ndrange

3 files changed, 63 insertions, 21 deletions

diff --git a/backend/src/driver/cl_gen_gpu_state.cpp b/backend/src/driver/cl_gen_gpu_state.cpp
index addaa930..9b8f1124 100644
--- a/backend/src/driver/cl_gen_gpu_state.cpp
+++ b/backend/src/driver/cl_gen_gpu_state.cpp
@@ -171,6 +171,23 @@ void GenGPUState::setStack(uint32_t offset, uint32_t size, uint8_t bti)
   this->bindBuf(this->stack_b.bo, offset, 0, size, bti);
 }
 
+int GenGPUState::setScratch(uint32_t per_thread_size)
+{
+  uint32_t total = per_thread_size * this->max_threads;
+  /* Per Bspec, scratch should 2X the desired size, otherwise luxmark may hang */
+  if (IS_HASWELL(device_id))
+      total *= 2;
+
+  this->per_thread_scratch = per_thread_size;
+
+  this->scratch_b.bo = drm_intel_bo_alloc(bufmgr, "SCRATCH_BO", total, 4096);
+  if (this->scratch_b.bo == NULL)
+    return -1;
+
+  return 0;
+}
+
+
 bool GenGPUState::stateInit(uint32_t max_threads, uint32_t size_cs_entry)
 {
   drm_intel_bo *bo = NULL;
@@ -180,6 +197,7 @@ bool GenGPUState::stateInit(uint32_t max_threads, uint32_t size_cs_entry)
   this->img_bitmap = 0;
   this->img_index_base = 3;
   this->sampler_bitmap = ~((1 << max_sampler_n) - 1);
+  this->per_thread_scratch = 0;
 
   /* URB */
   this->curb.num_cs_entries = 64;
diff --git a/backend/src/driver/cl_gen_gpu_state.hpp b/backend/src/driver/cl_gen_gpu_state.hpp
index eafac148..54849f94 100644
--- a/backend/src/driver/cl_gen_gpu_state.hpp
+++ b/backend/src/driver/cl_gen_gpu_state.hpp
@@ -169,6 +169,7 @@ struct GenGPUState {
   void bindBuf(drm_intel_bo *buf, uint32_t offset, uint32_t internal_offset,
                size_t size, uint8_t bti);
   void setStack(uint32_t offset, uint32_t size, uint8_t bti);
+  int setScratch(uint32_t per_thread_size);
   bool stateInit(uint32_t max_threads, uint32_t size_cs_entry);
   bool allocConstantBuffer(uint32_t size, uint8_t bti);
   void batchStart(uint32_t use_slm);
diff --git a/backend/src/driver/cl_gen_kernel.cpp b/backend/src/driver/cl_gen_kernel.cpp
index 72a933fa..6bd6748f 100644
--- a/backend/src/driver/cl_gen_kernel.cpp
+++ b/backend/src/driver/cl_gen_kernel.cpp
@@ -391,7 +391,7 @@ static cl_int genGPUBindSurfaces(GenGPUState& gpuState, cl_command_queue queue,
   return CL_SUCCESS;
 }
 
-static cl_int genQueueBindImage(GenGPUState& gpuState, cl_kernel kernel, Kernel* ker)
+static cl_int genGPUBindImage(char* curbe, GenGPUState& gpuState, cl_kernel kernel, Kernel* ker)
 {
   uint32_t i;
   size_t image_sz = ker->getImageSize();
@@ -406,32 +406,40 @@ static cl_int genQueueBindImage(GenGPUState& gpuState, cl_kernel kernel, Kernel*
 
     image = cl_mem_to_image(kernel->args[id]->mem);
 
