Various features, fixes and hacks to make oneAPI-samples/Libraries/MPI/jacobian_solver SYCL run on PoCL-CPU and PoCL-R

CMakeLists.txt

@@ -1325,7 +1325,8 @@ cl_exp_pinned_buffers")
 set(HOST_DEVICE_FEATURES_30 "__opencl_c_3d_image_writes  __opencl_c_images \
 __opencl_c_atomic_order_acq_rel __opencl_c_atomic_order_seq_cst \
 __opencl_c_atomic_scope_device __opencl_c_program_scope_global_variables \
-__opencl_c_atomic_scope_all_devices __opencl_c_generic_address_space")
+__opencl_c_atomic_scope_all_devices __opencl_c_generic_address_space \
+__opencl_c_work_group_collective_functions")
 
 # Host CPU device: extensions only enabled when conformance is OFF
 if(NOT ENABLE_CONFORMANCE)

doc/sphinx/source/notes_6_0.rst

@@ -2,14 +2,6 @@
 Release Notes for PoCL 6.0
 **************************
 
-
-
-Minimal support for `cl_khr_priority_hints` and `cl_khr_throttle_hints` has been added.
-As the extension specification states that these hints provide no guarantees of
-any particular behavior (or lack thereof) they are treated as a no-op. However
-specifying them no longer causes `clCreateCommandQueueWithProperties` to return
-an error.
-
 ============================
 New device driver: cpu-tbb
 ============================
@@ -18,6 +10,16 @@ The cpu-tbb device driver uses the Intel oneAPI Threading Building Blocks (oneTB
 library for work-group and kernel-level task scheduling. Except for the
 task scheduler, the driver is identical to the original 'cpu' driver (pthread).
 
+=====================================
+Command queue priority/throttle hints
+=====================================
+
+Minimal support for `cl_khr_priority_hints` and `cl_khr_throttle_hints` has been added.
+As the extension specification states that these hints provide no guarantees of
+any particular behavior (or lack thereof) they are treated as a no-op. However
+specifying them no longer causes `clCreateCommandQueueWithProperties` to return
+an error.
+
 ===========================
 Driver-specific features
 ===========================
@@ -26,9 +28,13 @@ Driver-specific features
 CPU driver
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The 'cpu' driver gained support for using OpenMP for thread scheduling.
-Support is disabled by default, but can be enabled with CMake option. The
-'cpu-minimal' driver does not support OpenMP.
+ * Support for using OpenMP for task scheduling was added. It is disabled
+   by default, but can be enabled with CMake option. The 'cpu-minimal'
+   driver does not support OpenMP since it's supposed to be single-threaded.
+ * The CPU drivers can be now used for running SYCL programs compiled with
+   the oneAPI binary distributions of DPC++ by adding the following environment
+   settings: **POCL_DRIVER_VERSION_OVERRIDE=2023.16.7.0.21_160000 POCL_CPU_VENDOR_ID_OVERRIDE=32902**.
+ * Added support for the **__opencl_c_work_group_collective_functions** feature.
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Remote

doc/sphinx/source/using.rst

@@ -160,6 +160,14 @@ pocl.
  'cpu' device driver. The default is to determine this from the number of
  hardware threads available in the CPU.
 
+- **POCL_CPU_VENDOR_ID_OVERRIDE**
+
+ Overrides the vendor id reported by PoCL for the CPU drivers.
+ For example, setting the vendor id to be 32902 (0x8086) and setting the driver
+ version using **POCL_DRIVER_VER_OVERRIDE** to "2023.16.7.0.21_160000" (or such) can
+ be used to convince binary-distributed DPC++ compilers to compile and run SYCL
+ programs on the PoCL-CPU driver.
+
 - **POCL_DEBUG**
 
  Enables debug messages to stderr. This will be mostly messages from error
@@ -170,7 +178,8 @@ pocl.
 
  The old way (setting POCL_DEBUG to 1) has been updated to support categories.
  Using this limits the amount of debug messages produced. Current options are:
- error,warning,general,memory,llvm,events,cache,locking,refcounts,timing,hsa,tce,cuda,vulkan,proxy,all.
+ 'error', 'warning', 'general', 'memory', 'llvm', 'events', 'cache', 'locking',
+ 'refcounts', 'timing', 'hsa', 'tce', 'cuda', 'vulkan', 'proxy' and 'all'.
  Note: setting POCL_DEBUG to 1 still works and equals error+warning+general.
 
