1617a3babSopenharmony_ci#version 460
2617a3babSopenharmony_ci#extension GL_NV_ray_tracing : enable
3617a3babSopenharmony_ci
4617a3babSopenharmony_civec4 undeclared_errors(vec4 f4)
5617a3babSopenharmony_ci{
6617a3babSopenharmony_ci  vec4 result;
7617a3babSopenharmony_ci  gl_SubgroupSize;               // ERROR, extension not enabled (basic)
8617a3babSopenharmony_ci  gl_SubgroupInvocationID;       // ERROR, extension not enabled (basic)
9617a3babSopenharmony_ci  subgroupBarrier();             // ERROR, extension not enabled (basic)
10617a3babSopenharmony_ci  subgroupMemoryBarrier();       // ERROR, extension not enabled (basic)
11617a3babSopenharmony_ci  subgroupMemoryBarrierBuffer(); // ERROR, extension not enabled (basic)
12617a3babSopenharmony_ci  subgroupMemoryBarrierImage();  // ERROR, extension not enabled (basic)
13617a3babSopenharmony_ci  subgroupElect();               // ERROR, extension not enabled (basic)
14617a3babSopenharmony_ci  gl_NumSubgroups;               // ERROR, only defined in compute
15617a3babSopenharmony_ci  gl_SubgroupID;                 // ERROR, only defined in compute
16617a3babSopenharmony_ci  subgroupMemoryBarrierShared(); // ERROR, only defined in compute
17617a3babSopenharmony_ci
18617a3babSopenharmony_ci  subgroupAll(true);        // ERROR extension not enabled (vote)
19617a3babSopenharmony_ci  subgroupAny(false);       // ERROR extension not enabled (vote)
20617a3babSopenharmony_ci  subgroupAllEqual(f4);     // ERROR extension not enabled (vote)
21617a3babSopenharmony_ci
22617a3babSopenharmony_ci  gl_SubgroupEqMask;                        // ERROR extension not enabled (ballot)
23617a3babSopenharmony_ci  gl_SubgroupGeMask;                        // ERROR extension not enabled (ballot)
24617a3babSopenharmony_ci  gl_SubgroupGtMask;                        // ERROR extension not enabled (ballot)
25617a3babSopenharmony_ci  gl_SubgroupLeMask;                        // ERROR extension not enabled (ballot)
26617a3babSopenharmony_ci  gl_SubgroupLtMask;                        // ERROR extension not enabled (ballot)
27617a3babSopenharmony_ci  subgroupBroadcast(f4, 0);                 // ERROR extension not enabled (ballot)
28617a3babSopenharmony_ci  subgroupBroadcastFirst(f4);               // ERROR extension not enabled (ballot)
29617a3babSopenharmony_ci  uvec4 ballot = subgroupBallot(false);     // ERROR extension not enabled (ballot)
30617a3babSopenharmony_ci  subgroupInverseBallot(uvec4(0x1));        // ERROR extension not enabled (ballot)
31617a3babSopenharmony_ci  subgroupBallotBitExtract(ballot, 0);      // ERROR extension not enabled (ballot)
32617a3babSopenharmony_ci  subgroupBallotBitCount(ballot);           // ERROR extension not enabled (ballot)
33617a3babSopenharmony_ci  subgroupBallotInclusiveBitCount(ballot);  // ERROR extension not enabled (ballot)
34617a3babSopenharmony_ci  subgroupBallotExclusiveBitCount(ballot);  // ERROR extension not enabled (ballot)
35617a3babSopenharmony_ci  subgroupBallotFindLSB(ballot);            // ERROR extension not enabled (ballot)
36617a3babSopenharmony_ci  subgroupBallotFindMSB(ballot);            // ERROR extension not enabled (ballot)
37617a3babSopenharmony_ci
38617a3babSopenharmony_ci  subgroupShuffle(f4, 0);           // ERROR extension not enabled (shuffle)
39617a3babSopenharmony_ci  subgroupShuffleXor(f4, 0x1);      // ERROR extension not enabled (shuffle)
40617a3babSopenharmony_ci  subgroupShuffleUp(f4, 1);         // ERROR extension not enabled (shuffle_relative)
41617a3babSopenharmony_ci  subgroupShuffleDown(f4, 1);       // ERROR extension not enabled (shuffle_relative)
42617a3babSopenharmony_ci
