAlone/StandAlone.cpp

@@ -199,6 +199,7 @@ const char* DefaultConfig =
     "MaxTransformFeedbackInterleavedComponents 64\n"
     "MaxCullDistances 8\n"
     "MaxCombinedClipAndCullDistances 8\n"
+    "MaxSamples 4\n"
 
     "nonInductiveForLoops 1\n"
     "whileLoops 1\n"
@@ -407,6 +408,8 @@ void ProcessConfigFile()
             Resources.maxCullDistances = value;
         else if (strcmp(token, "MaxCombinedClipAndCullDistances") == 0)
             Resources.maxCombinedClipAndCullDistances = value;
+        else if (strcmp(token, "MaxSamples") == 0)
+            Resources.maxSamples = value;
 
         else if (strcmp(token, "nonInductiveForLoops") == 0)
             Resources.limits.nonInductiveForLoops = (value != 0);

Test/430.vert

@@ -157,3 +157,26 @@ void fooBarrier()
 }
 
 buffer vec4 v;  // ERROR
+
+uniform sampler2DMS s2dms;
+uniform usampler2DMSArray us2dmsa;
+layout(rgba32i) uniform iimage2DMS ii2dms;
+layout(rgba32f) uniform image2DMSArray i2dmsa;
+
+void fooq()
+{
+    int s = textureSamples(s2dms); // ERROR
+    s += textureSamples(us2dmsa);  // ERROR
+    s += imageSamples(ii2dms);     // ERROR
+    s += imageSamples(i2dmsa);     // ERROR
+}
+
+#extension GL_ARB_shader_texture_image_samples : enable
+
+void fooq2()
+{
+    int s = textureSamples(s2dms);
+    s += textureSamples(us2dmsa); 
+    s += imageSamples(ii2dms);    
+    s += imageSamples(i2dmsa);    
+}

Test/450.frag

@@ -16,7 +16,7 @@ void main()
     v4 = fwidthCoarse(in4) + fwidthFine(in4);
 
     float cull = gl_CullDistance[2];
-    float consts = gl_MaxCullDistances + gl_MaxCombinedClipAndCullDistances;
+    float consts = gl_MaxCullDistances + gl_MaxCombinedClipAndCullDistances + gl_MaxSamples;
 
     if (gl_HelperInvocation)
         ++v4;
@@ -33,3 +33,17 @@ void main()
     uint um = mix(uin, uin, b);
     ivec3 im3 = mix(ivec3(uin), ivec3(uin), bvec3(b));
 }
+
+uniform sampler2DMS s2dms;
+uniform usampler2DMSArray us2dmsa;
+layout(rgba32i) uniform iimage2DMS ii2dms;
+layout(rgba32f) uniform image2DMSArray i2dmsa;
+
+void foo()
+{
+    int s = textureSamples(s2dms);
+    s += textureSamples(us2dmsa);
+    s += imageSamples(ii2dms);
+    s += imageSamples(i2dmsa);
+    float f = imageAtomicExchange(i2dmsa, ivec3(in3), 2, 4.5);
+}

Test/baseResults/430.vert.out

@@ -49,11 +49,16 @@ ERROR: 0:150: 'barrier' : no matching overloaded function found
 ERROR: 0:154: 'memoryBarrierShared' : no matching overloaded function found 
 ERROR: 0:156: 'groupMemoryBarrier' : no matching overloaded function found 
 ERROR: 0:159: 'buffer' : buffers can be declared only as blocks 
-ERROR: 49 compilation errors.  No code generated.
+ERROR: 0:168: 'textureSamples and imageSamples' : not supported for this version or the enabled extensions 
+ERROR: 0:169: 'textureSamples and imageSamples' : not supported for this version or the enabled extensions 
+ERROR: 0:170: 'textureSamples and imageSamples' : not supported for this version or the enabled extensions 
+ERROR: 0:171: 'textureSamples and imageSamples' : not supported for this version or the enabled extensions 
+ERROR: 53 compilation errors.  No code generated.
 
