[SYCL][Joint Matrix] Implement bfloat16 32x64x32 combination support

intel · Oct 7, 2024 · d2c18ca · d2c18ca
1 parent 6e095fd
commit d2c18ca
Show file tree

Hide file tree

Showing 4 changed files with 7 additions and 0 deletions.
diff --git a/sycl/source/detail/device_info.hpp b/sycl/source/detail/device_info.hpp
@@ -858,6 +858,8 @@ struct get_device_info_impl<
            matrix_type::fp32, matrix_type::fp32},
           {0, 0, 0, 32, 64, 16, matrix_type::bf16, matrix_type::bf16,
            matrix_type::fp32, matrix_type::fp32},
+          {0, 0, 0, 32, 64, 32, matrix_type::bf16, matrix_type::bf16,
+           matrix_type::fp32, matrix_type::fp32},
           {8, 0, 0, 0, 16, 8, matrix_type::tf32, matrix_type::tf32,
            matrix_type::fp32, matrix_type::fp32},
       };

diff --git a/sycl/test-e2e/Matrix/element_wise_all_ops_impl.hpp b/sycl/test-e2e/Matrix/element_wise_all_ops_impl.hpp
@@ -228,7 +228,9 @@ int main() {
 // This combination is not currently supported for sub group size = 32 in IGC
 #if (!defined(SG_SZ) || SG_SZ != 32)
         test_ewops_ab<bfloat16, 32, 16, use::a, layout::row_major, 1>();
+        test_ewops_ab<bfloat16, 32, 32, use::a, layout::row_major, 1>();
         test_ewops_ab<bfloat16, 16, 64, use::b, layout::ext_intel_packed, 2>();
+        test_ewops_ab<bfloat16, 32, 64, use::b, layout::ext_intel_packed, 2>();
         test_ewops_c<float, 1, 64>();
         test_ewops_c<float, 32, 64>();
 #endif

diff --git a/sycl/test-e2e/Matrix/element_wise_ops_impl.hpp b/sycl/test-e2e/Matrix/element_wise_ops_impl.hpp
@@ -134,6 +134,7 @@ int main() {
       passed &= test<bfloat16, float, 16, 16, 16, 2, class pvc_bf16_16x16x16>();
       passed &= test<bfloat16, float, 1, 64, 16, 2, class pvc_bf16_1x64x16>();
       passed &= test<bfloat16, float, 32, 64, 16, 2, class pvc_bf16_32x64x16>();
+      passed &= test<bfloat16, float, 32, 64, 32, 2, class pvc_bf16_32x64x32>();
 #endif
       break;
     }

diff --git a/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB_impl.hpp b/sycl/test-e2e/Matrix/joint_matrix_rowmajorA_rowmajorB_impl.hpp
@@ -129,6 +129,8 @@ int main() {
                               float>();
         res += gemm_row_major<32, 64, 16, class bf16_32x64x16, bfloat16,
                               bfloat16, float>();
+        res += gemm_row_major<32, 64, 32, class bf16_32x64x32, bfloat16,
+                              bfloat16, float>();
       }
       break;
     }