Generate separate stores for partially swizzled memory stores
Full vector and fully specified vector swizzle stores are not affected by this change, only partial swizzles ie swizzles with fewer components than the vector being stored to. Previously the vector being stored to loaded and any components not specified in the swizzle were used to create a full store to the vector. While this change generates more SPIR-V instructions, it is necessary for correctness. Fixes #2518.
This commit is contained in:
Jeremy Hayes
parent
9158061398
commit
6d5b40f051
50 changed files with 31343 additions and 26594 deletions
|
|
@ -1,7 +1,7 @@
|
|||
spv.400.frag
|
||||
// Module Version 10000
|
||||
// Generated by (magic number): 8000a
|
||||
// Id's are bound by 1118
|
||||
// Id's are bound by 1122
|
||||
|
||||
Capability Shader
|
||||
Capability Geometry
|
||||
|
|
@ -11,7 +11,7 @@ spv.400.frag
|
|||
Capability SampledRect
|
||||
1: ExtInstImport "GLSL.std.450"
|
||||
MemoryModel Logical GLSL450
|
||||
EntryPoint Fragment 4 "main" 13 1027 1033 1038 1050 1076 1097 1099 1105 1107 1116
|
||||
EntryPoint Fragment 4 "main" 13 1027 1033 1038 1054 1080 1101 1103 1109 1111 1120
|
||||
ExecutionMode 4 OriginUpperLeft
|
||||
Source GLSL 400
|
||||
SourceExtension "GL_ARB_separate_shader_objects"
|
||||
|
|
@ -42,16 +42,16 @@ spv.400.frag
|
|||
Name 1027 "i"
|
||||
Name 1033 "c2D"
|
||||
Name 1038 "gl_ClipDistance"
|
||||
Name 1050 "uoutp"
|
||||
Name 1054 "samp2dr"
|
||||
Name 1076 "ioutp"
|
||||
Name 1080 "isamp2DA"
|
||||
Name 1097 "gl_FragCoord"
|
||||
Name 1099 "vl2"
|
||||
Name 1105 "uo"
|
||||
Name 1107 "u"
|
||||
Name 1115 "id"
|
||||
Name 1116 "gl_PrimitiveID"
|
||||
Name 1054 "uoutp"
|
||||
Name 1058 "samp2dr"
|
||||
Name 1080 "ioutp"
|
||||
Name 1084 "isamp2DA"
|
||||
Name 1101 "gl_FragCoord"
|
||||
Name 1103 "vl2"
|
||||
Name 1109 "uo"
|
||||
Name 1111 "u"
|
||||
Name 1119 "id"
|
||||
Name 1120 "gl_PrimitiveID"
|
||||
Decorate 13(outp) Location 1
|
||||
Decorate 17(u2drs) DescriptorSet 0
|
||||
Decorate 17(u2drs) Binding 3
|
||||
|
|
@ -61,19 +61,19 @@ spv.400.frag
|
|||
Decorate 1027(i) Location 1
|
||||
Decorate 1033(c2D) Location 0
|
||||
Decorate 1038(gl_ClipDistance) BuiltIn ClipDistance
|
||||
Decorate 1050(uoutp) Location 3
|
||||
Decorate 1054(samp2dr) DescriptorSet 0
|
||||
Decorate 1054(samp2dr) Binding 1
|
||||
Decorate 1076(ioutp) Location 2
|
||||
Decorate 1080(isamp2DA) DescriptorSet 0
|
||||
Decorate 1080(isamp2DA) Binding 2
|
||||
Decorate 1097(gl_FragCoord) BuiltIn FragCoord
|
||||
Decorate 1099(vl2) Location 6
|
||||
Decorate 1105(uo) Location 0
|
||||
Decorate 1107(u) Flat
|
||||
Decorate 1107(u) Location 2
|
||||
Decorate 1116(gl_PrimitiveID) Flat
|
||||
Decorate 1116(gl_PrimitiveID) BuiltIn PrimitiveId
|
||||