 
 Shader version: 430
 Requested GL_ARB_enhanced_layouts
+Requested GL_ARB_shader_texture_image_samples
 in xfb mode
 ERROR: node is still EOpNull!
 0:14  Function Definition: foo( (void)
@@ -88,6 +93,46 @@ ERROR: node is still EOpNull!
 0:155      MemoryBarrierImage (void)
 0:156      Constant:
 0:156        0.000000
+0:166  Function Definition: fooq( (void)
+0:166    Function Parameters: 
+0:168    Sequence
+0:168      Sequence
+0:168        move second child to first child (int)
+0:168          's' (int)
+0:168          Function Call: textureSamples(s21; (int)
+0:168            's2dms' (uniform sampler2DMS)
+0:169      add second child into first child (int)
+0:169        's' (int)
+0:169        Function Call: textureSamples(usA21; (int)
+0:169          'us2dmsa' (uniform usampler2DMSArray)
+0:170      add second child into first child (int)
+0:170        's' (int)
+0:170        Function Call: imageSamples(iI21; (int)
+0:170          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:171      add second child into first child (int)
+0:171        's' (int)
+0:171        Function Call: imageSamples(IA21; (int)
+0:171          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:176  Function Definition: fooq2( (void)
+0:176    Function Parameters: 
+0:178    Sequence
+0:178      Sequence
+0:178        move second child to first child (int)
+0:178          's' (int)
+0:178          Function Call: textureSamples(s21; (int)
+0:178            's2dms' (uniform sampler2DMS)
+0:179      add second child into first child (int)
+0:179        's' (int)
+0:179        Function Call: textureSamples(usA21; (int)
+0:179          'us2dmsa' (uniform usampler2DMSArray)
+0:180      add second child into first child (int)
+0:180        's' (int)
+0:180        Function Call: imageSamples(iI21; (int)
+0:180          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:181      add second child into first child (int)
+0:181        's' (int)
+0:181        Function Call: imageSamples(IA21; (int)
+0:181          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 0:?   Linker Objects
 0:?     'v4' (layout(location=3 ) 4-component vector of float)
 0:?     'uv4' (layout(location=4 ) uniform 4-component vector of float)
@@ -129,6 +174,10 @@ ERROR: node is still EOpNull!
 0:?     'bbinst5' (out block{layout(xfb_buffer=1 xfb_offset=0 ) out 4-component vector of float bbv1, layout(xfb_buffer=1 xfb_offset=64 xfb_stride=80 ) out 4-component vector of float bbv2})
 0:?     'sharedv' (shared 4-component vector of float)
 0:?     'v' (buffer 4-component vector of float)
+0:?     's2dms' (uniform sampler2DMS)
+0:?     'us2dmsa' (uniform usampler2DMSArray)
+0:?     'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:?     'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 0:?     'gl_VertexID' (gl_VertexId int)
 0:?     'gl_InstanceID' (gl_InstanceId int)
 
@@ -141,6 +190,7 @@ ERROR:     xfb_buffer 3, xfb_stride 64, minimum stride needed: 80
 
 Shader version: 430
 Requested GL_ARB_enhanced_layouts
+Requested GL_ARB_shader_texture_image_samples
 in xfb mode
 ERROR: node is still EOpNull!
 0:14  Function Definition: foo( (void)
@@ -175,6 +225,46 @@ ERROR: node is still EOpNull!
 0:155      MemoryBarrierImage (void)
 0:156      Constant:
 0:156        0.000000
+0:166  Function Definition: fooq( (void)
+0:166    Function Parameters: 
+0:168    Sequence
+0:168      Sequence
+0:168        move second child to first child (int)
+0:168          's' (int)
+0:168          Function Call: textureSamples(s21; (int)
+0:168            's2dms' (uniform sampler2DMS)
+0:169      add second child into first child (int)
+0:169        's' (int)
+0:169        Function Call: textureSamples(usA21; (int)
+0:169          'us2dmsa' (uniform usampler2DMSArray)
+0:170      add second child into first child (int)
+0:170        's' (int)
+0:170        Function Call: imageSamples(iI21; (int)
+0:170          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:171      add second child into first child (int)
+0:171        's' (int)
+0:171        Function Call: imageSamples(IA21; (int)
+0:171          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:176  Function Definition: fooq2( (void)
+0:176    Function Parameters: 
+0:178    Sequence
+0:178      Sequence
+0:178        move second child to first child (int)
+0:178          's' (int)
+0:178          Function Call: textureSamples(s21; (int)
+0:178            's2dms' (uniform sampler2DMS)
+0:179      add second child into first child (int)
+0:179        's' (int)
+0:179        Function Call: textureSamples(usA21; (int)
+0:179          'us2dmsa' (uniform usampler2DMSArray)
+0:180      add second child into first child (int)
+0:180        's' (int)
+0:180        Function Call: imageSamples(iI21; (int)
+0:180          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:181      add second child into first child (int)
+0:181        's' (int)
+0:181        Function Call: imageSamples(IA21; (int)
+0:181          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 0:?   Linker Objects
 0:?     'v4' (layout(location=3 ) 4-component vector of float)
 0:?     'uv4' (layout(location=4 ) uniform 4-component vector of float)
@@ -216,6 +306,10 @@ ERROR: node is still EOpNull!
 0:?     'bbinst5' (out block{layout(xfb_buffer=1 xfb_offset=0 ) out 4-component vector of float bbv1, layout(xfb_buffer=1 xfb_offset=64 xfb_stride=80 ) out 4-component vector of float bbv2})
 0:?     'sharedv' (shared 4-component vector of float)
 0:?     'v' (buffer 4-component vector of float)
+0:?     's2dms' (uniform sampler2DMS)
+0:?     'us2dmsa' (uniform usampler2DMSArray)
+0:?     'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:?     'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 0:?     'gl_VertexID' (gl_VertexId int)
 0:?     'gl_InstanceID' (gl_InstanceId int)
 