43617a3babSopenharmony_ci  result = subgroupAdd(f4);         // ERROR, extension not enabled (arith)
44617a3babSopenharmony_ci  subgroupMul(f4);                  // ERROR, extension not enabled (arith)
45617a3babSopenharmony_ci  subgroupMin(f4);                  // ERROR, extension not enabled (arith)
46617a3babSopenharmony_ci  subgroupMax(f4);                  // ERROR, extension not enabled (arith)
47617a3babSopenharmony_ci  subgroupAnd(ballot);              // ERROR, extension not enabled (arith)
48617a3babSopenharmony_ci  subgroupOr(ballot);               // ERROR, extension not enabled (arith)
49617a3babSopenharmony_ci  subgroupXor(ballot);              // ERROR, extension not enabled (arith)
50617a3babSopenharmony_ci  subgroupInclusiveAdd(f4);         // ERROR, extension not enabled (arith)
51617a3babSopenharmony_ci  subgroupInclusiveMul(f4);         // ERROR, extension not enabled (arith)
52617a3babSopenharmony_ci  subgroupInclusiveMin(f4);         // ERROR, extension not enabled (arith)
53617a3babSopenharmony_ci  subgroupInclusiveMax(f4);         // ERROR, extension not enabled (arith)
54617a3babSopenharmony_ci  subgroupInclusiveAnd(ballot);     // ERROR, extension not enabled (arith)
55617a3babSopenharmony_ci  subgroupInclusiveOr(ballot);      // ERROR, extension not enabled (arith)
56617a3babSopenharmony_ci  subgroupInclusiveXor(ballot);     // ERROR, extension not enabled (arith)
57617a3babSopenharmony_ci  subgroupExclusiveAdd(f4);         // ERROR, extension not enabled (arith)
58617a3babSopenharmony_ci  subgroupExclusiveMul(f4);         // ERROR, extension not enabled (arith)
59617a3babSopenharmony_ci  subgroupExclusiveMin(f4);         // ERROR, extension not enabled (arith)
60617a3babSopenharmony_ci  subgroupExclusiveMax(f4);         // ERROR, extension not enabled (arith)
61617a3babSopenharmony_ci  subgroupExclusiveAnd(ballot);     // ERROR, extension not enabled (arith)
62617a3babSopenharmony_ci  subgroupExclusiveOr(ballot);      // ERROR, extension not enabled (arith)
63617a3babSopenharmony_ci  subgroupExclusiveXor(ballot);     // ERROR, extension not enabled (arith)
64617a3babSopenharmony_ci
65617a3babSopenharmony_ci  subgroupClusteredAdd(f4, 2);      // ERROR, extension not enabled (clustered)
66617a3babSopenharmony_ci  subgroupClusteredMul(f4, 2);      // ERROR, extension not enabled (clustered)
67617a3babSopenharmony_ci  subgroupClusteredMin(f4, 2);      // ERROR, extension not enabled (clustered)
68617a3babSopenharmony_ci  subgroupClusteredMax(f4, 2);      // ERROR, extension not enabled (clustered)
69617a3babSopenharmony_ci  subgroupClusteredAnd(ballot, 2);  // ERROR, extension not enabled (clustered)
70617a3babSopenharmony_ci  subgroupClusteredOr(ballot, 2);   // ERROR, extension not enabled (clustered)
71617a3babSopenharmony_ci  subgroupClusteredXor(ballot, 2);  // ERROR, extension not enabled (clustered)
72617a3babSopenharmony_ci
73617a3babSopenharmony_ci  subgroupQuadBroadcast(f4, 0);     // ERROR, extension not enabled (quad)
74617a3babSopenharmony_ci  subgroupQuadSwapHorizontal(f4);   // ERROR, extension not enabled (quad)
75617a3babSopenharmony_ci  subgroupQuadSwapVertical(f4);     // ERROR, extension not enabled (quad)
76617a3babSopenharmony_ci  subgroupQuadSwapDiagonal(f4);     // ERROR, extension not enabled (quad)
77617a3babSopenharmony_ci
78617a3babSopenharmony_ci  uvec4 parti = subgroupPartitionNV(f4);                // ERROR, extension not enabled (partitioned)
79617a3babSopenharmony_ci  subgroupPartitionedAddNV(f4, parti);                  // ERROR, extension not enabled (partitioned)
80617a3babSopenharmony_ci  subgroupPartitionedMulNV(f4, parti);                  // ERROR, extension not enabled (partitioned)