Decorate 1054(uoutp) Location 3
|
||||
Decorate 1058(samp2dr) DescriptorSet 0
|
||||
Decorate 1058(samp2dr) Binding 1
|
||||
Decorate 1080(ioutp) Location 2
|
||||
Decorate 1084(isamp2DA) DescriptorSet 0
|
||||
Decorate 1084(isamp2DA) Binding 2
|
||||
Decorate 1101(gl_FragCoord) BuiltIn FragCoord
|
||||
Decorate 1103(vl2) Location 6
|
||||
Decorate 1109(uo) Location 0
|
||||
Decorate 1111(u) Flat
|
||||
Decorate 1111(u) Location 2
|
||||
Decorate 1120(gl_PrimitiveID) Flat
|
||||
Decorate 1120(gl_PrimitiveID) BuiltIn PrimitiveId
|
||||
2: TypeVoid
|
||||
3: TypeFunction 2
|
||||
10: TypeFloat 32
|
||||
|
|
@ -161,46 +161,46 @@ spv.400.frag
|
|||
1038(gl_ClipDistance): 1037(ptr) Variable Input
|
||||
1039: TypePointer Input 10(float)
|
||||
1043: TypeVector 10(float) 3
|
||||
1048: TypeVector 32(int) 4
|
||||
1049: TypePointer Output 1048(ivec4)
|
||||
1050(uoutp): 1049(ptr) Variable Output
|
||||
1051: TypeImage 32(int) Rect sampled format:Unknown
|
||||
1052: TypeSampledImage 1051
|
||||
1053: TypePointer UniformConstant 1052
|
||||
1054(samp2dr): 1053(ptr) Variable UniformConstant
|
||||
1057: 32(int) Constant 4
|
||||
1058: TypeArray 24(ivec2) 1057
|
||||
1059: 24(ivec2) ConstantComposite 966 970
|
||||
1060: 23(int) Constant 15
|
||||
1061: 23(int) Constant 16
|
||||
1062: 24(ivec2) ConstantComposite 1060 1061
|
||||
1063: 23(int) Constant 4294967294
|
||||
1064: 23(int) Constant 0
|
||||
1065: 24(ivec2) ConstantComposite 1063 1064
|
||||
1066: 1058 ConstantComposite 1059 27 1062 1065
|
||||
1074: TypeVector 23(int) 4
|
||||
1075: TypePointer Output 1074(ivec4)
|
||||
1076(ioutp): 1075(ptr) Variable Output
|
||||
1077: TypeImage 23(int) 2D array sampled format:Unknown
|
||||
1078: TypeSampledImage 1077
|
||||
1079: TypePointer UniformConstant 1078
|
||||
1080(isamp2DA): 1079(ptr) Variable UniformConstant
|
||||
1082: 10(float) Constant 1036831949
|
||||
1083: 1043(fvec3) ConstantComposite 1082 1082 1082
|
||||
1084: 24(ivec2) ConstantComposite 966 966
|
||||
1096: TypePointer Input 11(fvec4)
|
||||
1097(gl_FragCoord): 1096(ptr) Variable Input
|
||||
1099(vl2): 1096(ptr) Variable Input
|
||||
1104: TypePointer Output 32(int)
|
||||
1105(uo): 1104(ptr) Variable Output
|
||||
1106: TypePointer Input 32(int)
|
||||
1107(u): 1106(ptr) Variable Input
|
||||
1114: TypePointer Function 23(int)
|
||||
1116(gl_PrimitiveID): 1026(ptr) Variable Input
|
||||
1052: TypeVector 32(int) 4
|
||||
1053: TypePointer Output 1052(ivec4)
|
||||
1054(uoutp): 1053(ptr) Variable Output
|
||||
1055: TypeImage 32(int) Rect sampled format:Unknown
|
||||
1056: TypeSampledImage 1055
|
||||
1057: TypePointer UniformConstant 1056
|
||||
1058(samp2dr): 1057(ptr) Variable UniformConstant
|
||||
1061: 32(int) Constant 4
|
||||
1062: TypeArray 24(ivec2) 1061
|
||||