Test/baseResults/450.frag.out

@@ -58,7 +58,7 @@ Shader version: 450
 0:19        move second child to first child (float)
 0:19          'consts' (float)
 0:19          Constant:
-0:19            16.000000
+0:19            20.000000
 0:21      Test condition and select (void)
 0:21        Condition
 0:21        'gl_HelperInvocation' (in bool)
@@ -99,12 +99,47 @@ Shader version: 450
 0:34                'uin' (uint)
 0:34            Construct bvec3 (3-component vector of bool)
 0:34              'b' (bool)
+0:42  Function Definition: foo( (void)
+0:42    Function Parameters: 
+0:44    Sequence
+0:44      Sequence
+0:44        move second child to first child (int)
+0:44          's' (int)
+0:44          Function Call: textureSamples(s21; (int)
+0:44            's2dms' (uniform sampler2DMS)
+0:45      add second child into first child (int)
+0:45        's' (int)
+0:45        Function Call: textureSamples(usA21; (int)
+0:45          'us2dmsa' (uniform usampler2DMSArray)
+0:46      add second child into first child (int)
+0:46        's' (int)
+0:46        Function Call: imageSamples(iI21; (int)
+0:46          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:47      add second child into first child (int)
+0:47        's' (int)
+0:47        Function Call: imageSamples(IA21; (int)
+0:47          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:48      Sequence
+0:48        move second child to first child (float)
+0:48          'f' (float)
+0:48          Function Call: imageAtomicExchange(IA21;vi3;i1;f1; (float)
+0:48            'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:48            Convert float to int (3-component vector of int)
+0:48              'in3' (smooth in 3-component vector of float)
+0:48            Constant:
+0:48              2 (const int)
+0:48            Constant:
+0:48              4.500000
 0:?   Linker Objects
 0:?     'in1' (smooth in float)
 0:?     'in2' (smooth in 2-component vector of float)
 0:?     'in3' (smooth in 3-component vector of float)
 0:?     'in4' (smooth in 4-component vector of float)
 0:?     'gl_CullDistance' (smooth in implicitly-sized array of float)
+0:?     's2dms' (uniform sampler2DMS)
+0:?     'us2dmsa' (uniform usampler2DMSArray)
+0:?     'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:?     'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 
 
