@@ -189,89 +189,96 @@ int main() {
189189 // This combination is not currently supported for sub group size = 32 in
190190 // IGC
191191#if (!defined(SG_SZ) || SG_SZ != 32)
192- // 8x16x16 float/bfloat16
193- std::cout << " 8x16x16 float/bfloat16" << std::endl;
194- // A
195- test_get_coord_op<bfloat16, float , /* TM*/ 8 , /* TK*/ 16 , use::a,
196- layout::row_major, 1 >();
197- // B
198- test_get_coord_op<bfloat16, float , /* TK*/ 16 , /* TN*/ 16 , use::b,
199- layout::ext_intel_packed, 2 >();
192+
200193 test_get_coord_op<bfloat16, float , /* TK*/ 16 , /* TN*/ 16 , use::b,
201194 layout::row_major, 1 >();
202- // Accumulator
203- test_get_coord_op<bfloat16, float , /* TM*/ 8 , /* TN*/ 16 , use::accumulator,
204- layout::row_major, 1 >();
205- test_get_coord_op<float , float , /* TM*/ 8 , /* TN*/ 16 , use::accumulator,
195+ test_get_coord_op<int8_t , int32_t , /* TK*/ 32 , /* TN*/ 16 , use::b,
206196 layout::row_major, 1 >();
207197
208- // 16x16x16 float/bfloat16
209- std::cout << " 16x16x16 float/bfloat16" << std::endl;
210- // A
211- test_get_coord_op<bfloat16, float , /* TM*/ 16 , /* TK*/ 16 , use::a,
212- layout::row_major, 1 >();
213- // B - duplicate from 8x16x16
214- // Accumulator
215- test_get_coord_op<bfloat16, float , /* TM*/ 16 , /* TN*/ 16 , use::accumulator,
216- layout::row_major, 1 >();
217- test_get_coord_op<float , float , /* TM*/ 16 , /* TN*/ 16 , use::accumulator,
218- layout::row_major, 1 >();
198+ // These will be uncommented once the IGC task in GSD-11139 is merged
199+ // // 8x16x16 float/bfloat16
200+ // std::cout << "8x16x16 float/bfloat16" << std::endl;
201+ // // A
202+ // test_get_coord_op<bfloat16, float, /*TM*/ 8, /*TK*/ 16, use::a,
203+ // layout::row_major, 1>();
204+ // // B
205+ // test_get_coord_op<bfloat16, float, /*TK*/ 16, /*TN*/ 16, use::b,
206+ // layout::ext_intel_packed, 2>();
207+ // test_get_coord_op<bfloat16, float, /*TK*/ 16, /*TN*/ 16, use::b,
208+ // layout::row_major, 1>();
209+ // // Accumulator
210+ // test_get_coord_op<bfloat16, float, /*TM*/ 8, /*TN*/ 16, use::accumulator,
211+ // layout::row_major, 1>();
212+ // test_get_coord_op<float, float, /*TM*/ 8, /*TN*/ 16, use::accumulator,
213+ // layout::row_major, 1>();
219214
220- // 1x64x16 float/bfloat16
221- std::cout << " 1x64x16 float/bfloat16" << std::endl;
222- // A
223- test_get_coord_op<bfloat16, float , /* TM*/ 1 , /* TK*/ 16 , use::a,
224- layout::row_major, 1 >();
225- // B
226- test_get_coord_op<bfloat16, float , /* TK*/ 16 , /* TN*/ 64 , use::b,
227- layout::ext_intel_packed, 2 >();
228- test_get_coord_op<bfloat16, float , /* TK*/ 16 , /* TN*/ 64 , use::b,
229- layout::row_major, 1 >();
230- // Accumulator
231- test_get_coord_op<bfloat16, float , /* TM*/ 1 , /* TN*/ 64 , use::accumulator,
232- layout::row_major, 1 >();
233- test_get_coord_op<float , float , /* TM*/ 1 , /* TN*/ 64 , use::accumulator,
234- layout::row_major, 1 >();
215+ // // 16x16x16 float/bfloat16
216+ // std::cout << "16x16x16 float/bfloat16" << std::endl;
217+ // // A
218+ // test_get_coord_op<bfloat16, float, /*TM*/ 16, /*TK*/ 16, use::a,
219+ // layout::row_major, 1>();
220+ // // B - duplicate from 8x16x16
221+ // // Accumulator
222+ // test_get_coord_op<bfloat16, float, /*TM*/ 16, /*TN*/ 16, use::accumulator,
223+ // layout::row_major, 1>();
224+ // test_get_coord_op<float, float, /*TM*/ 16, /*TN*/ 16, use::accumulator,
225+ // layout::row_major, 1>();
235226
236- // 1x64x32 float/bfloat16
237- std::cout << " 1x64x32 float/bfloat16" << std::endl;
238- // A
239- test_get_coord_op<bfloat16, float , /* TM*/ 1 , /* TK*/ 32 , use::a,
240- layout::row_major, 1 >();
241- // B
242- test_get_coord_op<bfloat16, float , /* TK*/ 32 , /* TN*/ 64 , use::b,
243- layout::ext_intel_packed, 2 >();
244- test_get_coord_op<bfloat16, float , /* TK*/ 32 , /* TN*/ 64 , use::b,