 - **POCL_DEBUG_LLVM_PASSES**
@@ -221,6 +230,11 @@ pocl.
  POCL_TTASIM0_PARAMETERS will be passed to the first ttasim driver instantiated
  and POCL_TTASIM1_PARAMETERS to the second one.
 
+- **POCL_DRIVER_VERSION_OVERRIDE**
+
+  Can be used to override the driver version reported by PoCL.
+  See **POCL_CPU_VENDOR_ID_OVERRIDE** for an example use case.
+
 - **POCL_EXTRA_BUILD_FLAGS**
 
  Adds the contents of the environment variable to all clBuildProgram() calls.

examples/boxadd/boxadd.c

@@ -112,5 +112,9 @@ main (int argc, char **argv)
   CHECK_CL_ERROR (clReleaseContext (context));
   CHECK_CL_ERROR (clUnloadPlatformCompiler (platform));
 
+  free (srcA);
+  free (srcB);
+  free (dst);
+
   return err;
 }

examples/matadd/matadd.c

@@ -108,5 +108,9 @@ main (int argc, char **argv)
   CHECK_CL_ERROR (clReleaseContext (context));
   CHECK_CL_ERROR (clUnloadPlatformCompiler (platform));
 
+  free (srcA);
+  free (srcB);
+  free (dst);
+
   return err;
 }

include/pocl.h

@@ -47,9 +47,11 @@
 /* detects restrict, variadic macros etc */
 #include "pocl_compiler_features.h"
 
-/* The maximum file, directory and path name lengths. TODO: These should be
-   detected from the filesystem properties of the execution platform. */
-#define POCL_MAX_DIRNAME_LENGTH 255
+/* The maximum file, directory and path name lengths.
+   NOTE: GDB seems to fail to load symbols from .so files which have
+   longer pathnames than 511, thus the quite small dir/filename length
+   limiter. */
+#define POCL_MAX_DIRNAME_LENGTH 64
 #define POCL_MAX_FILENAME_LENGTH (POCL_MAX_DIRNAME_LENGTH)
 #define POCL_MAX_PATHNAME_LENGTH 4096
 

lib/CL/clCreateProgramWithIL.c

@@ -128,16 +128,13 @@ CL_API_SUFFIX__VERSION_2_1
   POCL_GOTO_ERROR_COND ((length == 0), CL_INVALID_VALUE);
 
   int is_spirv = 0;
-#ifdef ENABLE_SPIRV
   int is_spirv_kernel
       = pocl_bitcode_is_spirv_execmodel_kernel ((const char *)il, length);
   is_spirv += is_spirv_kernel;
-#endif
-#ifdef ENABLE_VULKAN
+
   int is_spirv_shader
       = pocl_bitcode_is_spirv_execmodel_shader ((const char *)il, length);
   is_spirv += is_spirv_shader;
-#endif
 
   POCL_GOTO_ERROR_ON (
       (!is_spirv), CL_INVALID_VALUE,

lib/CL/clGetDeviceInfo.c

@@ -61,14 +61,14 @@ POname(clGetDeviceInfo)(cl_device_id   device,
   case CL_DEVICE_IMAGE_SUPPORT:
     POCL_RETURN_GETINFO(cl_bool, device->image_support);
   case CL_DEVICE_TYPE:
-    POCL_RETURN_GETINFO(cl_device_type, device->type);   
+    POCL_RETURN_GETINFO (cl_device_type, device->type);
   case CL_DEVICE_VENDOR_ID:
     POCL_RETURN_GETINFO(cl_uint, device->vendor_id);
   case CL_DEVICE_MAX_COMPUTE_UNITS:
     POCL_RETURN_GETINFO(cl_uint, device->max_compute_units);
   case CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS          :
     POCL_RETURN_GETINFO(cl_uint, device->max_work_item_dimensions);
-  case CL_DEVICE_MAX_WORK_GROUP_SIZE               : 
+  case CL_DEVICE_MAX_WORK_GROUP_SIZE:
     {
       size_t max_wg_size = device->max_work_group_size;
       POCL_RETURN_GETINFO(size_t, max_wg_size);
@@ -342,7 +342,7 @@ POname(clGetDeviceInfo)(cl_device_id   device,
   case CL_DEVICE_NON_UNIFORM_WORK_GROUP_SUPPORT:
     POCL_RETURN_GETINFO (cl_bool, device->non_uniform_work_group_support);
   case CL_DEVICE_WORK_GROUP_COLLECTIVE_FUNCTIONS_SUPPORT:
-    POCL_RETURN_GETINFO (cl_bool, CL_FALSE);
+    POCL_RETURN_GETINFO (cl_bool, device->wg_collective_func_support);
   case CL_DEVICE_GENERIC_ADDRESS_SPACE_SUPPORT:
     POCL_RETURN_GETINFO (cl_bool, device->generic_as_support);
   case CL_DEVICE_DEVICE_ENQUEUE_CAPABILITIES:

