|
| 1 | +# Test plan for dynamic linking of device code |
| 2 | + |
| 3 | +This is a test plan for device code dynamic linking feature described in |
| 4 | +https://github.com/intel/llvm/blob/sycl/sycl/doc/SharedLibraries.md document. |
| 5 | + |
| 6 | +## 1. Testing scope |
| 7 | + |
| 8 | +### 1.1. Device & BE coverage |
| 9 | + |
| 10 | +All of the tests described below are performed on Intel CPU and GPU devices. |
| 11 | +For GPU device both OpenCL and L0 backends are used. |
| 12 | + |
| 13 | +**NOTE**: The feature isn't supported on FPGA device. |
| 14 | + |
| 15 | +### 1.2. AOT & JIT compilation |
| 16 | + |
| 17 | +All the tests are compiled using AOT, JIT and mixed (AOT with JIT) compilation. |
| 18 | +Mixed compilation is the case when application is built using AOT and used |
| 19 | +shared library uses JIT or vice versa. |
| 20 | + |
| 21 | +### 1.3. Device code split modes |
| 22 | + |
| 23 | +All the tests are compiled using different combination of device code split |
| 24 | +mode. |
| 25 | +There are four possible device code split modes that can be requested via |
| 26 | +`fsycl-device-code-split` compiler's option: |
| 27 | +- `per_kernel` - device code module is created for each SYCL kernel |
| 28 | +- `per_source` - device code module is created for each source |
| 29 | +- `off` - no device code split |
| 30 | +- `auto` - use heuristic to select the best way of splitting device code |
| 31 | + |
| 32 | +For each test the main application and all shared libraries are compiled with |
| 33 | +all the variations of device code split modes. This results in four binaries for |
| 34 | +each application or library. |
| 35 | +Each tests covers all the combinations of device code split modes used for |
| 36 | +the application and libraries, i.e. all binaries of the application are checked |
| 37 | +with all binaries for shared libraries. |
| 38 | +For one application and one shared library examples of the combinations will be: |
| 39 | +- Application compiled with `per_kernel` and library compiled with `per_kernel` |
| 40 | +- Application compiled with `per_kernel` and library compiled with `per_source` |
| 41 | +- Application compiled with `per_kernel` and library compiled with `off` |
| 42 | +- Application compiled with `per_kernel` and library compiled with `auto` |
| 43 | + |
| 44 | +- Application compiled with `per_source` and library compiled with `per_kernel` |
| 45 | +- Application compiled with `per_source` and library compiled with `per_source` |
| 46 | +- Application compiled with `per_source` and library compiled with `off` |
| 47 | +- Application compiled with `per_source` and library compiled with `auto` |
| 48 | + |
| 49 | +and etc. Total 16 combinations. |
| 50 | + |
| 51 | +## 2. Tests |
| 52 | + |
| 53 | +All the tests consist of at least one main application and at least one shared |
| 54 | +library. |
| 55 | +For all the tests simple calculation is performed by the application's kernel(s) |
| 56 | +that used `SYCL_EXTERNAL` from the shared library. Example : addition of |
| 57 | +vectors. |
| 58 | +The result of the calculation must be checked on host. |
| 59 | + |
| 60 | +### 2.1. Basic test with shared library |
| 61 | + |
| 62 | +- Several kernels and `SYCL_EXTERNAL` functions are defined in both library and |
| 63 | + application. |
| 64 | +- At least one `SYCL_EXTERNAL` function defined in shared library is used from |
| 65 | + the application's kernel. |
| 66 | +- At least one `SYCL_EXTERNAL` function defined in shared library is used in |
| 67 | + kernel inside the library. |
| 68 | +- Defined kernels are submitted and ran. |
| 69 | + |
| 70 | +### 2.2. Test with dependent shared libraries |
| 71 | + |
| 72 | +The test consists of the main application and **two** shared libraries libA and |
| 73 | +libB and performs the following: |
| 74 | +- Application defines kernel(s) with dependency on libA. |
| 75 | +- libA defines `SYCL_EXTERNAL` function(s) with dependency on libB. |
| 76 | +- libB defines `SYCL_EXTERNAL` function(s) without external dependencies. |
| 77 | +- Defined kernels are submitted and ran. |
| 78 | + |
| 79 | +### 2.3. Tests with function pointers |
| 80 | + |
| 81 | +#### 2.3.1. Variation 1 |
| 82 | + |
| 83 | +- Application defines a function with `indirectly_callable` attribute. |
| 84 | +- Shared library defines host API that can receive function pointer as an |
| 85 | + argument and runs a kernel that calls through passed function pointer. |
| 86 | +- Address of `indirectly_callable` function is taken and passed to shared |
| 87 | + library's host API. |
| 88 | + |
| 89 | +#### 2.3.2. Variation 2 |
| 90 | + |
| 91 | +- Shared library defines a function with `indirectly_callable` attribute. |
| 92 | +- Shared library defines host API that can return a function pointer to |
| 93 | + a function with `indirectly_callable` attribute. |
| 94 | +- Application used host API of shared library and calls through received |
| 95 | + function pointer from a kernel. |
| 96 | + |
| 97 | +### 2.4. Test with shared library without device code |
| 98 | + |
| 99 | +The test aimed to check that support for device code inside shared library |
| 100 | +doesn't break support for host-only shared library. |
| 101 | + |
| 102 | +- Shared library defines host-only non-`SYCL_EXTERNAL` function. |
| 103 | +- The function from shared library is used by host part of the application. |
| 104 | +- Application defines and runs a kernel with a simple calculation. |
| 105 | + |
| 106 | +### 2.5. Tests with run-time loaded shared library |
| 107 | + |
| 108 | +Tests aimed to check that support for dynamic linking of device code doesn't |
| 109 | +break run-time loading of shared library. |
| 110 | + |
| 111 | +#### 2.5.1. Variation 1 |
| 112 | + |
| 113 | +The test makes sure that run-time loading of host-only shared libraries is |
| 114 | +not broken. |
| 115 | + |
| 116 | +- Shared library defines several host-only external functions |
| 117 | +- Shared library doesn't contain any device code |
| 118 | +- There is no explicit dependencies between application and library |
| 119 | +- Shared library is loaded via dlopen/LoadLibrary functionality |
| 120 | +- Functions from the shared library are used by host part of the application |
| 121 | + to perform a simple calculation |
| 122 | +- Result of the simple calculation is checked |
| 123 | + |
| 124 | +#### 2.5.2. Variation 2 |
| 125 | + |
| 126 | +The test makes sure that run-time loading of shared libraries with embedded |
| 127 | +device code is not broken. |
| 128 | + |
| 129 | +- Shared library defines several host-only external functions, several |
| 130 | + `SYCL_EXTERNAL` functions and several kernels |
| 131 | +- There is no explicit dependencies between application and library on |
| 132 | + device and host |
| 133 | +- Shared library is loaded via dlopen/LoadLibrary functionality |
| 134 | +- Host functions from the shared library are used by host part of the |
| 135 | + application to perform a simple calculation |
| 136 | +- Result of the simple calculation is checked |
| 137 | + |
| 138 | +### 2.6. Test with two versions of shared library |
| 139 | + |
| 140 | +The test is compiled with version 1 of the shared library and ran with |
| 141 | +shared library of version 2. |
| 142 | +Code can be distributed in the same way as in basic test, but `SYCL_EXTERNAL` |
| 143 | +functions defined in shared library of version 2, should have a different |
| 144 | +definition comparing to version 1, so it is possible to check that correct |
| 145 | +version of the shared library were used. |
| 146 | + |
| 147 | +### 2.7. Test with kernel bundle |
| 148 | + |
| 149 | +The tests aim to check that operations with kernel bundles work correctly in |
| 150 | +co-existence with dynamic linking feature. |
| 151 | + |
| 152 | +#### 2.7.1. Variation 1 |
| 153 | + |
| 154 | +- Shared library defines at least one `SYCL_EXTERNAL` function |
| 155 | +- Application defines a kernel with dependency on `SYCL_EXTERNAL` function(s) |
| 156 | + from the shared library |
| 157 | +- Application gets a `kernel_bundle` object that contains kernel with |
| 158 | + dependency (can be done via getting `kernel_id` object and using it as a |
| 159 | + parameter for `get_kernel_bundle` free function overload). |
| 160 | +- Application builds the `kernel_bundle` object and uses it to run the kernel |
| 161 | + with dependency |
| 162 | + |
| 163 | +#### 2.7.2. Variation 2 |
| 164 | + |
| 165 | +- Shared library defines at least one `SYCL_EXTERNAL` function |
| 166 | +- Shared library defines several kernels, at least one kernel from shared |
| 167 | + library uses `SYCL_EXTERNAL` function |
| 168 | +- Application defines a kernel with dependency on `SYCL_EXTERNAL` function(s) |
| 169 | + from the shared library |
| 170 | +- Shared library provides a host function that returns `kernel_bundle` object |
| 171 | + containing kernels from the shared library |
| 172 | +- Application gets a `kernel_bundle` object that contains kernel with |
| 173 | + dependency and links it using link API with `kernel_bundle` that contains |
| 174 | + kernels from shared library |
| 175 | +- Resulting `kernel_bundle` object is used to run the kernel with dependency |
| 176 | + on shared library |
| 177 | + |
| 178 | +### 2.8. Test with several applications |
| 179 | + |
| 180 | +Same as [Basic test with shared library](#Basic-test-with-shared-library) but |
| 181 | +two or more applications using same shared library are ran at the same time. |
0 commit comments