1063: 24(ivec2) ConstantComposite 966 970
|
||||
1064: 23(int) Constant 15
|
||||
1065: 23(int) Constant 16
|
||||
1066: 24(ivec2) ConstantComposite 1064 1065
|
||||
1067: 23(int) Constant 4294967294
|
||||
1068: 23(int) Constant 0
|
||||
1069: 24(ivec2) ConstantComposite 1067 1068
|
||||
1070: 1062 ConstantComposite 1063 27 1066 1069
|
||||
1078: TypeVector 23(int) 4
|
||||
1079: TypePointer Output 1078(ivec4)
|
||||
1080(ioutp): 1079(ptr) Variable Output
|
||||
1081: TypeImage 23(int) 2D array sampled format:Unknown
|
||||
1082: TypeSampledImage 1081
|
||||
1083: TypePointer UniformConstant 1082
|
||||
1084(isamp2DA): 1083(ptr) Variable UniformConstant
|
||||
1086: 10(float) Constant 1036831949
|
||||
1087: 1043(fvec3) ConstantComposite 1086 1086 1086
|
||||
1088: 24(ivec2) ConstantComposite 966 966
|
||||
1100: TypePointer Input 11(fvec4)
|
||||
1101(gl_FragCoord): 1100(ptr) Variable Input
|
||||
1103(vl2): 1100(ptr) Variable Input
|
||||
1108: TypePointer Output 32(int)
|
||||
1109(uo): 1108(ptr) Variable Output
|
||||
1110: TypePointer Input 32(int)
|
||||
1111(u): 1110(ptr) Variable Input
|
||||
1118: TypePointer Function 23(int)
|
||||
1120(gl_PrimitiveID): 1026(ptr) Variable Input
|
||||
4(main): 2 Function None 3
|
||||
5: Label
|
||||
1019(v): 1018(ptr) Variable Function
|
||||
1115(id): 1114(ptr) Variable Function
|
||||
1119(id): 1118(ptr) Variable Function
|
||||
1028: 23(int) Load 1027(i)
|
||||
1030: 1029(ptr) AccessChain 1025(arrayedSampler) 1028
|
||||
1031: 1021 Load 1030
|
||||
|
|
@ -213,50 +213,56 @@ spv.400.frag
|
|||
Store 1042 1041
|
||||
1044: 11(fvec4) Load 1019(v)
|
||||
1045: 1043(fvec3) VectorShuffle 1044 1044 1 2 3
|
||||
1046: 11(fvec4) Load 13(outp)
|
||||
1047: 11(fvec4) VectorShuffle 1046 1045 0 4 5 6
|
||||
Store 13(outp) 1047
|
||||
1055: 1052 Load 1054(samp2dr)
|
||||
1056: 20(fvec2) Load 1033(c2D)
|
||||
1067: 1048(ivec4) ImageGather 1055 1056 970 ConstOffsets 1066
|
||||
Store 1050(uoutp) 1067
|
||||
1068: 1029(ptr) AccessChain 1025(arrayedSampler) 1064
|
||||
1069: 1021 Load 1068
|
||||
1070: 20(fvec2) Load 1033(c2D)
|
||||
1071: 11(fvec4) ImageGather 1069 1070 1064
|
||||
1072: 11(fvec4) Load 13(outp)
|
||||
1073: 11(fvec4) FAdd 1072 1071
|
||||
Store 13(outp) 1073
|
||||
1081: 1078 Load 1080(isamp2DA)
|
||||
1085: 1074(ivec4) ImageGather 1081 1083 25 ConstOffset 1084
|
||||
Store 1076(ioutp) 1085
|
||||
1086: 1078 Load 1080(isamp2DA)
|
||||
1087: 1074(ivec4) ImageGather 1086 1083 25 ConstOffset 1084
|
||||
1088: 1074(ivec4) Load 1076(ioutp)
|
||||
1089: 1074(ivec4) IAdd 1088 1087
|
||||
Store 1076(ioutp) 1089
|
||||
1090: 1078 Load 1080(isamp2DA)
|
||||
1091: 23(int) Load 1027(i)
|
||||
1092: 24(ivec2) CompositeConstruct 1091 1091
|
||||
1093: 1074(ivec4) ImageGather 1090 1083 1064 Offset 1092