81617a3babSopenharmony_ci  subgroupPartitionedMinNV(f4, parti);                  // ERROR, extension not enabled (partitioned)
82617a3babSopenharmony_ci  subgroupPartitionedMaxNV(f4, parti);                  // ERROR, extension not enabled (partitioned)
83617a3babSopenharmony_ci  subgroupPartitionedAndNV(ballot, parti);              // ERROR, extension not enabled (partitioned)
84617a3babSopenharmony_ci  subgroupPartitionedOrNV(ballot, parti);               // ERROR, extension not enabled (partitioned)
85617a3babSopenharmony_ci  subgroupPartitionedXorNV(ballot, parti);              // ERROR, extension not enabled (partitioned)
86617a3babSopenharmony_ci  subgroupPartitionedInclusiveAddNV(f4, parti);         // ERROR, extension not enabled (partitioned)
87617a3babSopenharmony_ci  subgroupPartitionedInclusiveMulNV(f4, parti);         // ERROR, extension not enabled (partitioned)
88617a3babSopenharmony_ci  subgroupPartitionedInclusiveMinNV(f4, parti);         // ERROR, extension not enabled (partitioned)
89617a3babSopenharmony_ci  subgroupPartitionedInclusiveMaxNV(f4, parti);         // ERROR, extension not enabled (partitioned)
90617a3babSopenharmony_ci  subgroupPartitionedInclusiveAndNV(ballot, parti);     // ERROR, extension not enabled (partitioned)
91617a3babSopenharmony_ci  subgroupPartitionedInclusiveOrNV(ballot, parti);      // ERROR, extension not enabled (partitioned)
92617a3babSopenharmony_ci  subgroupPartitionedInclusiveXorNV(ballot, parti);     // ERROR, extension not enabled (partitioned)
93617a3babSopenharmony_ci  subgroupPartitionedExclusiveAddNV(f4, parti);         // ERROR, extension not enabled (partitioned)
94617a3babSopenharmony_ci  subgroupPartitionedExclusiveMulNV(f4, parti);         // ERROR, extension not enabled (partitioned)
95617a3babSopenharmony_ci  subgroupPartitionedExclusiveMinNV(f4, parti);         // ERROR, extension not enabled (partitioned)
96617a3babSopenharmony_ci  subgroupPartitionedExclusiveMaxNV(f4, parti);         // ERROR, extension not enabled (partitioned)
97617a3babSopenharmony_ci  subgroupPartitionedExclusiveAndNV(ballot, parti);     // ERROR, extension not enabled (partitioned)
98617a3babSopenharmony_ci  subgroupPartitionedExclusiveOrNV(ballot, parti);      // ERROR, extension not enabled (partitioned)
99617a3babSopenharmony_ci  subgroupPartitionedExclusiveXorNV(ballot, parti);     // ERROR, extension not enabled (partitioned)
100617a3babSopenharmony_ci
101617a3babSopenharmony_ci  return result;
102617a3babSopenharmony_ci}
103617a3babSopenharmony_ci
104617a3babSopenharmony_ci
105617a3babSopenharmony_cilayout(location = 0) callableDataNV vec4 data0;
106617a3babSopenharmony_cilayout(location = 1) callableDataInNV dataBlock {
107617a3babSopenharmony_ci	uint data1;
108617a3babSopenharmony_ci};
109617a3babSopenharmony_civoid main()
110617a3babSopenharmony_ci{
111617a3babSopenharmony_ci	uvec3 id = gl_LaunchIDNV;
112617a3babSopenharmony_ci	uvec3 size = gl_LaunchSizeNV;
113617a3babSopenharmony_ci	data1 = 256U;
114617a3babSopenharmony_ci	executeCallableNV(2,1);
115617a3babSopenharmony_ci}
116617a3babSopenharmony_ci
117617a3babSopenharmony_ci
118617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_basic: enable
119617a3babSopenharmony_civoid basic_works (void)
120617a3babSopenharmony_ci{
121617a3babSopenharmony_ci  gl_SubgroupSize;
122617a3babSopenharmony_ci  gl_SubgroupInvocationID;
123617a3babSopenharmony_ci  subgroupBarrier();
124617a3babSopenharmony_ci  subgroupMemoryBarrier();
125617a3babSopenharmony_ci  subgroupMemoryBarrierBuffer();
126617a3babSopenharmony_ci  subgroupMemoryBarrierImage();