-
-
-#if 0
-    set_image_info(k->curbe, &k->images[i], image);
-    if(k->vme) {
-      if( (image->fmt.image_channel_order != CL_R) || (image->fmt.image_channel_data_type != CL_UNORM_INT8) )
-        return CL_IMAGE_FORMAT_NOT_SUPPORTED;
-      cl_gpgpu_bind_image_for_vme(gpgpu, k->images[i].idx, image->base.bo, image->offset + k->args[id].mem->offset,
-                                  image->intel_fmt, image->image_type, image->bpp,
-                                  image->w, image->h, image->depth,
-                                  image->row_pitch, image->slice_pitch, (cl_gpgpu_tiling)image->tiling);
-    } else
-      cl_gpgpu_bind_image(gpgpu, k->images[i].idx, image->base.bo, image->offset + k->args[id].mem->offset,
-                          image->intel_fmt, image->image_type, image->bpp,
-                          image->w, image->h, image->depth,
-                          image->row_pitch, image->slice_pitch, (cl_gpgpu_tiling)image->tiling);
+    if (images[i].wSlot >= 0)
+      *(uint32_t*)(curbe + images[i].wSlot) = image->w;
+    if (images[i].hSlot >= 0)
+      *(uint32_t*)(curbe + images[i].hSlot) = image->h;
+    if (images[i].depthSlot >= 0)
+      *(uint32_t*)(curbe + images[i].depthSlot) = image->depth;
+    if (images[i].channelOrderSlot >= 0)
+      *(uint32_t*)(curbe + images[i].channelOrderSlot) = image->fmt.image_channel_order;
+    if (images[i].dataTypeSlot >= 0)
+      *(uint32_t*)(curbe + images[i].dataTypeSlot) = image->fmt.image_channel_data_type;
+    /*
+        gpuState.bindImage(images[i].idx, image->base.bo,
+                            image->offset + k->args[id].mem->offset, image->intel_fmt,
+                            image->image_type, image->bpp, image->w, image->h, image->depth,
+                            image->row_pitch, image->slice_pitch, (cl_gpgpu_tiling)image->tiling);
     // TODO, this workaround is for GEN7/GEN75 only, we may need to do it in the driver layer
     // on demand.
     if (image->image_type == CL_MEM_OBJECT_IMAGE1D_ARRAY)
       cl_gpgpu_bind_image(gpgpu, k->images[i].idx + BTI_WORKAROUND_IMAGE_OFFSET, image->base.bo, image->offset + k->args[id].mem->offset,
-                          image->intel_fmt, image->image_type, image->bpp,
-                          image->w, image->h, image->depth,
-                          image->row_pitch, image->slice_pitch, (cl_gpgpu_tiling)image->tiling);
+    					  image->intel_fmt, image->image_type, image->bpp,
+    					  image->w, image->h, image->depth,
+    					  image->row_pitch, image->slice_pitch, (cl_gpgpu_tiling)image->tiling);
 
-#endif
+    */
+    return CL_SUCCESS;
   }
+}
+
+static cl_int genGPUBindSamplers(GenGPUState& gpuState, cl_kernel kernel, Kernel* ker)
+{
+  size_t sz = ker->getSamplerSize();
+  uint32_t *samplers = (uint32_t *)alloca(GEN_MAX_SAMPLERS * sizeof(uint32_t));
+  ker->getSamplerData(samplers);
+  gpuState.bindSamplers(samplers, sz);
   return CL_SUCCESS;
 }
 
@@ -505,5 +513,20 @@ cl_int genEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const u
   if (err != CL_SUCCESS)
     return err;
 
+  if (ker->getImageSize()) {
+    err = genGPUBindImage(curbe, *ndRange->gpuState, kernel, ker);
+    if (err != CL_SUCCESS)
+      return err;
+  }
+
+  if (ker->getSamplerSize()) {
+    err = genGPUBindSamplers(*ndRange->gpuState, kernel, ker);
+    if (err != CL_SUCCESS)
+      return err;
+  }
+  err = ndRange->gpuState->setScratch(scratch_sz);
+  if (err != CL_SUCCESS)
+    return err;
+
 }