|
||||
1094: 1074(ivec4) Load 1076(ioutp)
|
||||
1095: 1074(ivec4) IAdd 1094 1093
|
||||
Store 1076(ioutp) 1095
|
||||
1098: 11(fvec4) Load 1097(gl_FragCoord)
|
||||
1100: 11(fvec4) Load 1099(vl2)
|
||||
1101: 11(fvec4) FAdd 1098 1100
|
||||
1102: 11(fvec4) Load 13(outp)
|
||||
1103: 11(fvec4) FAdd 1102 1101
|
||||
Store 13(outp) 1103
|
||||
1108: 32(int) Load 1107(u)
|
||||
1109: 23(int) Load 1027(i)
|
||||
1110: 32(int) Bitcast 1109
|
||||
1111: 32(int) UMod 1108 1110
|
||||
Store 1105(uo) 1111
|
||||
1112: 2 FunctionCall 6(foo23()
|
||||
1113: 2 FunctionCall 8(doubles()
|
||||
1117: 23(int) Load 1116(gl_PrimitiveID)
|
||||
Store 1115(id) 1117
|
||||
1046: 34(ptr) AccessChain 13(outp) 954
|
||||
1047: 10(float) CompositeExtract 1045 0
|
||||
Store 1046 1047
|
||||
1048: 34(ptr) AccessChain 13(outp) 958
|
||||
1049: 10(float) CompositeExtract 1045 1
|
||||
Store 1048 1049
|
||||
1050: 34(ptr) AccessChain 13(outp) 962
|
||||
1051: 10(float) CompositeExtract 1045 2
|
||||
Store 1050 1051
|
||||
1059: 1056 Load 1058(samp2dr)
|
||||
1060: 20(fvec2) Load 1033(c2D)
|
||||
1071: 1052(ivec4) ImageGather 1059 1060 970 ConstOffsets 1070
|
||||
Store 1054(uoutp) 1071
|
||||
1072: 1029(ptr) AccessChain 1025(arrayedSampler) 1068
|
||||
1073: 1021 Load 1072
|
||||
1074: 20(fvec2) Load 1033(c2D)
|
||||
1075: 11(fvec4) ImageGather 1073 1074 1068
|
||||
1076: 11(fvec4) Load 13(outp)
|
||||
1077: 11(fvec4) FAdd 1076 1075
|
||||
Store 13(outp) 1077
|
||||
1085: 1082 Load 1084(isamp2DA)
|
||||
1089: 1078(ivec4) ImageGather 1085 1087 25 ConstOffset 1088
|
||||
Store 1080(ioutp) 1089
|
||||
1090: 1082 Load 1084(isamp2DA)
|
||||
1091: 1078(ivec4) ImageGather 1090 1087 25 ConstOffset 1088
|
||||
1092: 1078(ivec4) Load 1080(ioutp)
|
||||
1093: 1078(ivec4) IAdd 1092 1091
|
||||
Store 1080(ioutp) 1093
|
||||
1094: 1082 Load 1084(isamp2DA)
|
||||
1095: 23(int) Load 1027(i)
|
||||
1096: 24(ivec2) CompositeConstruct 1095 1095
|
||||
1097: 1078(ivec4) ImageGather 1094 1087 1068 Offset 1096
|
||||
1098: 1078(ivec4) Load 1080(ioutp)
|
||||
1099: 1078(ivec4) IAdd 1098 1097
|
||||
Store 1080(ioutp) 1099
|
||||
1102: 11(fvec4) Load 1101(gl_FragCoord)
|
||||
1104: 11(fvec4) Load 1103(vl2)
|
||||
1105: 11(fvec4) FAdd 1102 1104
|
||||
1106: 11(fvec4) Load 13(outp)
|
||||
1107: 11(fvec4) FAdd 1106 1105
|
||||
Store 13(outp) 1107
|
||||
1112: 32(int) Load 1111(u)
|
||||
1113: 23(int) Load 1027(i)
|
||||
1114: 32(int) Bitcast 1113
|
||||
1115: 32(int) UMod 1112 1114
|
||||
Store 1109(uo) 1115
|
||||
1116: 2 FunctionCall 6(foo23()
|
||||
1117: 2 FunctionCall 8(doubles()
|
||||
1121: 23(int) Load 1120(gl_PrimitiveID)
|
||||
Store 1119(id) 1121
|
||||
Return
|
||||
FunctionEnd
|
||||
6(foo23(): 2 Function None 3
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue