mewin/glslang
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

GL_NV_integer_cooperative_matrix support

Browse Source
This commit is contained in:
Jeff Bolz
2019-08-22 20:28:00 -05:00
parent a3bc04b278
commit 387657e4cf
15 changed files with 3580 additions and 2797 deletions
Download Patch File Download Diff File Expand all files Collapse all files

418
Test/baseResults/spv.intcoopmat.comp.out Normal file
View File

@@ -0,0 +1,418 @@
spv.intcoopmat.comp
// Module Version 10000
// Generated by (magic number): 80007
// Id's are bound by 262
Capability Shader
Capability Float16
Capability Int8
Capability StorageBuffer8BitAccess
Capability VulkanMemoryModelKHR
Capability PhysicalStorageBufferAddressesEXT
Capability CooperativeMatrixNV
Extension "SPV_EXT_physical_storage_buffer"
Extension "SPV_KHR_8bit_storage"
Extension "SPV_KHR_storage_buffer_storage_class"
Extension "SPV_KHR_vulkan_memory_model"
Extension "SPV_NV_cooperative_matrix"
1: ExtInstImport "GLSL.std.450"
MemoryModel PhysicalStorageBuffer64EXT VulkanKHR
EntryPoint GLCompute 4 "main"
ExecutionMode 4 LocalSize 64 1 1
Source GLSL 450
SourceExtension "GL_EXT_buffer_reference"
SourceExtension "GL_EXT_shader_explicit_arithmetic_types"
SourceExtension "GL_KHR_memory_scope_semantics"
SourceExtension "GL_NV_cooperative_matrix"
SourceExtension "GL_NV_integer_cooperative_matrix"
Name 4 "main"
Name 14 "ineg(i81;"
Name 13 "m"
Name 21 "umul(u81;"
Name 20 "m"
Name 35 "mu"
Name 39 "mi"
Name 55 "mf16_0"
Name 61 "mf32_0"
Name 64 "mf16_1"
Name 67 "mf32_1"
Name 71 "x"
Name 81 "tempArg"
Name 85 "Block"
MemberName 85(Block) 0 "y"
MemberName 85(Block) 1 "x"
Name 87 "block"
Name 98 "tempArg"
Name 103 "Block16"
MemberName 103(Block16) 0 "y"
MemberName 103(Block16) 1 "x"
MemberName 103(Block16) 2 "b"
Name 106 "Block"
MemberName 106(Block) 0 "y"
MemberName 106(Block) 1 "x"
Name 108 "block8"
Name 115 "tempArg"
Name 128 "D"
Name 129 "A"
Name 131 "B"
Name 133 "C"
Name 137 "l"
Name 142 "a"
Name 146 "md1"
Name 156 "Y"
Name 157 "Z"
Name 161 "muC2"
Name 169 "miC2"
Name 176 "tempArg"
Name 182 "tempArg"
Name 188 "p1"
Name 189 "param"
Name 192 "p2"
Name 193 "param"
Name 207 "tempArg"
Name 212 "shmatrix"
Name 217 "ms"
Name 225 "miC"
Name 226 "muC"
Name 231 "iarr"
Name 236 "iarr2"
Name 241 "uarr"
Name 246 "uarr2"
Name 251 "S"
MemberName 251(S) 0 "a"
MemberName 251(S) 1 "b"
MemberName 251(S) 2 "c"
Name 256 "SC"
Name 261 "scm"
Decorate 83 ArrayStride 4
Decorate 84 ArrayStride 4
MemberDecorate 85(Block) 0 Offset 0
MemberDecorate 85(Block) 1 Offset 4194304
Decorate 85(Block) Block
Decorate 87(block) DescriptorSet 0
Decorate 87(block) Binding 0
Decorate 99 ArrayStride 1
Decorate 101 ArrayStride 1
MemberDecorate 103(Block16) 0 Offset 0
MemberDecorate 103(Block16) 1 Offset 1048576
MemberDecorate 103(Block16) 2 Offset 1048584
Decorate 103(Block16) Block
Decorate 104 ArrayStride 4
Decorate 105 ArrayStride 4
MemberDecorate 106(Block) 0 Offset 0
MemberDecorate 106(Block) 1 Offset 4194304
Decorate 106(Block) Block
Decorate 108(block8) DescriptorSet 0
Decorate 108(block8) Binding 0
Decorate 156(Y) SpecId 0
Decorate 223 BuiltIn WorkgroupSize
Decorate 256(SC) SpecId 2
2: TypeVoid
3: TypeFunction 2
6: TypeInt 8 1
7: TypeInt 32 0
8: 7(int) Constant 3
9: 7(int) Constant 8
10: TypeCooperativeMatrixNV 6(int8_t) 8 9 9
11: TypePointer Function 10
12: TypeFunction 10 11(ptr)
16: TypeInt 8 0
17: TypeCooperativeMatrixNV 16(int8_t) 8 9 9
18: TypePointer Function 17
19: TypeFunction 17 18(ptr)
28: 16(int8_t) Constant 2
32: 7(int) Constant 16
33: TypeCooperativeMatrixNV 16(int8_t) 8 32 9
34: TypePointer Function 33
36: 33 ConstantComposite 28
37: TypeCooperativeMatrixNV 6(int8_t) 8 32 9
38: TypePointer Function 37
40: 6(int8_t) Constant 2
41: 37 ConstantComposite 40
52: TypeFloat 16
53: TypeCooperativeMatrixNV 52(float16_t) 8 32 9
54: TypePointer Function 53
58: TypeFloat 32
59: TypeCooperativeMatrixNV 58(float) 8 32 9
60: TypePointer Function 59
70: TypePointer Function 16(int8_t)
72: TypeInt 32 1
73: 72(int) Constant 1
76: 72(int) Constant 0
79: TypePointer Function 6(int8_t)
82: 7(int) Constant 1048576
83: TypeArray 7(int) 82
84: TypeRuntimeArray 7(int)
85(Block): TypeStruct 83 84
86: TypePointer StorageBuffer 85(Block)
87(block): 86(ptr) Variable StorageBuffer
88: 7(int) Constant 5
89: TypePointer StorageBuffer 7(int)
91: 7(int) Constant 128
92: TypeBool
93: 92(bool) ConstantFalse
99: TypeArray 6(int8_t) 82
100: 7(int) Constant 1
101: TypeArray 6(int8_t) 100
TypeForwardPointer 102 PhysicalStorageBufferEXT
103(Block16): TypeStruct 99 101 102
104: TypeArray 7(int) 82
105: TypeRuntimeArray 7(int)
106(Block): TypeStruct 104 105
102: TypePointer PhysicalStorageBufferEXT 106(Block)
107: TypePointer StorageBuffer 103(Block16)
108(block8): 107(ptr) Variable StorageBuffer
109: TypePointer StorageBuffer 6(int8_t)
116: 72(int) Constant 2
117: TypePointer StorageBuffer 102(ptr)
120: TypePointer PhysicalStorageBufferEXT 7(int)
136: TypePointer Function 72(int)
138: 7(int) SpecConstantOp 5362 33
139: 72(int) SpecConstantOp 128 138 76
140: TypeArray 37 88
141: TypePointer Function 140
143: 72(int) Constant 3
144: 6(int8_t) Constant 1
150: 72(int) Constant 1234
155: 72(int) Constant 8
156(Y): 72(int) SpecConstant 2
157(Z): 72(int) SpecConstantOp 132 155 156(Y)
158: TypeCooperativeMatrixNV 16(int8_t) 8 157(Z) 9
159: TypeArray 158 8
160: TypePointer Private 159
161(muC2): 160(ptr) Variable Private
162: TypePointer Private 158
166: TypeCooperativeMatrixNV 6(int8_t) 8 157(Z) 9
167: TypeArray 166 8
168: TypePointer Private 167
169(miC2): 168(ptr) Variable Private
170: TypePointer Private 6(int8_t)
174: TypePointer Private 16(int8_t)
204: 16(int8_t) Constant 4
208: TypeVector 7(int) 4
209: 7(int) Constant 32
210: TypeArray 208(ivec4) 209
211: TypePointer Workgroup 210
212(shmatrix): 211(ptr) Variable Workgroup
213: 7(int) Constant 2
214: TypePointer Workgroup 208(ivec4)
221: TypeVector 7(int) 3
222: 7(int) Constant 64
223: 221(ivec3) ConstantComposite 222 100 100
224: TypePointer Private 166
225(miC): 224(ptr) Variable Private
226(muC): 162(ptr) Variable Private
227: 7(int) SpecConstantOp 5362 166
228: 72(int) SpecConstantOp 128 227 76
229: TypeArray 72(int) 228
230: TypePointer Private 229
231(iarr): 230(ptr) Variable Private
232: 7(int) SpecConstantOp 5362 166
233: 72(int) SpecConstantOp 128 232 76
234: TypeArray 72(int) 233
235: TypePointer Private 234
236(iarr2): 235(ptr) Variable Private
237: 7(int) SpecConstantOp 5362 158
238: 72(int) SpecConstantOp 128 237 76
239: TypeArray 72(int) 238
240: TypePointer Private 239
241(uarr): 240(ptr) Variable Private
242: 7(int) SpecConstantOp 5362 158
243: 72(int) SpecConstantOp 128 242 76
244: TypeArray 72(int) 243
245: TypePointer Private 244
246(uarr2): 245(ptr) Variable Private
247: TypeCooperativeMatrixNV 72(int) 8 157(Z) 9
248: 247 ConstantComposite 73
249: 16(int8_t) Constant 1
250: 17 ConstantComposite 249
251(S): TypeStruct 72(int) 72(int) 72(int)
252: 72(int) Constant 12
253: 72(int) Constant 23
254: 72(int) Constant 34
255: 251(S) ConstantComposite 252 253 254
256(SC): 72(int) SpecConstant 1
257: TypeCooperativeMatrixNV 7(int) 8 256(SC) 256(SC)
258: TypeArray 257 256(SC)
259: TypeArray 258 256(SC)
260: TypePointer Private 259
261(scm): 260(ptr) Variable Private
4(main): 2 Function None 3
5: Label
35(mu): 34(ptr) Variable Function
39(mi): 38(ptr) Variable Function
55(mf16_0): 54(ptr) Variable Function
61(mf32_0): 60(ptr) Variable Function
64(mf16_1): 54(ptr) Variable Function
67(mf32_1): 60(ptr) Variable Function
71(x): 70(ptr) Variable Function
81(tempArg): 38(ptr) Variable Function
98(tempArg): 34(ptr) Variable Function
115(tempArg): 38(ptr) Variable Function
128(D): 34(ptr) Variable Function
129(A): 34(ptr) Variable Function
131(B): 18(ptr) Variable Function
133(C): 34(ptr) Variable Function
137(l): 136(ptr) Variable Function
142(a): 141(ptr) Variable Function
146(md1): 136(ptr) Variable Function
176(tempArg): 38(ptr) Variable Function
182(tempArg): 34(ptr) Variable Function
188(p1): 11(ptr) Variable Function
189(param): 11(ptr) Variable Function
192(p2): 18(ptr) Variable Function
193(param): 18(ptr) Variable Function
207(tempArg): 38(ptr) Variable Function
217(ms): 38(ptr) Variable Function
Store 35(mu) 36
Store 39(mi) 41
42: 33 Load 35(mu)
43: 33 Load 35(mu)
44: 33 IAdd 42 43
Store 35(mu) 44
45: 33 Load 35(mu)
46: 33 Load 35(mu)
47: 33 ISub 45 46
Store 35(mu) 47
48: 37 Load 39(mi)
49: 37 SNegate 48
Store 39(mi) 49
50: 37 Load 39(mi)
51: 37 MatrixTimesScalar 50 40
Store 39(mi) 51
56: 33 Load 35(mu)
57: 53 ConvertUToF 56
Store 55(mf16_0) 57
62: 33 Load 35(mu)
63: 59 ConvertUToF 62
Store 61(mf32_0) 63
65: 37 Load 39(mi)
66: 53 ConvertSToF 65
Store 64(mf16_1) 66
68: 37 Load 39(mi)
69: 59 ConvertSToF 68
Store 67(mf32_1) 69
74: 70(ptr) AccessChain 35(mu) 73
75: 16(int8_t) Load 74
Store 71(x) 75
77: 16(int8_t) Load 71(x)
78: 6(int8_t) Bitcast 77
80: 79(ptr) AccessChain 39(mi) 76
Store 80 78
90: 89(ptr) AccessChain 87(block) 73 32
94: 37 CooperativeMatrixLoadNV 90 91 93 MakePointerVisibleKHR NonPrivatePointerKHR 88
Store 81(tempArg) 94
95: 37 Load 81(tempArg)
Store 39(mi) 95
96: 37 Load 39(mi)
97: 89(ptr) AccessChain 87(block) 73 32
CooperativeMatrixStoreNV 97 96 91 93 MakePointerAvailableKHR NonPrivatePointerKHR 88
110: 109(ptr) AccessChain 108(block8) 73 32
111: 33 CooperativeMatrixLoadNV 110 91 93 MakePointerVisibleKHR NonPrivatePointerKHR 88
Store 98(tempArg) 111
112: 33 Load 98(tempArg)
Store 35(mu) 112
113: 33 Load 35(mu)
114: 109(ptr) AccessChain 108(block8) 73 32
CooperativeMatrixStoreNV 114 113 91 93 MakePointerAvailableKHR NonPrivatePointerKHR 88
118: 117(ptr) AccessChain 108(block8) 116
119: 102(ptr) Load 118 MakePointerVisibleKHR NonPrivatePointerKHR 88
121: 120(ptr) AccessChain 119 73 32
122: 37 CooperativeMatrixLoadNV 121 91 93 Aligned MakePointerVisibleKHR NonPrivatePointerKHR 16 88
Store 115(tempArg) 122
123: 37 Load 115(tempArg)
Store 39(mi) 123
124: 37 Load 39(mi)
125: 117(ptr) AccessChain 108(block8) 116
126: 102(ptr) Load 125 MakePointerVisibleKHR NonPrivatePointerKHR 88
127: 120(ptr) AccessChain 126 73 32
CooperativeMatrixStoreNV 127 124 91 93 Aligned MakePointerAvailableKHR NonPrivatePointerKHR 16 88
130: 33 Load 129(A)
132: 17 Load 131(B)
134: 33 Load 133(C)
135: 33 CooperativeMatrixMulAddNV 130 132 134
Store 128(D) 135
Store 137(l) 139
145: 79(ptr) AccessChain 142(a) 143 76
Store 145 144
Store 146(md1) 73
147: 37 Load 39(mi)
148: 37 Load 39(mi)
149: 37 IAdd 148 147
Store 39(mi) 149
151: 6(int8_t) CompositeExtract 149 1234
152: 72(int) SConvert 151
153: 72(int) Load 146(md1)
154: 72(int) IAdd 153 152
Store 146(md1) 154
163: 162(ptr) AccessChain 161(muC2) 73
164: 158 Load 163
165: 162(ptr) AccessChain 161(muC2) 76
Store 165 164
171: 170(ptr) AccessChain 169(miC2) 116 76
172: 6(int8_t) Load 171
173: 16(int8_t) Bitcast 172
175: 174(ptr) AccessChain 161(muC2) 73 76
Store 175 173
177: 89(ptr) AccessChain 87(block) 76 32
178: 37 CooperativeMatrixLoadNV 177 91 93 MakePointerVisibleKHR NonPrivatePointerKHR 88
Store 176(tempArg) 178
179: 37 Load 176(tempArg)
Store 39(mi) 179
180: 37 Load 39(mi)
181: 89(ptr) AccessChain 87(block) 76 32
CooperativeMatrixStoreNV 181 180 91 93 MakePointerAvailableKHR NonPrivatePointerKHR 88
183: 109(ptr) AccessChain 108(block8) 76 32
184: 33 CooperativeMatrixLoadNV 183 91 93 MakePointerVisibleKHR NonPrivatePointerKHR 88
Store 182(tempArg) 184
185: 33 Load 182(tempArg)
Store 35(mu) 185
186: 33 Load 35(mu)
187: 109(ptr) AccessChain 108(block8) 76 32
CooperativeMatrixStoreNV 187 186 91 93 MakePointerAvailableKHR NonPrivatePointerKHR 88
190: 10 Load 188(p1)
Store 189(param) 190
191: 10 FunctionCall 14(ineg(i81;) 189(param)
Store 188(p1) 191
194: 17 Load 192(p2)
Store 193(param) 194
195: 17 FunctionCall 21(umul(u81;) 193(param)
Store 192(p2) 195
196: 10 Load 188(p1)
197: 10 Load 188(p1)
198: 10 SDiv 197 196
Store 188(p1) 198
199: 17 Load 192(p2)
200: 17 Load 192(p2)
201: 17 UDiv 200 199
Store 192(p2) 201
202: 10 Load 188(p1)
203: 10 MatrixTimesScalar 202 40
Store 188(p1) 203
205: 17 Load 192(p2)
206: 17 MatrixTimesScalar 205 204
Store 192(p2) 206
215: 214(ptr) AccessChain 212(shmatrix) 100
216: 37 CooperativeMatrixLoadNV 215 213 93 MakePointerVisibleKHR NonPrivatePointerKHR 213
Store 207(tempArg) 216
218: 37 Load 207(tempArg)
Store 217(ms) 218
219: 37 Load 217(ms)
220: 214(ptr) AccessChain 212(shmatrix) 100
CooperativeMatrixStoreNV 220 219 213 93 MakePointerAvailableKHR NonPrivatePointerKHR 213
Return
FunctionEnd
14(ineg(i81;): 10 Function None 12
13(m): 11(ptr) FunctionParameter
15: Label
23: 10 Load 13(m)
24: 10 SNegate 23
ReturnValue 24
FunctionEnd
21(umul(u81;): 17 Function None 19
20(m): 18(ptr) FunctionParameter
22: Label
27: 17 Load 20(m)
29: 17 MatrixTimesScalar 27 28
ReturnValue 29
FunctionEnd

117
Test/spv.intcoopmat.comp Normal file
View File

@@ -0,0 +1,117 @@
#version 450 core
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_NV_cooperative_matrix : enable
#extension GL_NV_integer_cooperative_matrix : enable
#extension GL_EXT_shader_explicit_arithmetic_types : enable
#extension GL_EXT_buffer_reference : enable
layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
const int X = 8;
layout(constant_id = 0) const int Y = 2;
const int Z = X*Y;
icoopmatNV<8, gl_ScopeSubgroup, Z, 8> miC;
icoopmatNV<8, gl_ScopeSubgroup, Z, 8> miC2[3];
ucoopmatNV<8, gl_ScopeSubgroup, Z, 8> muC;
ucoopmatNV<8, gl_ScopeSubgroup, Z, 8> muC2[3];
int iarr[miC.length()];
int iarr2[miC2[1].length()];
int uarr[muC.length()];
int uarr2[muC2[1].length()];
const icoopmatNV<32, gl_ScopeSubgroup, Z, 8> mD = icoopmatNV<32, gl_ScopeSubgroup, Z, 8>(1);