127617a3babSopenharmony_ci  subgroupElect();
128617a3babSopenharmony_ci}
129617a3babSopenharmony_ci
130617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_ballot: enable
131617a3babSopenharmony_civoid ballot_works(vec4 f4) {
132617a3babSopenharmony_ci  gl_SubgroupEqMask;
133617a3babSopenharmony_ci  gl_SubgroupGeMask;
134617a3babSopenharmony_ci  gl_SubgroupGtMask;
135617a3babSopenharmony_ci  gl_SubgroupLeMask;
136617a3babSopenharmony_ci  gl_SubgroupLtMask;
137617a3babSopenharmony_ci  subgroupBroadcast(f4, 0);
138617a3babSopenharmony_ci  subgroupBroadcastFirst(f4);
139617a3babSopenharmony_ci  uvec4 ballot = subgroupBallot(false);
140617a3babSopenharmony_ci  subgroupInverseBallot(uvec4(0x1));
141617a3babSopenharmony_ci  subgroupBallotBitExtract(ballot, 0);
142617a3babSopenharmony_ci  subgroupBallotBitCount(ballot);
143617a3babSopenharmony_ci  subgroupBallotInclusiveBitCount(ballot);
144617a3babSopenharmony_ci  subgroupBallotExclusiveBitCount(ballot);
145617a3babSopenharmony_ci  subgroupBallotFindLSB(ballot);
146617a3babSopenharmony_ci  subgroupBallotFindMSB(ballot);
147617a3babSopenharmony_ci}
148617a3babSopenharmony_ci
149617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_vote: enable
150617a3babSopenharmony_civoid vote_works(vec4 f4)
151617a3babSopenharmony_ci{
152617a3babSopenharmony_ci  subgroupAll(true);
153617a3babSopenharmony_ci  subgroupAny(false);
154617a3babSopenharmony_ci  subgroupAllEqual(f4);
155617a3babSopenharmony_ci}
156617a3babSopenharmony_ci
157617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_shuffle: enable
158617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_shuffle_relative: enable
159617a3babSopenharmony_civoid shuffle_works(vec4 f4)
160617a3babSopenharmony_ci{
161617a3babSopenharmony_ci  subgroupShuffle(f4, 0);
162617a3babSopenharmony_ci  subgroupShuffleXor(f4, 0x1);
163617a3babSopenharmony_ci  subgroupShuffleUp(f4, 1);
164617a3babSopenharmony_ci  subgroupShuffleDown(f4, 1);
165617a3babSopenharmony_ci}
166617a3babSopenharmony_ci
167617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_arithmetic: enable
168617a3babSopenharmony_civoid arith_works(vec4 f4)
169617a3babSopenharmony_ci{
170617a3babSopenharmony_ci  uvec4 ballot;
171617a3babSopenharmony_ci  subgroupAdd(f4);
172617a3babSopenharmony_ci  subgroupMul(f4);
173617a3babSopenharmony_ci  subgroupMin(f4);
174617a3babSopenharmony_ci  subgroupMax(f4);
175617a3babSopenharmony_ci  subgroupAnd(ballot);
176617a3babSopenharmony_ci  subgroupOr(ballot);
177617a3babSopenharmony_ci  subgroupXor(ballot);
178617a3babSopenharmony_ci  subgroupInclusiveAdd(f4);
179617a3babSopenharmony_ci  subgroupInclusiveMul(f4);
180617a3babSopenharmony_ci  subgroupInclusiveMin(f4);
181617a3babSopenharmony_ci  subgroupInclusiveMax(f4);
182617a3babSopenharmony_ci  subgroupInclusiveAnd(ballot);
183617a3babSopenharmony_ci  subgroupInclusiveOr(ballot);
184617a3babSopenharmony_ci  subgroupInclusiveXor(ballot);
185617a3babSopenharmony_ci  subgroupExclusiveAdd(f4);
186617a3babSopenharmony_ci  subgroupExclusiveMul(f4);
187617a3babSopenharmony_ci  subgroupExclusiveMin(f4);
188617a3babSopenharmony_ci  subgroupExclusiveMax(f4);
189617a3babSopenharmony_ci  subgroupExclusiveAnd(ballot);
190617a3babSopenharmony_ci  subgroupExclusiveOr(ballot);
191617a3babSopenharmony_ci  subgroupExclusiveXor(ballot);
192617a3babSopenharmony_ci}
193617a3babSopenharmony_ci
194617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_clustered: enable
195617a3babSopenharmony_civoid clustered_works(vec4 f4)
196617a3babSopenharmony_ci{
197617a3babSopenharmony_ci  uvec4 ballot = uvec4(0x55,0,0,0);
198617a3babSopenharmony_ci  subgroupClusteredAdd(f4, 2);
