[#475][Refactor] Remove UVA + umalloc (#507)

libocca · Sep 7, 2021 · 845c524 · 845c524
1 parent 9446605
commit 845c524
Show file tree

Hide file tree

Showing 122 changed files with 139 additions and 2,139 deletions.
diff --git a/docs/guide/_sidebar.md b/docs/guide/_sidebar.md
@@ -8,7 +8,6 @@
   - [Introduction](/guide/occa/introduction)
   - [Device Streams](/guide/occa/device-streams)
   - [Background Device](/guide/occa/background-device)
-  - [Unified Memory](/guide/occa/unified-memory)
 
 - **OKL**
   - [Introduction](/guide/okl/introduction)

diff --git a/docs/guide/occa/unified-memory.md b/docs/guide/occa/unified-memory.md
diff --git a/examples/c/01_add_vectors/main.c b/examples/c/01_add_vectors/main.c
@@ -69,12 +69,13 @@ int main(int argc, const char **argv) {
   //========================================================
 
   // Allocate memory on the device
-  o_a  = occaDeviceTypedMalloc(device, entries, occaDtypeFloat, NULL, occaDefault);
-  o_b  = occaDeviceTypedMalloc(device, entries, occaDtypeFloat, NULL, occaDefault);
+  o_a  = occaDeviceTypedMalloc(device, entries, occaDtypeFloat, a, occaDefault);
+  o_b  = occaDeviceTypedMalloc(device, entries, occaDtypeFloat, b, occaDefault);
 
-  // We can also allocate memory without a dtype
+  // We can also allocate memory without a dtype and manually copy the data over
   // WARNING: This will disable runtime type checking
   o_ab = occaDeviceMalloc(device, entries * sizeof(float), NULL, occaDefault);
+  occaCopyPtrToMem(o_ab, ab, occaAllBytes, 0, occaDefault);
 
   // Setup properties that can be passed to the kernel
   occaJson props = occaCreateJson();
@@ -86,10 +87,6 @@ int main(int argc, const char **argv) {
                                      "addVectors",
                                      props);
 
-  // Copy memory to the device
-  occaCopyPtrToMem(o_a, a, entries*sizeof(float), 0, occaDefault);
-  occaCopyPtrToMem(o_b, b, occaAllBytes         , 0, occaDefault);
-
   // Launch device kernel
   occaKernelRun(addVectors,
                 occaInt(entries), o_a, o_b, o_ab);
@@ -132,7 +129,7 @@ occaJson parseArgs(int argc, const char **argv) {
     "      shortname: 'd',"
     "      description: 'Device properties (default: \"{mode: \\'Serial\\'}\")',"
     "      with_arg: true,"
-    "      default_value: { mode: 'Serial' },"
+    "      default_value: \"{ mode: 'Serial' }\","
     "    },"
     "    {"
     "      name: 'verbose',"

diff --git a/examples/c/02_background_device/main.c b/examples/c/02_background_device/main.c
@@ -37,33 +37,30 @@ int main(int argc, const char **argv) {
   int entries = 5;
   int i;
 
-  float *a  = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
-  float *b  = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
-  float *ab = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
+  float *a  = (float*) malloc(entries*sizeof(float));
+  float *b  = (float*) malloc(entries*sizeof(float));
+  float *ab = (float*) malloc(entries*sizeof(float));
 
   for (i = 0; i < entries; ++i) {
     a[i]  = i;
     b[i]  = 1 - i;
     ab[i] = 0;
   }
 
+  // Allocate memory on the background device
+  occaMemory o_a  = occaTypedMalloc(entries, occaDtypeFloat, a, occaDefault);
+  occaMemory o_b  = occaTypedMalloc(entries, occaDtypeFloat, b, occaDefault);
+  occaMemory o_ab = occaTypedMalloc(entries, occaDtypeFloat, ab, occaDefault);
+
   occaKernel addVectors = occaBuildKernel("addVectors.okl",
                                           "addVectors",
                                           occaDefault);
 