lib/CL/devices/common.c

@@ -1984,7 +1984,7 @@ pocl_setup_ils_with_version (cl_device_id dev)
     }
 }
 
-static const cl_name_version OPENCL_FEATURES[] = {
+static const cl_name_version OPENCL_C_FEATURES[] = {
   { CL_MAKE_VERSION (3, 0, 0), "__opencl_c_3d_image_writes" },
   { CL_MAKE_VERSION (3, 0, 0), "__opencl_c_images" },
   { CL_MAKE_VERSION (3, 0, 0), "__opencl_c_read_write_images" },
@@ -2013,15 +2013,15 @@ static const cl_name_version OPENCL_FEATURES[] = {
   { CL_MAKE_VERSION (3, 0, 0), "__opencl_c_ext_fp64_local_atomic_min_max" },
 };
 
-const size_t OPENCL_FEATURES_NUM
-    = sizeof (OPENCL_FEATURES) / sizeof (OPENCL_FEATURES[0]);
+const size_t OPENCL_C_FEATURES_NUM
+    = sizeof (OPENCL_C_FEATURES) / sizeof (OPENCL_C_FEATURES[0]);
 
 void
 pocl_setup_features_with_version (cl_device_id dev)
 {
   cl_name_version *tmp = NULL;
   unsigned ret = pocl_space_delim_string_to_cl_name_version_array (
-      &tmp, dev->features, OPENCL_FEATURES, OPENCL_FEATURES_NUM);
+      &tmp, dev->features, OPENCL_C_FEATURES, OPENCL_C_FEATURES_NUM);
 
   dev->num_opencl_features_with_version = ret;
   dev->opencl_features_with_version = tmp;

lib/CL/devices/common_utils.c

@@ -105,9 +105,8 @@ align_ptr (char *p)
 
 #define FALLBACK_MAX_THREAD_COUNT 8
 
-/* initializes CPU-specific device info struct members, that cannot / should
+/* Initializes CPU-specific device info default, that cannot / should
    not be initialized in pocl_init_default_device_infos() */
-
 cl_int
 pocl_cpu_init_common (cl_device_id device)
 {
@@ -139,6 +138,7 @@ pocl_cpu_init_common (cl_device_id device)
   device->features = HOST_DEVICE_FEATURES_30;
   device->run_program_scope_variables_pass = CL_TRUE;
   device->generic_as_support = CL_TRUE;
+  device->wg_collective_func_support = CL_TRUE;
 
   pocl_setup_opencl_c_with_version (device, CL_TRUE);
   pocl_setup_features_with_version (device);

lib/CL/devices/cpuinfo.c

@@ -305,15 +305,16 @@ pocl_cpuinfo_get_cpu_name_and_vendor(cl_device_id device)
   /* default vendor and vendor_id, in case it cannot be found by other means */
   device->vendor = cpuvendor_default;
   if (device->vendor_id == 0)
-    device->vendor_id = CL_KHRONOS_VENDOR_ID_POCL;
+    device->vendor_id = pocl_get_int_option ("POCL_CPU_VENDOR_ID_OVERRIDE",
+                                             CL_KHRONOS_VENDOR_ID_POCL);
 
   /* read contents of /proc/cpuinfo */
   if (access (cpuinfo, R_OK) != 0)
     return;
 
   FILE *f = fopen (cpuinfo, "r");
   char contents[MAX_CPUINFO_SIZE];
-  int num_read = fread (contents, 1, MAX_CPUINFO_SIZE - 1, f);            
+  int num_read = fread (contents, 1, MAX_CPUINFO_SIZE - 1, f);
   fclose(f);
   contents[num_read]='\0';
 

lib/CL/devices/devices.c

@@ -669,7 +669,9 @@ pocl_init_devices ()
              a shared global memory. */
           dev->global_mem_id = dev_index;
           POCL_INIT_OBJECT (dev);
-          dev->driver_version = POCL_VERSION_FULL;
+          dev->driver_version = pocl_get_string_option (
+              "POCL_DRIVER_VERSION_OVERRIDE", POCL_VERSION_FULL);
+
           if (dev->version == NULL)
             dev->version = "OpenCL 2.0 pocl";
 

