refine program

4 files changed, 23 insertions, 18 deletions

diff --git a/backend/src/driver/cl_gen_command_queue.cpp b/backend/src/driver/cl_gen_command_queue.cpp
index 57fee252..5ee7730e 100644
--- a/backend/src/driver/cl_gen_command_queue.cpp
+++ b/backend/src/driver/cl_gen_command_queue.cpp
@@ -266,7 +266,7 @@ cl_int GenReleaseCommandQueue(cl_command_queue queue)
     return CL_INVALID_VALUE;
   }
 
-  GBE_FREE(gpuQueue);
+  GBE_DELETE(gpuQueue);
   setGenCommandQueuePrivate(queue, NULL);
   return CL_SUCCESS;
 }
diff --git a/backend/src/driver/cl_gen_device_id.cpp b/backend/src/driver/cl_gen_device_id.cpp
index 005be34b..b665d6b9 100644
--- a/backend/src/driver/cl_gen_device_id.cpp
+++ b/backend/src/driver/cl_gen_device_id.cpp
@@ -818,7 +818,7 @@ cl_int GenDriverInit(cl_platform_id platform)
   GenGPUDevice* gpuDev = GBE_NEW(GenGPUDevice);
   if (gpuDev->gen_ver < 7) {
     gen_device = NULL;
-    GBE_FREE(gpuDev);
+    GBE_DELETE(gpuDev);
     err = CL_DEVICE_NOT_FOUND;
     return err;
   }
@@ -826,7 +826,7 @@ cl_int GenDriverInit(cl_platform_id platform)
   initGenDevice(gpuDev);
   if (!gen_device) {
     gen_device = NULL;
-    GBE_FREE(gpuDev);
+    GBE_DELETE(gpuDev);
     err = CL_DEVICE_NOT_FOUND;
     return err;
   }
@@ -838,7 +838,7 @@ cl_int GenDriverInit(cl_platform_id platform)
   setGenDevicePrivate(gen_device, gpuDev);
   /* Check and set the extension for device. */
   if (!checkDeviceExtension(platform)) {
-    GBE_FREE(gpuDev);
+    GBE_DELETE(gpuDev);
     setGenDevicePrivate(gen_device, NULL);
     gen_device = NULL;
     err = CL_DEVICE_NOT_FOUND;
diff --git a/backend/src/driver/cl_gen_kernel.cpp b/backend/src/driver/cl_gen_kernel.cpp
index 5f828b29..cfde43e5 100644
--- a/backend/src/driver/cl_gen_kernel.cpp
+++ b/backend/src/driver/cl_gen_kernel.cpp
@@ -759,14 +759,14 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
 
   err = genGPUBindSurfaces(*ndRange->gpuState, queue, kernel, ker);
   if (err != CL_SUCCESS) {
-    GBE_FREE(ndRange);
+    GBE_DELETE(ndRange);
     return err;
   }
 
   if (ker->getImageSize()) {
     err = genGPUBindImage(curbe, *ndRange->gpuState, kernel, ker);
     if (err != CL_SUCCESS) {
-      GBE_FREE(ndRange);
+      GBE_DELETE(ndRange);
       return err;
     }
   }
@@ -774,13 +774,13 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   if (ker->getSamplerSize()) {
     err = genGPUBindSamplers(*ndRange->gpuState, kernel, ker);
     if (err != CL_SUCCESS) {
-      GBE_FREE(ndRange);
+      GBE_DELETE(ndRange);
       return err;
     }
   }
 
   if (ndRange->gpuState->setScratch(scratch_sz) != 0) {
-    GBE_FREE(ndRange);
+    GBE_DELETE(ndRange);
     return CL_OUT_OF_RESOURCES;
   }
 
@@ -788,7 +788,7 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   genGPUSetStack(*ndRange->gpuState, ker, queue->device, gpuDev);
 
   if (genUploadConstantBuffer(queue, curbe, *ndRange->gpuState, kernel, ker, prog) != 0) {
-    GBE_FREE(ndRange);
+    GBE_DELETE(ndRange);
     return CL_OUT_OF_RESOURCES;
   }
 
@@ -801,7 +801,7 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
     GBE_ASSERT(cst_sz > 0);
     char *final_curbe = (char*)alloca(thread_n * cst_sz);
     if (final_curbe == NULL) {
-      GBE_FREE(ndRange);
+      GBE_DELETE(ndRange);
       return CL_OUT_OF_HOST_MEMORY;
     }
 
@@ -810,12 +810,12 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
     }
 
     if (genSetVaryingPayload(ker, final_curbe, local_wk_sz, simd_sz, cst_sz, thread_n) != true) {
-      GBE_FREE(ndRange);
+      GBE_DELETE(ndRange);
       return CL_OUT_OF_HOST_MEMORY;
     }
 
     if (ndRange->gpuState->uploadCurbes(final_curbe, thread_n*cst_sz, thread_n, cst_sz) != true) {
-      GBE_FREE(ndRange);
+      GBE_DELETE(ndRange);
       return CL_OUT_OF_RESOURCES;
     }
   }
@@ -831,7 +831,7 @@ cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
 
   /* Final,enqueue it in the queue worker thread. */
   if (gpuQueue->enqueueWorkItem(ndRange) == false) {
-    GBE_FREE(ndRange);
+    GBE_DELETE(ndRange);
     return CL_INVALID_COMMAND_QUEUE;
   }
 
diff --git a/backend/src/driver/cl_gen_program.cpp b/backend/src/driver/cl_gen_program.cpp
index 1da69cd3..49de5754 100644
--- a/backend/src/driver/cl_gen_program.cpp
+++ b/backend/src/driver/cl_gen_program.cpp
@@ -35,14 +35,20 @@ using namespace gbe;
 extern "C"
 cl_int GenBuildProgram(cl_program program, const cl_device_id device)
 {
-  GBE_ASSERT(getGenProgramPrivate(program, device) == NULL);
+  if (getGenProgramPrivate(program, device) != NULL) { // Rebuild ?
+    Program* prog = reinterpret_cast<Program*>(getGenProgramPrivate(program, device));
+    GBE_DELETE(prog);
+    setGenProgramPrivate(program, device, NULL);
+  }
 
   int index = findIndexByDevice(program->ctx, device);
   int device_id = (reinterpret_cast<GenGPUDevice*>(getGenDevicePrivate(device)))->device_id;
   GBE_ASSERT(index >= 0);
 
-  Program* p = ProgramNewFromSource(device_id, program->source, CL_BUILD_LOG_MAX_SZ, program->build_opts,
-                                    program->build_log[index], &program->build_log_sz[index]);
+  Program* p = ProgramNewFromSource(device_id, program->source, CL_BUILD_LOG_MAX_SZ,
+                                    program->device_status[index].build_opts,
+                                    program->device_status[index].build_log,
+                                    &program->device_status[index].build_log_sz);
   if (p == NULL) {
     return CL_BUILD_PROGRAM_FAILURE;
   }
@@ -94,8 +100,7 @@ cl_int GenReleaseProgram(cl_program program, const cl_device_id device)
   if (p == NULL)
     return CL_INVALID_VALUE;
 
-  GBE_FREE(p);
-
+  GBE_DELETE(p);
   setGenProgramPrivate(program, device, NULL);
   return CL_SUCCESS;
 }