const ucoopmatNV<8, gl_ScopeSubgroup, 8, 8> mD2 = ucoopmatNV<8, gl_ScopeSubgroup, 8, 8>(1);
struct S { int a; int b; int c; };
const S s = S(12, 23, 34);
layout(set = 0, binding = 0, buffer_reference) coherent buffer Block {
uint y[1024*1024];
uint x[];
} block;
layout(set = 0, binding = 0) coherent buffer Block16 {
int8_t y[1024*1024];
int8_t x[];
Block b;
} block8;
icoopmatNV<8, gl_ScopeSubgroup, 8, 8> ineg(icoopmatNV<8, gl_ScopeSubgroup, 8, 8> m) { return -m; }
ucoopmatNV<8, gl_ScopeSubgroup, 8, 8> umul(ucoopmatNV<8, gl_ScopeSubgroup, 8, 8> m) { return m * uint8_t(2); }
layout(constant_id = 2) const int SC = 1;
ucoopmatNV<32, gl_ScopeSubgroup, SC, SC> scm[SC][SC];
// sized for icoopmatNV<8, gl_ScopeSubgroup, 16, 16>
shared uvec4 shmatrix[16*16*2/16];
void main()
{
ucoopmatNV<8, gl_ScopeSubgroup, 16, (2>1?8:4)> mu = ucoopmatNV<8, gl_ScopeSubgroup, 16, (2>1?8:4)>(2);
icoopmatNV<8, gl_ScopeSubgroup, 16, (2>1?8:4)> mi = icoopmatNV<8, gl_ScopeSubgroup, 16, (2>1?8:4)>(2);
mu = mu + mu;
mu = mu - mu;
mi = -mi;
mi = mi * int8_t(2);
fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> mf16_0 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(mu);
fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> mf32_0 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(mu);
fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> mf16_1 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(mi);
fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> mf32_1 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(mi);
uint8_t x = mu[1];
mi[0] = int8_t(x);
coopMatLoadNV(mi, block.x, 16, 128, false);
coopMatStoreNV(mi, block.x, 16, 128, false);
coopMatLoadNV(mu, block8.x, 16, 128, false);
coopMatStoreNV(mu, block8.x, 16, 128, false);
coopMatLoadNV(mi, block8.b.x, 16, 128, false);
coopMatStoreNV(mi, block8.b.x, 16, 128, false);
ucoopmatNV<8, gl_ScopeSubgroup, 16, 8> A;
ucoopmatNV<8, gl_ScopeSubgroup, 8, 8> B;
ucoopmatNV<8, gl_ScopeSubgroup, 16, 8> C;
ucoopmatNV<8, gl_ScopeSubgroup, 16, 8> D;
D = coopMatMulAddNV(A, B, C);
int l = D.length();
icoopmatNV<8, gl_ScopeSubgroup, 16, (2>1?8:4)> a[5];
a[3][0] = int8_t(1);
int md1 = mD[1];
md1 += (mi += mi)[1234];
muC2[0] = muC2[1];
muC2[1][0] = (miC2[2][0]);
coopMatLoadNV(mi, block.y, 16, 128, false);
coopMatStoreNV(mi, block.y, 16, 128, false);
coopMatLoadNV(mu, block8.y, 16, 128, false);
coopMatStoreNV(mu, block8.y, 16, 128, false);
icoopmatNV<8, gl_ScopeSubgroup, 8, 8> p1;
ucoopmatNV<8, gl_ScopeSubgroup, 8, 8> p2;
p1 = ineg(p1);
p2 = umul(p2);
p1 /= p1;
p2 /= p2;
p1 *= int8_t(2);
p2 *= uint8_t(4);
icoopmatNV<8, gl_ScopeSubgroup, 16, 8> ms;
coopMatLoadNV(ms, shmatrix, 1, 2, false);
coopMatStoreNV(ms, shmatrix, 1, 2, false);
}