 Linked fragment stage:
@@ -167,7 +202,7 @@ Shader version: 450
 0:19        move second child to first child (float)
 0:19          'consts' (float)
 0:19          Constant:
-0:19            16.000000
+0:19            20.000000
 0:21      Test condition and select (void)
 0:21        Condition
 0:21        'gl_HelperInvocation' (in bool)
@@ -208,10 +243,45 @@ Shader version: 450
 0:34                'uin' (uint)
 0:34            Construct bvec3 (3-component vector of bool)
 0:34              'b' (bool)
+0:42  Function Definition: foo( (void)
+0:42    Function Parameters: 
+0:44    Sequence
+0:44      Sequence
+0:44        move second child to first child (int)
+0:44          's' (int)
+0:44          Function Call: textureSamples(s21; (int)
+0:44            's2dms' (uniform sampler2DMS)
+0:45      add second child into first child (int)
+0:45        's' (int)
+0:45        Function Call: textureSamples(usA21; (int)
+0:45          'us2dmsa' (uniform usampler2DMSArray)
+0:46      add second child into first child (int)
+0:46        's' (int)
+0:46        Function Call: imageSamples(iI21; (int)
+0:46          'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:47      add second child into first child (int)
+0:47        's' (int)
+0:47        Function Call: imageSamples(IA21; (int)
+0:47          'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:48      Sequence
+0:48        move second child to first child (float)
+0:48          'f' (float)
+0:48          Function Call: imageAtomicExchange(IA21;vi3;i1;f1; (float)
+0:48            'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
+0:48            Convert float to int (3-component vector of int)
+0:48              'in3' (smooth in 3-component vector of float)
+0:48            Constant:
+0:48              2 (const int)
+0:48            Constant:
+0:48              4.500000
 0:?   Linker Objects
 0:?     'in1' (smooth in float)
 0:?     'in2' (smooth in 2-component vector of float)
 0:?     'in3' (smooth in 3-component vector of float)
 0:?     'in4' (smooth in 4-component vector of float)
 0:?     'gl_CullDistance' (smooth in 3-element array of float)
+0:?     's2dms' (uniform sampler2DMS)
+0:?     'us2dmsa' (uniform usampler2DMSArray)
+0:?     'ii2dms' (layout(rgba32i ) uniform iimage2DMS)
+0:?     'i2dmsa' (layout(rgba32f ) uniform image2DMSArray)
 

Test/baseResults/test.conf

@@ -80,6 +80,7 @@ MaxTransformFeedbackBuffers 4
 MaxTransformFeedbackInterleavedComponents 64
 MaxCullDistances 8
 MaxCombinedClipAndCullDistances 8
+MaxSamples 4
 nonInductiveForLoops 1
 whileLoops 1
 doWhileLoops 1

glslang/Include/ResourceLimits.h

@@ -132,6 +132,7 @@ struct TBuiltInResource {
     int maxTransformFeedbackInterleavedComponents;
     int maxCullDistances;
     int maxCombinedClipAndCullDistances;
+    int maxSamples;
 
     TLimits limits;
 };

glslang/MachineIndependent/Initialize.cpp

@@ -1697,6 +1697,18 @@ void TBuiltIns::addQueryFunctions(TSampler sampler, TString& typeName, int versi
         commonBuiltins.append(",int);\n");
     else
         commonBuiltins.append(");\n");
+
+    // GL_ARB_shader_texture_image_samples
+    // TODO: spec issue? there are no memory qualifiers; how to query a writeonly/readonly image, etc?
+    if (profile != EEsProfile && version >= 430 && sampler.ms) {
+        commonBuiltins.append("int ");
+        if (sampler.image)
+            commonBuiltins.append("imageSamples(");
+        else
+            commonBuiltins.append("textureSamples(");
+        commonBuiltins.append(typeName);
+        commonBuiltins.append(");\n");
+    }
 }
 
 //
@@ -1764,6 +1776,15 @@ void TBuiltIns::addImageFunctions(TSampler sampler, TString& typeName, int versi
             commonBuiltins.append(", ");
             commonBuiltins.append(dataType);
             commonBuiltins.append(");\n");
+        } else {
+            // not int or uint
+            // GL_ARB_ES3_1_compatibility
+            // TODO: spec issue: are there restrictions on the kind of layout() that can be used?  what about dropping memory qualifiers?
+            if (version >= 450) {
+                commonBuiltins.append("float imageAtomicExchange(coherent ");
+                commonBuiltins.append(imageParams);
+                commonBuiltins.append(", float);\n");
+            }
         }
     }
 }
@@ -2356,6 +2377,8 @@ void TBuiltIns::initialize(const TBuiltInResource &resources, int version, EProf
         s.append(builtInConstant);
         snprintf(builtInConstant, maxSize, "const int gl_MaxGeometryAtomicCounterBuffers = %d;", resources.       maxGeometryAtomicCounterBuffers);
         s.append(builtInConstant);
+
+        s.append("\n");
     }
 