245- layout::row_major, 1 >();
246- // Accumulator
247- test_get_coord_op<bfloat16, float , /* TM*/ 1 , /* TN*/ 64 , use::accumulator,
248- layout::row_major, 1 >();
249- test_get_coord_op<float , float , /* TM*/ 1 , /* TN*/ 64 , use::accumulator,
250- layout::row_major, 1 >();
227+ // // 1x64x16 float/bfloat16
228+ // std::cout << "1x64x16 float/bfloat16" << std::endl;
229+ // // A
230+ // test_get_coord_op<bfloat16, float, /*TM*/ 1, /*TK*/ 16 , use::a,
231+ // layout::row_major, 1>();
232+ // // B
233+ // test_get_coord_op<bfloat16, float, /*TK*/ 16 , /*TN*/ 64, use::b,
234+ // layout::ext_intel_packed, 2>();
235+ // test_get_coord_op<bfloat16, float, /*TK*/ 16 , /*TN*/ 64, use::b,
236+ // layout::row_major, 1>();
237+ // // Accumulator
238+ // test_get_coord_op<bfloat16, float, /*TM*/ 1, /*TN*/ 64, use::accumulator,
239+ // layout::row_major, 1>();
240+ // test_get_coord_op<float, float, /*TM*/ 1, /*TN*/ 64, use::accumulator,
241+ // layout::row_major, 1>();
251242
252- // 32x64x16 float/bfloat16
253- std::cout << " 32x64x16 float/bfloat16" << std::endl;
254- // A
255- test_get_coord_op<bfloat16, float , /* TM*/ 32 , /* TK*/ 16 , use::a,
256- layout::row_major, 1 >();
257- // B - duplicate from 1x64x16
258- // Accumulator
259- test_get_coord_op<bfloat16, float , /* TM*/ 32 , /* TN*/ 64 , use::accumulator,
260- layout::row_major, 1 >();
261- test_get_coord_op<float , float , /* TM*/ 32 , /* TN*/ 64 , use::accumulator,
262- layout::row_major, 1 >();
243+ // // 1x64x32 float/bfloat16
244+ // std::cout << "1x64x32 float/bfloat16" << std::endl;
245+ // // A
246+ // test_get_coord_op<bfloat16, float, /*TM*/ 1, /*TK*/ 32, use::a,
247+ // layout::row_major, 1>();
248+ // // B
249+ // test_get_coord_op<bfloat16, float, /*TK*/ 32, /*TN*/ 64, use::b,
250+ // layout::ext_intel_packed, 2>();
251+ // test_get_coord_op<bfloat16, float, /*TK*/ 32, /*TN*/ 64, use::b,
252+ // layout::row_major, 1>();
253+ // // Accumulator
254+ // test_get_coord_op<bfloat16, float, /*TM*/ 1, /*TN*/ 64, use::accumulator,
255+ // layout::row_major, 1>();
256+ // test_get_coord_op<float, float, /*TM*/ 1, /*TN*/ 64, use::accumulator,
257+ // layout::row_major, 1>();
263258
264- // 32x64x32 float/bfloat16
265- std::cout << " 32x64x32 float/bfloat16" << std::endl;
266- // A
267- test_get_coord_op<bfloat16, float , /* TM*/ 32 , /* TK*/ 32 , use::a,
268- layout::row_major, 1 >();
269- // B - duplicate from 1x64x32
270- // Accumulator
271- test_get_coord_op<bfloat16, float , /* TM*/ 32 , /* TN*/ 64 , use::accumulator,
272- layout::row_major, 1 >();
273- test_get_coord_op<float , float , /* TM*/ 32 , /* TN*/ 64 , use::accumulator,
274- layout::row_major, 1 >();
259+ // // 32x64x16 float/bfloat16
260+ // std::cout << "32x64x16 float/bfloat16" << std::endl;
261+ // // A
262+ // test_get_coord_op<bfloat16, float, /*TM*/ 32, /*TK*/ 16, use::a,
263+ // layout::row_major, 1>();
264+ // // B - duplicate from 1x64x16
265+ // // Accumulator
266+ // test_get_coord_op<bfloat16, float, /*TM*/ 32, /*TN*/ 64, use::accumulator,
267+ // layout::row_major, 1>();
268+ // test_get_coord_op<float, float, /*TM*/ 32, /*TN*/ 64, use::accumulator,
269+ // layout::row_major, 1>();
270+
271+ // // 32x64x32 float/bfloat16
272+ // std::cout << "32x64x32 float/bfloat16" << std::endl;
273+ // // A
274+ // test_get_coord_op<bfloat16, float, /*TM*/ 32, /*TK*/ 32, use::a,
275+ // layout::row_major, 1>();
276+ // // B - duplicate from 1x64x32
277+ // // Accumulator
278+ // test_get_coord_op<bfloat16, float, /*TM*/ 32, /*TN*/ 64, use::accumulator,
279+ // layout::row_major, 1>();
280+ // test_get_coord_op<float, float, /*TM*/ 32, /*TN*/ 64, use::accumulator,
281+ // layout::row_major, 1>();
275282#endif
276283 break ;
277284 }
0 commit comments