-  // Arrays a, b, and ab are now resident
-  //   on [device]
   occaKernelRun(addVectors,
-                occaInt(entries), occaPtr(a), occaPtr(b), occaPtr(ab));
-
-  // b is not const in the kernel, so we can use
-  //   dontSync(b) to manually force b to not sync
-  occaDontSync(b);
+                occaInt(entries), o_a, o_b, o_ab);
 
-  // Finish work queued up in [device],
-  //   synchronizing a, b, and ab and
-  //   making it safe to use them again
-  occaFinish();
+  // Copy result to the host
+  occaCopyMemToPtr(ab, o_ab, occaAllBytes, 0, occaDefault);
 
   for (i = 0; i < entries; ++i) {
     printf("%d = %f\n", i, ab[i]);
@@ -74,11 +71,17 @@ int main(int argc, const char **argv) {
     }
   }
 
+  // Free host memory
+  free(a);
+  free(b);
+  free(ab);
+
+  // Free device memory and occa objects
   occaFree(&args);
   occaFree(&addVectors);
-  occaFreeUvaPtr(a);
-  occaFreeUvaPtr(b);
-  occaFreeUvaPtr(ab);
+  occaFree(&o_a);
+  occaFree(&o_b);
+  occaFree(&o_ab);
 
   return 0;
 }
@@ -94,7 +97,7 @@ occaJson parseArgs(int argc, const char **argv) {
     "      shortname: 'd',"
     "      description: 'Device properties (default: \"{ mode: \\'Serial\\' }\")',"
     "      with_arg: true,"
-    "      default_value: { mode: 'Serial' },"
+    "      default_value: \"{ mode: 'Serial' }\","
     "    },"
     "    {"
     "      name: 'verbose',"

diff --git a/examples/c/03_generic_inline_kernel/main.c b/examples/c/03_generic_inline_kernel/main.c
@@ -27,16 +27,21 @@ int main(int argc, const char **argv) {
   int entries = 5;
   int i;
 
-  float *a  = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
-  float *b  = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
-  float *ab = (float*) occaTypedUMalloc(entries, occaDtypeFloat, NULL, occaDefault);
+  float *a  = (float*) malloc(entries*sizeof(float));
+  float *b  = (float*) malloc(entries*sizeof(float));
+  float *ab = (float*) malloc(entries*sizeof(float));
 
   for (i = 0; i < entries; ++i) {
     a[i]  = i;
     b[i]  = 1 - i;
     ab[i] = 0;
   }
 
+  // Allocate memory on the background device
+  occaMemory o_a  = occaTypedMalloc(entries, occaDtypeFloat, a, occaDefault);
+  occaMemory o_b  = occaTypedMalloc(entries, occaDtypeFloat, b, occaDefault);
+  occaMemory o_ab = occaTypedMalloc(entries, occaDtypeFloat, ab, occaDefault);
+
   occaJson props = occaCreateJson();
   occaJsonObjectSet(props,
                     "defines/TILE_SIZE",
@@ -46,10 +51,10 @@ int main(int argc, const char **argv) {
 
   // Build the variable scope used inside the inlined OKL code
   occaScopeAddConst(scope, "entries", occaInt(entries));
-  occaScopeAddConst(scope, "a", occaPtr(a));
-  occaScopeAddConst(scope, "b", occaPtr(b));
+  occaScopeAddConst(scope, "a", o_a);
+  occaScopeAddConst(scope, "b", o_b);
   // We can name our scoped variales anything
-  occaScopeAdd(scope, "output", occaPtr(ab));
+  occaScopeAdd(scope, "output", o_ab);
   // We can also add unused variables to the scope which could be
   // useful while debugging
   occaScopeAdd(scope, "debugValue", occaInt(42));
@@ -66,7 +71,8 @@ int main(int argc, const char **argv) {
     )
   );
 
-  occaFinish();
+  // Copy result to the host
+  occaCopyMemToPtr(ab, o_ab, occaAllBytes, 0, occaDefault);
 
   for (i = 0; i < entries; ++i) {
     printf("%d = %f\n", i, ab[i]);
@@ -77,12 +83,18 @@ int main(int argc, const char **argv) {
     }
   }
 
+  // Free host memory
+  free(a);
+  free(b);
+  free(ab);
+
+  // Free device memory and occa objects
   occaFree(&args);
   occaFree(&props);
   occaFree(&scope);
-  occaFreeUvaPtr(a);
-  occaFreeUvaPtr(b);
-  occaFreeUvaPtr(ab);
+  occaFree(&o_a);
+  occaFree(&o_b);
+  occaFree(&o_ab);
 
   return 0;
 }
@@ -98,7 +110,7 @@ occaJson parseArgs(int argc, const char **argv) {
     "      shortname: 'd',"
     "      description: 'Device properties (default: \"{ mode: \\'Serial\\' }\")',"
     "      with_arg: true,"
-    "      default_value: { mode: 'Serial' },"
+    "      default_value: \"{ mode: 'Serial' }\","
     "    },"
     "    {"
     "      name: 'verbose',"

diff --git a/examples/c/04_reduction/main.c b/examples/c/04_reduction/main.c
@@ -114,7 +114,7 @@ occaJson parseArgs(int argc, const char **argv) {
     "      shortname: 'd',"
     "      description: 'Device properties (default: \"{ mode: \\'Serial\\' }\")',"
     "      with_arg: true,"
-    "      default_value: { mode: 'Serial' },"
+    "      default_value: \"{ mode: 'Serial' }\","
     "    },"
     "    {"
     "      name: 'verbose',"

diff --git a/examples/cpp/04_generic_inline_kernel/main.cpp b/examples/cpp/04_generic_inline_kernel/main.cpp
@@ -19,22 +19,31 @@ int main(int argc, const char **argv) {
 
   int entries = 5;
 
-  float *a  = occa::umalloc<float>(entries);
-  float *b  = occa::umalloc<float>(entries);
-  float *ab = occa::umalloc<float>(entries);
+  float *a  = new float[entries];
+  float *b  = new float[entries];
+  float *ab = new float[entries];
 
   for (int i = 0; i < entries; ++i) {
     a[i]  = i;
     b[i]  = 1 - i;
     ab[i] = 0;
   }
 
+  // Uses the background device
+  occa::array<float> array_a(entries);
+  occa::array<float> array_b(entries);
+  occa::array<float> array_ab(entries);
+
+  // Copy over host data
+  array_a.copyFrom(a);
+  array_b.copyFrom(b);
+  array_ab.fill(0);
+
   occa::scope scope({
     {"entries", entries},
-    // Const-ness of variables is passed through, which can be useful for the compiler
-    {"a", (const float*) a},
-    {"b", (const float*) b},
-    {"ab", ab}
+    {"a", array_a},
+    {"b", array_b},
+    {"ab", array_ab}
   }, {
     // Define TILE_SIZE at compile-time
     {"defines/TILE_SIZE", 16}
@@ -46,7 +55,7 @@ int main(int argc, const char **argv) {
     }
   ));
 
-  occa::finish();
+  array_ab.copyTo(ab);
 
   for (int i = 0; i < entries; ++i) {
     std::cout << i << ": " << ab[i] << '\n';
@@ -57,9 +66,10 @@ int main(int argc, const char **argv) {
     }
   }
 
-  occa::free(a);
-  occa::free(b);
-  occa::free(ab);
+  // Free host memory
+  delete [] a;
+  delete [] b;
+  delete [] ab;
 
   return 0;
 }

diff --git a/examples/cpp/07_memory_management/.gitignore b/examples/cpp/07_memory_management/.gitignore
diff --git a/examples/cpp/07_memory_management/CMakeLists.txt b/examples/cpp/07_memory_management/CMakeLists.txt
diff --git a/examples/cpp/07_memory_management/README.md b/examples/cpp/07_memory_management/README.md