Remove acl_device_program_info_t backpointer from acl_device_def_t

Author: sophimao

include/acl.h

@@ -593,20 +593,19 @@ typedef struct acl_device_def_autodiscovery_t {
 } acl_device_def_autodiscovery_t;
 
 typedef struct acl_device_def_t {
-  // Back pointer. This is only valid if this def is owned by a dev_prog. In
-  // tests, this is not the case.
-  acl_device_program_info dev_prog;
-  unsigned int physical_device_id; /* The ID of the physical device that the we
-                                      need to "talk" to */
+  // Information obtained from MMD when acl_hal_mmd tries the device
+  unsigned int physical_device_id; // The ID of the physical device that the we
+                                   // need to "talk" to
   unsigned int concurrent_reads;   // # of reads that can happen at one time
   unsigned int concurrent_writes;  // # of writes that can happen at one time
-  unsigned int
-      max_inflight_mem_ops; // max # of memory ops that can happen concurrently
+  unsigned int max_inflight_mem_ops; // max # of memory ops that can happen
+                                     // concurrently
   unsigned int host_capabilities;
   unsigned int shared_capabilities;
   unsigned int device_capabilities;
   size_t min_host_mem_alignment;
 
+  // autodiscovery information that changes every reprogram
   acl_device_def_autodiscovery_t autodiscovery_def;
 } acl_device_def_t;
 

include/acl_types.h

@@ -1264,9 +1264,12 @@ typedef struct _cl_device_id {
   cl_uint vendor_id;
   const char *version;
   const char *driver_version;
-  acl_device_def_t def;
   size_t min_local_mem_size; // min size of all local mems for accelerators in
                              // this device.
+  unsigned int address_bits; // cache address bits to avoid GetDeviceInfo calls
+
+  // MMD and autodiscovery information
+  acl_device_def_t def;
 
   // Indicates the number of contexts in which the device is currently opened.
   int opened_count;
@@ -1280,8 +1283,6 @@ typedef struct _cl_device_id {
   // this device are released)
   context_mode_lock mode_lock;
 
-  unsigned int address_bits; // cache address bits to avoid GetDeviceInfo calls
-
   int present; // Is the device present in the host system?
 
   // Error notification callback.
@@ -1296,15 +1297,19 @@ typedef struct _cl_device_id {
   // In non-embedded mode, we can reprogram the device during runtime of
   // the host program.
   // A device can only be programmed with one cl_program at a time.
+  // Two operations that can trigger device reprogram are
+  //   1. Eager reprogram of the first binary to clCreateProgram*
+  //   2. Kernel launch that use a different binary than the one on device
   //
   // These two fields keep track of what is programmed on the device:
   //
   //    loaded_bin:   What acl_device_binary_t is loaded *right now*.
+  //                  (updated in acl_program_device)
   //
   //    last_bin:     What acl_device_binary_t will be loaded on the device
-  //    after
-  //                   after the last operation on the device op queue
-  //                   has completed?
+  //                  after the last operation on the device op queue has
+  //                  completed? (updated when submitting command for the
+  //                  above-mentioned reprogram operations)
   //
   const acl_device_binary_t *loaded_bin = nullptr;
   const acl_device_binary_t *last_bin = nullptr;

src/acl_device_program_info.cpp

@@ -24,7 +24,6 @@ static std::string l_get_hashed_kernel_name(const std::string &kernel_name) {
 
 acl_device_program_info_t::acl_device_program_info_t() {
   device_binary.set_dev_prog(this);
-  device_binary.get_devdef().dev_prog = this;
 }
 
 acl_device_program_info_t::~acl_device_program_info_t() {
@@ -53,7 +52,6 @@ acl_device_program_info_t::add_split_binary(const std::string &hashed_name) {
 
   auto &dev_bin = m_split_device_binaries[hashed_name];
   dev_bin.set_dev_prog(this);
-  dev_bin.get_devdef().dev_prog = this;
   return dev_bin;
 }
 

test/acl_globals_test.cpp

@@ -149,8 +149,7 @@ static acl_system_def_t acltest_simple_system = {
 
     // Device definitions.
     1,
-    {{nullptr,
-      0,
+    {{0,
       1,
       1,
       1, /* half duplex memory transfers */
@@ -556,8 +555,7 @@ static acl_system_def_t acltest_complex_system = {
     /* num_devices */ 5,
     {// All of these have kernels 0 and 2 in common.  This is used for
      // testing clCreateKernels
-     {nullptr,
-      0,
+     {0,
       1,
       1,
       1, /* half duplex memory transfers */
@@ -591,8 +589,7 @@ static acl_system_def_t acltest_complex_system = {
                 0}},
           },
       }},
-     {nullptr,
-      1,
+     {1,
       1,
       1,
       2, /* full duplex memory transfers */
@@ -626,8 +623,7 @@ static acl_system_def_t acltest_complex_system = {
                 0}},
           },
       }},
-     {nullptr,
-      2,
+     {2,
       1,
       1,
       2, /* full duplex memory transfers */
@@ -651,8 +647,7 @@ static acl_system_def_t acltest_complex_system = {
                /* burst_interleaved */ 1,
            },
        }}},
-     {nullptr,
-      3,
+     {3,
       1,
       1,
       1, /* half duplex memory transfers */
@@ -676,8 +671,7 @@ static acl_system_def_t acltest_complex_system = {
                /* burst_interleaved */ 1,
            },
        }}},
-     {nullptr,
-      4,
+     {4,
       1,
       1,
       1, /* half duplex memory transfers */