ndrange

2 files changed, 54 insertions, 8 deletions

diff --git a/backend/src/driver/cl_gen_kernel.cpp b/backend/src/driver/cl_gen_kernel.cpp
index 8e2baf3d..3e1877c2 100644
--- a/backend/src/driver/cl_gen_kernel.cpp
+++ b/backend/src/driver/cl_gen_kernel.cpp
@@ -351,6 +351,45 @@ static int32_t genFillCurbe(cl_kernel kernel, Kernel* ker, char* curbe, const ui
   return slm_offset;
 }
 
+cl_int genGPUBindSurfaces(GenGPUState& gpuState, cl_command_queue queue, cl_kernel kernel, Kernel* ker)
+{
+  /* Bind all user buffers (given by clSetKernelArg) */
+  uint32_t i;
+  enum gbe_arg_type arg_type; /* kind of argument */
+  for (i = 0; i < ker->getArgNum(); ++i) {
+    int32_t offset; // location of the address in the curbe
+    arg_type = ker->getArgType(i);
+    cl_mem mem = kernel->args[i]->mem;
+    if (arg_type != GBE_ARG_GLOBAL_PTR || !mem)
+      continue;
+
+    offset = ker->getCurbeOffset(GBE_CURBE_KERNEL_ARGUMENT, i);
+    if (offset < 0)
+      continue;
+
+    cl_mem_buffer buffer = cl_mem_to_buffer(mem);
+    uint8_t bti = ker->getArgBTI(i);
+    GenGPUMem* genMem = reinterpret_cast<GenGPUMem*>(getGenMemPrivate(mem, queue->device));
+    if (genMem == NULL)
+      return CL_INVALID_VALUE;
+
+    if (genMem->bo == NULL) { // We have not allocated buffer, first time use this mem.
+      if (genMem->genAllocMemBo(mem) == false) {
+        return CL_MEM_OBJECT_ALLOCATION_FAILURE;
+      }
+    }
+
+    if (genMem->alignedHostPtr == NULL) {
+      gpuState.bindBuf(genMem->bo, offset, buffer->sub_offset, genMem->realSize, bti);
+    } else {
+      gpuState.bindBuf(genMem->bo, offset, (char*)mem->host_ptr - (char*)genMem->alignedHostPtr
+                       + buffer->sub_offset, genMem->realSize, bti);
+    }
+  }
+
+  return CL_SUCCESS;
+}
+
 extern "C"
 cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const uint32_t work_dim,
                                const size_t *global_wk_off, const size_t *global_wk_sz,
@@ -374,10 +413,8 @@ cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   size_t cst_sz = ker->getCurbeSize();
   int32_t scratch_sz = ker->getScratchSize();
   size_t thread_n = 0u;
-  int printf_num = 0;
   cl_int err = CL_SUCCESS;
   size_t global_size = global_wk_sz[0] * global_wk_sz[1] * global_wk_sz[2];
-  void* printf_info = NULL;
   bool use_slm = ker->getUseSLM();
   size_t local_sz = 0u;
 
@@ -395,7 +432,7 @@ cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
 
   thread_n = (local_sz + simd_sz - 1) / simd_sz;
 
-  if (scratch_sz > gpuDev->scratch_mem_size) {
+  if ((uint32_t)scratch_sz > gpuDev->scratch_mem_size) {
     printf("Out of scratch memory %d.", scratch_sz);
     return CL_OUT_OF_RESOURCES;
   }
@@ -406,7 +443,7 @@ cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   if (curbe) {
     int32_t slm_sz = genFillCurbe(kernel, ker, curbe, work_dim,
                                   global_wk_off, global_wk_sz, local_wk_sz, thread_n);
-    if (slm_sz > queue->device->local_mem_size) {
+    if ((uint32_t)slm_sz > queue->device->local_mem_size) {
       printf("Out of shared local memory %d.\n", slm_sz);
       return CL_OUT_OF_RESOURCES;
     }
@@ -417,8 +454,11 @@ cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   err = ndRange->gpuState->stateInit(
           queue->device->max_compute_unit * gpuDev->max_thread_per_unit, cst_sz / 32);
 
-//  if (queue->props & CL_QUEUE_PROFILING_ENABLE)
-//      err = cl_gpgpu_state_init(gpgpu, ctx->device->max_compute_unit * ctx->device->max_thread_per_unit, cst_sz / 32, 0);
+  /* We may allocate buffers, begin to lock.*/
+  GenGPUCommandQueueLocker locker(gpuQueue);
+  err = genGPUBindSurfaces(*ndRange->gpuState, queue, kernel, ker);
+  if (err != CL_SUCCESS)
+    return err;
 
 }
 
diff --git a/backend/src/driver/cl_gen_mem.cpp b/backend/src/driver/cl_gen_mem.cpp
index a2d57b60..33e1d5ca 100644
--- a/backend/src/driver/cl_gen_mem.cpp
+++ b/backend/src/driver/cl_gen_mem.cpp
@@ -146,8 +146,14 @@ bool GenGPUMem::genAllocMemBo(cl_mem mem)
   }
 
   /* Copy the data if required */
-  if (mem->flags & CL_MEM_COPY_HOST_PTR) {
-    drm_intel_bo_subdata(this->bo, 0, mem->size, mem->host_ptr);
+  if ((mem->flags & CL_MEM_COPY_HOST_PTR) ||
+    (mem->flags & CL_MEM_USE_HOST_PTR && this->alignedHostPtr == NULL)) {
+    if (IS_IMAGE(mem)) {
+      drm_intel_bo_subdata(this->bo, 0, mem->size, mem->host_ptr);
+    } else {
+      cl_mem_buffer buffer = cl_mem_to_buffer(mem);
+      drm_intel_bo_subdata(this->bo, buffer->sub_offset, mem->size, mem->host_ptr);
+    }
   }
 
   return true;