lib/CL/devices/remote/remote.c

@@ -686,7 +686,7 @@ setup_relevant_devices (cl_program program, cl_device_id device,
   remote_server_data_t *server
       = ((remote_device_data_t *)device->data)->server;
   unsigned num_relevant_devices = 0;
-  char program_bc_path[POCL_MAX_FILENAME_LENGTH];
+  char program_bc_path[POCL_MAX_PATHNAME_LENGTH];
   unsigned i, j;
 
   for (i = 0; i < program->num_devices; ++i)

lib/CL/pocl_cache.c

@@ -92,12 +92,11 @@ void pocl_cache_program_path(char*        path,
   program_device_dir (path, program, device_i, "");
 }
 
-// required in llvm API
 void pocl_cache_program_bc_path(char*        program_bc_path,
                                 cl_program   program,
                                 unsigned     device_i) {
-    program_device_dir(program_bc_path, program,
-                       device_i, POCL_PROGRAM_BC_FILENAME);
+  program_device_dir (program_bc_path, program,
+                      device_i, POCL_PROGRAM_BC_FILENAME);
 }
 
 void
@@ -208,9 +207,9 @@ pocl_cache_kernel_cachedir (char *kernel_cachedir_path, cl_program program,
 {
   int bytes_written;
   char tempstring[POCL_MAX_PATHNAME_LENGTH];
-  char file_name[POCL_MAX_DIRNAME_LENGTH + 1];
+  char file_name[POCL_MAX_FILENAME_LENGTH + 1];
 
-  pocl_hash_clipped_name (kernel_name, POCL_MAX_DIRNAME_LENGTH, &file_name[0]);
+  pocl_hash_clipped_name (kernel_name, POCL_MAX_FILENAME_LENGTH, &file_name[0]);
 
   bytes_written
       = snprintf (tempstring, POCL_MAX_PATHNAME_LENGTH, "/%s", file_name);

lib/CL/pocl_cl.h

@@ -853,6 +853,7 @@ struct _cl_device_id {
   size_t preferred_wg_size_multiple;
   cl_bool non_uniform_work_group_support;
   cl_bool generic_as_support;
+  cl_bool wg_collective_func_support;
   cl_uint preferred_vector_width_char;
   cl_uint preferred_vector_width_short;
   cl_uint preferred_vector_width_int;

lib/kernel/host/CMakeLists.txt

@@ -139,6 +139,7 @@ vload_store_half_f16c.c
 vstore.cl
 vstore_half.cl
 wait_group_events.cl
+work_group.c
 write_image.cl
 
 ###################################################################

lib/kernel/subgroups.c

@@ -29,37 +29,11 @@
 
 #include <math.h>
 
-/**
- * \brief Internal pseudo function which allocates space from the work-group
- * thread's stack (basically local memory) for each work-item.
- *
- * It's expanded in WorkitemLoops.cc to an alloca().
- *
- * @param element_size The size of an element to allocate (for all WIs in the
- * WG).
- * @param align The alignment of the start of chunk.
- * @param extra_bytes extra bytes to add to the allocation, some functions need
- * extra space
- * @return pointer to the allocated stack space (freed at unwind).
- */
-void *__pocl_work_group_alloca (size_t element_size, size_t align,
-                                size_t extra_bytes);
-
-/**
- * \brief Internal pseudo function which allocates space from the work-group
- * thread's stack (basically local memory).
- *
- * It's expanded in WorkitemLoops.cc to an alloca().
- *
- * @param bytes The size of data to allocate in bytes.
- * @param align The alignment of the start of chunk.
- * @return pointer to the allocated stack space (freed at unwind).
- */
-void *__pocl_local_mem_alloca (size_t bytes, size_t align);
-
-size_t _CL_OVERLOADABLE get_local_size (unsigned int dimindx);
+#include "work_group_alloca.h"
 
 size_t _CL_OVERLOADABLE get_local_id (unsigned int dimindx);
+size_t _CL_OVERLOADABLE get_local_linear_id (void);
+size_t _CL_OVERLOADABLE get_local_size (unsigned int dimindx);
 
 /* Magic variable that is expanded in Workgroup.cc */
 extern uint _pocl_sub_group_size;
@@ -89,8 +63,6 @@ get_enqueued_num_sub_groups (void)
   return 1;
 }
 
-size_t _CL_OVERLOADABLE get_local_linear_id (void);
-
 uint _CL_OVERLOADABLE
 get_sub_group_id (void)
 {