199617a3babSopenharmony_ci  subgroupClusteredMul(f4, 2);
200617a3babSopenharmony_ci  subgroupClusteredMin(f4, 2);
201617a3babSopenharmony_ci  subgroupClusteredMax(f4, 2);
202617a3babSopenharmony_ci  subgroupClusteredAnd(ballot, 2);
203617a3babSopenharmony_ci  subgroupClusteredOr(ballot, 2);
204617a3babSopenharmony_ci  subgroupClusteredXor(ballot, 2);
205617a3babSopenharmony_ci}
206617a3babSopenharmony_ci
207617a3babSopenharmony_ci#extension GL_KHR_shader_subgroup_quad: enable
208617a3babSopenharmony_civoid quad_works(vec4 f4)
209617a3babSopenharmony_ci{
210617a3babSopenharmony_ci  subgroupQuadBroadcast(f4, 0);
211617a3babSopenharmony_ci  subgroupQuadSwapHorizontal(f4);
212617a3babSopenharmony_ci  subgroupQuadSwapVertical(f4);
213617a3babSopenharmony_ci  subgroupQuadSwapDiagonal(f4);
214617a3babSopenharmony_ci}
215617a3babSopenharmony_ci
216617a3babSopenharmony_ci#extension GL_NV_shader_subgroup_partitioned: enable
217617a3babSopenharmony_civoid partitioned_works(vec4 f4)
218617a3babSopenharmony_ci{
219617a3babSopenharmony_ci  uvec4 parti = subgroupPartitionNV(f4);
220617a3babSopenharmony_ci  uvec4 ballot = uvec4(0x55,0,0,0);
221617a3babSopenharmony_ci  subgroupPartitionedAddNV(f4, parti);
222617a3babSopenharmony_ci  subgroupPartitionedMulNV(f4, parti);
223617a3babSopenharmony_ci  subgroupPartitionedMinNV(f4, parti);
224617a3babSopenharmony_ci  subgroupPartitionedMaxNV(f4, parti);
225617a3babSopenharmony_ci  subgroupPartitionedAndNV(ballot, parti);
226617a3babSopenharmony_ci  subgroupPartitionedOrNV(ballot, parti);
227617a3babSopenharmony_ci  subgroupPartitionedXorNV(ballot, parti);
228617a3babSopenharmony_ci  subgroupPartitionedInclusiveAddNV(f4, parti);
229617a3babSopenharmony_ci  subgroupPartitionedInclusiveMulNV(f4, parti);
230617a3babSopenharmony_ci  subgroupPartitionedInclusiveMinNV(f4, parti);
231617a3babSopenharmony_ci  subgroupPartitionedInclusiveMaxNV(f4, parti);
232617a3babSopenharmony_ci  subgroupPartitionedInclusiveAndNV(ballot, parti);
233617a3babSopenharmony_ci  subgroupPartitionedInclusiveOrNV(ballot, parti);
234617a3babSopenharmony_ci  subgroupPartitionedInclusiveXorNV(ballot, parti);
235617a3babSopenharmony_ci  subgroupPartitionedExclusiveAddNV(f4, parti);
236617a3babSopenharmony_ci  subgroupPartitionedExclusiveMulNV(f4, parti);
237617a3babSopenharmony_ci  subgroupPartitionedExclusiveMinNV(f4, parti);
238617a3babSopenharmony_ci  subgroupPartitionedExclusiveMaxNV(f4, parti);
239617a3babSopenharmony_ci  subgroupPartitionedExclusiveAndNV(ballot, parti);
240617a3babSopenharmony_ci  subgroupPartitionedExclusiveOrNV(ballot, parti);
241617a3babSopenharmony_ci  subgroupPartitionedExclusiveXorNV(ballot, parti);
242617a3babSopenharmony_ci}
243617a3babSopenharmony_ci
244617a3babSopenharmony_ci// tests for NV_shader_sm_builtins
245617a3babSopenharmony_civoid sm_builtins_err()
246617a3babSopenharmony_ci{
247617a3babSopenharmony_ci    gl_WarpsPerSMNV;    // ERROR, no extension
248617a3babSopenharmony_ci    gl_SMCountNV;       // ERROR, no extension
249617a3babSopenharmony_ci    gl_WarpIDNV;        // ERROR, no extension
250617a3babSopenharmony_ci    gl_SMIDNV;          // ERROR, no extension
251617a3babSopenharmony_ci}
252617a3babSopenharmony_ci
253617a3babSopenharmony_ci#ifdef GL_NV_shader_sm_builtins
254617a3babSopenharmony_ci#extension GL_NV_shader_sm_builtins : enable
255617a3babSopenharmony_ci#endif
256617a3babSopenharmony_ci
257617a3babSopenharmony_civoid sm_builtins()
258617a3babSopenharmony_ci{
259617a3babSopenharmony_ci    gl_WarpsPerSMNV;
260617a3babSopenharmony_ci    gl_SMCountNV;
261617a3babSopenharmony_ci    gl_WarpIDNV;
262617a3babSopenharmony_ci    gl_SMIDNV;
263617a3babSopenharmony_ci}
264