 
@@ -2380,6 +2403,8 @@ void TBuiltIns::initialize(const TBuiltInResource &resources, int version, EProf
         s.append(builtInConstant);
         snprintf(builtInConstant, maxSize, "const int gl_MaxComputeAtomicCounterBuffers = %d;", resources.maxComputeAtomicCounterBuffers);
         s.append(builtInConstant);
+
+        s.append("\n");
     }
 
     // GL_ARB_cull_distance
@@ -2390,6 +2415,12 @@ void TBuiltIns::initialize(const TBuiltInResource &resources, int version, EProf
         s.append(builtInConstant);
     }
 
+    // GL_ARB_ES3_1_compatibility
+    if (profile != EEsProfile && version >= 450) {
+        snprintf(builtInConstant, maxSize, "const int gl_MaxSamples = %d;", resources.maxSamples);
+        s.append(builtInConstant);
+    }
+
     s.append("\n");
 }
 

glslang/MachineIndependent/ParseHelper.cpp

@@ -1296,12 +1296,17 @@ void TParseContext::nonOpBuiltInCheck(TSourceLoc loc, const TFunction& fnCandida
             }
         }
     }
+
+    // GL_ARB_shader_texture_image_samples
+    if (fnCandidate.getName().compare(0, 14, "textureSamples") == 0 || fnCandidate.getName().compare(0, 12, "imageSamples") == 0)
+        profileRequires(loc, ~EEsProfile, 450, GL_ARB_shader_texture_image_samples, "textureSamples and imageSamples");
+
     if (fnCandidate.getName().compare(0, 11, "imageAtomic") == 0) {
         const TType& imageType = callNode.getSequence()[0]->getAsTyped()->getType();
         if (imageType.getSampler().type == EbtInt || imageType.getSampler().type == EbtUint) {
             if (imageType.getQualifier().layoutFormat != ElfR32i && imageType.getQualifier().layoutFormat != ElfR32ui)
                 error(loc, "only supported on image with format r32i or r32ui", fnCandidate.getName().c_str(), "");
-        } else
+        } else if (fnCandidate.getName().compare(0, 19, "imageAtomicExchange") != 0) 
             error(loc, "only supported on integer images", fnCandidate.getName().c_str(), "");
     }
 }

glslang/MachineIndependent/Versions.cpp

@@ -169,6 +169,7 @@ void TParseContext::initializeExtensionBehavior()
     extensionBehavior[GL_ARB_shader_image_load_store]      = EBhDisable;
     extensionBehavior[GL_ARB_shader_atomic_counters]       = EBhDisable;
     extensionBehavior[GL_ARB_derivative_control]           = EBhDisable;
+    extensionBehavior[GL_ARB_shader_texture_image_samples] = EBhDisable;
 //    extensionBehavior[GL_ARB_cull_distance]            = EBhDisable;    // present for 4.5, but need extension control over block members
 }
 
@@ -213,6 +214,7 @@ const char* TParseContext::getPreamble()
             "#define GL_ARB_shader_image_load_store 1\n"
             "#define GL_ARB_shader_atomic_counters 1\n"
             "#define GL_ARB_derivative_control 1\n"
+            "#define GL_ARB_shader_texture_image_samples 1\n"
 //            "#define GL_ARB_cull_distance 1\n"    // present for 4.5, but need extension control over block members
             ;
     }

glslang/MachineIndependent/Versions.h

@@ -93,6 +93,7 @@ const char* const GL_ARB_explicit_attrib_location = "GL_ARB_explicit_attrib_loca
 const char* const GL_ARB_shader_image_load_store      = "GL_ARB_shader_image_load_store";
 const char* const GL_ARB_shader_atomic_counters       = "GL_ARB_shader_atomic_counters";
 const char* const GL_ARB_derivative_control           = "GL_ARB_derivative_control";
+const char* const GL_ARB_shader_texture_image_samples = "GL_ARB_shader_texture_image_samples";
 //const char* const GL_ARB_cull_distance            = "GL_ARB_cull_distance";  // present for 4.5, but need extension control over block members
 
 } // end namespace glslang