[SYCL][Fusion] Improve kernel fusion e2e tests (#14366)

Improve the kernel fusion end-to-end tests:
* Remove the flag to embed IR on CUDA and AMD backend from the tests
that abort execution before requiring access to IR.
* Check for the number of kernel launches to make sure fusion did not
fail silently.

---------

Signed-off-by: Lukas Sommer <[email protected]>

Author: sommerlukas

sycl/test-e2e/KernelFusion/cancel_fusion.cpp

@@ -1,8 +1,12 @@
-// RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: %{build} -o %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test cancel fusion
 
+// As fusion is cancelled, this test launches two kernels.
+// CHECK-COUNT-2: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>

sycl/test-e2e/KernelFusion/complete_fusion.cpp

@@ -1,8 +1,12 @@
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test complete fusion without any internalization
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>

sycl/test-e2e/KernelFusion/cooperative_kernel.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=2 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test cooperative kernels are not fused

sycl/test-e2e/KernelFusion/different_nd_ranges.cpp

@@ -1,11 +1,13 @@
-// RUN: %{build} -o %t.out
-// RUN: env SYCL_RT_WARNING_LEVEL=1 SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=16:32:64 \
-// RUN:   %{run} %t.out 2>&1 | FileCheck %s
+// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: env SYCL_PI_TRACE=2 env SYCL_RT_WARNING_LEVEL=1 \
+// RUN: SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=16:32:64 %{run} %t.out 2>&1 \
+// RUN: | FileCheck %s --implicit-check-not "ERROR: JIT compilation for kernel fusion failed with message:"
 
 // Test complete fusion of kernels with different ND-ranges.
 
 // Kernels with different ND-ranges should be fused.
-// CHECK-NOT: Cannot fuse kernels with different offsets or local sizes
+// CHECK-COUNT-26: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
 
 #include <sycl/detail/core.hpp>
 
@@ -252,22 +254,6 @@ int main() {
   const auto R5 = {5ul};
   test({RangeDesc{{10}, R5}, RangeDesc{{20}, R5}, RangeDesc{{30}, R5}});
 
-  // Two 1-D kernels with different global sizes and a 2-D kernel with more
-  // work-items and specified (equal) local size.
-  test({RangeDesc{{10}, R2}, RangeDesc{{20}, R2}, RangeDesc{{10, 10}, {2, 1}}});
-
-  // Three 2-D kernels with different global sizes.
-  test({RangeDesc{{10, 15}, {2, 5}}, RangeDesc{{20, 10}, {2, 5}},
-        RangeDesc{{10, 5}, {2, 5}}});
-
-  // Three 3-D kernels with different global sizes.
-  test({RangeDesc{{10, 4, 2}, {5, 2, 1}}, RangeDesc{{20, 2, 4}, {5, 2, 1}},
-        RangeDesc{{10, 2, 4}, {5, 2, 1}}});
-
-  // 1-D, 2-D and 3-D kernels with different global sizes.
-  test({RangeDesc{{10}, R5}, RangeDesc{{10, 1}, {5, 1}},
-        RangeDesc{{10, 1, 1}, {5, 1, 1}}});
-
   // Test global sizes that trigger the rounded range kernel insertion.
   // Note that we lower the RR threshold when running this test.
   test({RangeDesc{67}, RangeDesc{87}, RangeDesc{64}});

sycl/test-e2e/KernelFusion/event_wait_cancel.cpp

@@ -1,5 +1,5 @@
 // REQUIRES: aspect-usm_shared_allocations
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 
 // Test validity of events after cancel_fusion.

sycl/test-e2e/KernelFusion/internal_explicit_dependency.cpp

@@ -1,10 +1,14 @@
 // REQUIRES: aspect-usm_shared_allocations
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test complete fusion where one kernel in the fusion list specifies an
 // explicit dependency (via events) on another kernel in the fusion list.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include "fusion_event_test_common.h"
 
 #include <sycl/detail/core.hpp>

sycl/test-e2e/KernelFusion/jit_caching_multitarget.cpp

@@ -1,5 +1,5 @@
 // REQUIRES: (gpu && (hip || cuda)), cpu
-// RUN: %{build} -fsycl-embed-ir -O2 -o %t.out
+// RUN: %{build} %{embed-ir} -O2 -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run-unfiltered-devices} %t.out 2>&1 | FileCheck %s --implicit-check-not "WRONG a VALUE" --implicit-check-not "WRONG b VALUE"
 
 // Test caching for JIT fused kernels when devices with different architectures

sycl/test-e2e/KernelFusion/math_function.cpp

@@ -1,8 +1,12 @@
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test fusion of a kernel using a math function.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 
 #include <sycl/builtins.hpp>

sycl/test-e2e/KernelFusion/non-kernel-cg.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=2 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test non-kernel device command groups are not fused

sycl/test-e2e/KernelFusion/non_unit_local_size.cpp

@@ -1,9 +1,13 @@
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test complete fusion with local internalization specified on the
 // accessors, where each work-item processes multiple data-items.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>

sycl/test-e2e/KernelFusion/sync_acc_mem_op.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_buffer_destruction.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_event_wait.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test fusion cancellation on event::wait() happening before

sycl/test-e2e/KernelFusion/sync_host_accessor.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_host_task.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_queue_destruction.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_queue_wait.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_second_queue.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Windows doesn't yet have full shutdown().

sycl/test-e2e/KernelFusion/sync_two_queues_requirement.cpp

@@ -1,5 +1,5 @@
 // For this test, complete_fusion must be supported.
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 
 // Test fusion cancellation for requirement between two active fusions.

sycl/test-e2e/KernelFusion/sync_usm_mem_op.cpp

@@ -1,4 +1,4 @@
-// RUN: %{build} %{embed-ir} -o %t.out
+// RUN: %{build} -o %t.out
 // RUN: env SYCL_RT_WARNING_LEVEL=1 %{run} %t.out 2>&1 | FileCheck %s
 // Windows doesn't yet have full shutdown().
 // UNSUPPORTED: ze_debug && windows

sycl/test-e2e/KernelFusion/usm_no_dependencies.cpp

@@ -1,9 +1,13 @@
 // REQUIRES: aspect-usm_shared_allocations
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test complete fusion using USM pointers.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>

sycl/test-e2e/KernelFusion/work_group_barrier.cpp

@@ -1,9 +1,13 @@
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test complete fusion with a combination of kernels that require a work-group
 // barrier to be inserted by fusion.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>

sycl/test-e2e/KernelFusion/wrapped_usm.cpp

@@ -1,9 +1,13 @@
 // REQUIRES: aspect-usm_shared_allocations
 // RUN: %{build} %{embed-ir} -o %t.out
-// RUN: %{run} %t.out
+// RUN: env SYCL_PI_TRACE=2 %{run} %t.out | FileCheck %s
 
 // Test complete fusion using an wrapped USM pointer as kernel functor argument.
 
+// The two kernels are fused, so only a single, fused kernel is launched.
+// CHECK-COUNT-1: piEnqueueKernelLaunch
+// CHECK-NOT: piEnqueueKernelLaunch
+
 #include <sycl/detail/core.hpp>
 #include <sycl/ext/codeplay/experimental/fusion_wrapper.hpp>
 #include <sycl/properties/all_properties.hpp>