[SYCL][Joint Matrix] Test combinations are queried Part 4 (#14019)

Supported matrix dimensions are queried from the device, and inform the
tests which tile sizes one can use.

This is a subset of all tests that are planned to be modified.

Test manually tested on PVC and SPR - no new regresssions

The following tests have been marked as XFAIL on all platforms. I
removed them from XMX8 folder. Once they are passing then they can be
modified to query the supported matrix dimensions form the device.
* joint_matrix_colA_rowB_colC.cpp
* joint_matrix_out_bounds.cpp
* joint_matrix_unaligned_k.cpp

Author: artemrad

sycl/test-e2e/Matrix/SG32/joint_matrix_bf16_fill_k_cache.cpp

@@ -5,21 +5,18 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
+// UNSUPPORTED: gpu-intel-dg2
+// REQUIRES: aspect-ext_intel_matrix
 // REQUIRES-INTEL-DRIVER: lin: 27501, win: 101.4943
 
-// RUN: %{build} -o %t_gpu.out -ffp-model=precise
-// RUN: %if gpu %{ %{run} %t_gpu.out %}
-
-// RUN: %{build}  -ffp-model=precise -o %t_cpu.out -DtM=16 -DtK=32 -DNCACHE1=32 -DKCACHE1=32
-// RUN: %if cpu %{ %{run} %t_cpu.out %}
+// RUN: %{build} -o %t.out -ffp-model=precise
+// RUN: %{run} %t.out
 
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "../common.hpp"
-#include <cstddef>
 
 #define SG_SZ 32
-constexpr size_t TN = 16;
 
 #include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

sycl/test-e2e/Matrix/SG32/joint_matrix_bf16_fill_k_cache_init.cpp

@@ -5,7 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix, gpu
+// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
+// UNSUPPORTED: gpu-intel-dg2
+// REQUIRES: aspect-ext_intel_matrix, gpu
 // REQUIRES-INTEL-DRIVER: lin: 27501, win: 101.4943
 
 // RUN: %{build} -o %t.out -DINIT_LIST -ffp-model=precise
@@ -14,9 +16,7 @@
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "../common.hpp"
-#include <cstddef>
 
 #define SG_SZ 32
-constexpr size_t TN = 16;
 
 #include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

sycl/test-e2e/Matrix/SG32/joint_matrix_bf16_fill_k_cache_unroll.cpp

@@ -5,23 +5,20 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
+// UNSUPPORTED: gpu-intel-dg2
+// REQUIRES: aspect-ext_intel_matrix
 // REQUIRES-INTEL-DRIVER: lin: 27501, win: 101.4943
 
-// RUN: %{build} -mllvm -inline-threshold=5000 -ffp-model=precise -o %t_gpu.out -DMANUAL_UNROLL
-// RUN: %if gpu %{ %{run} %t_gpu.out %}
-
-// RUN: %{build} -mllvm -inline-threshold=5000 -ffp-model=precise -o %t_cpu.out -DMANUAL_UNROLL -DtM=16 -DtK=32 -DNCACHE1=32 -DKCACHE1=32
-// RUN: %if cpu %{ %{run} %t_cpu.out %}
+// RUN: %{build} -mllvm -inline-threshold=5000 -ffp-model=precise -o %t.out -DMANUAL_UNROLL
+// RUN: %{run} %t.out
 
 // -mllvm -inline-threshold added as a workaround,
 // since IGC doesn't support some variants of IR for Joint Matrix currently
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "../common.hpp"
-#include <cstddef>
 
 #define SG_SZ 32
-constexpr size_t TN = 16;
 
 #include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

sycl/test-e2e/Matrix/SG32/joint_matrix_bf16_fill_k_cache_unroll_init.cpp

@@ -5,7 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix, gpu
+// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
+// UNSUPPORTED: gpu-intel-dg2
+// REQUIRES: aspect-ext_intel_matrix, gpu
 // REQUIRES-INTEL-DRIVER: lin: 27501, win: 101.4943
 
 // RUN: %{build} -mllvm -inline-threshold=5000 -ffp-model=precise -o %t_gpu.out -DINIT_LIST -DMANUAL_UNROLL
@@ -16,9 +18,7 @@
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "../common.hpp"
-#include <cstddef>
 
 #define SG_SZ 32
-constexpr size_t TN = 16;
 
 #include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

sycl/test-e2e/Matrix/XMX8/joint_matrix_bf16_fill_k_cache.cpp

@@ -5,19 +5,12 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// REQUIRES: aspect-ext_intel_matrix
 
-// RUN: %{build} -o %t_gpu.out -ffp-model=precise
-// RUN: %if gpu %{ %{run} %t_gpu.out %}
-
-// RUN: %{build}  -ffp-model=precise -o %t_cpu.out -DtM=16 -DtK=32 -DNCACHE1=32 -DKCACHE1=32
-// RUN: %if cpu %{ %{run} %t_cpu.out %}
+// RUN: %{build} -o %t.out -ffp-model=precise
+// RUN: %{run} %t.out
 
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "common.hpp"
-#include <cstddef>
-
-constexpr size_t TN = 16;
-
 #include "joint_matrix_bf16_fill_k_cache_impl.hpp"

sycl/test-e2e/Matrix/XMX8/joint_matrix_bf16_fill_k_cache_init.cpp

@@ -16,9 +16,4 @@
 // -ffp-model=precise is added to not depend on compiler defaults.
 
 #include "common.hpp"
-#include <cstddef>
-
-#define SG_SZ 16
-constexpr size_t TN = 16;
-
 #include "joint_matrix_bf16_fill_k_cache_impl.hpp"