Cycles: OpenCL split kernel refactor

This does a few things at once:

- Refactors host side split kernel logic into a new device
  agnostic class `DeviceSplitKernel`.
- Removes tile splitting, a new work pool implementation takes its place and
  allows as many threads as will fit in memory regardless of tile size, which
  can give performance gains.
- Refactors split state buffers into one buffer, as well as reduces the
  number of arguments passed to kernels. Means there's less code to deal
  with overall.
- Moves kernel logic out of OpenCL kernel files so they can later be used by
  other device types.
- Replaced OpenCL specific APIs with new generic versions
- Tiles can now be seen updating during rendering

1 files changed, 58 insertions, 20 deletions

diff --git a/intern/cycles/kernel/split/kernel_direct_lighting.h b/intern/cycles/kernel/split/kernel_direct_lighting.h
index 82ca18829d3..43b943710a1 100644
--- a/intern/cycles/kernel/split/kernel_direct_lighting.h
+++ b/intern/cycles/kernel/split/kernel_direct_lighting.h
@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "kernel_split_common.h"
+CCL_NAMESPACE_BEGIN
 
 /* Note on kernel_direct_lighting kernel.
  * This is the eighth kernel in the ray tracing logic. This is the seventh
@@ -47,20 +47,42 @@
  * QUEUE_SHADOW_RAY_CAST_DL_RAYS queue will be filled with rays for which a shadow_blocked function must be executed, after this
  * kernel call. Before this kernel call the QUEUE_SHADOW_RAY_CAST_DL_RAYS will be empty.
  */
-ccl_device char kernel_direct_lighting(
-        KernelGlobals *kg,
-        ShaderData *sd,                         /* Required for direct lighting */
-        ccl_global uint *rng_coop,              /* Required for direct lighting */
-        ccl_global PathState *PathState_coop,   /* Required for direct lighting */
-        ccl_global int *ISLamp_coop,            /* Required for direct lighting */
-        ccl_global Ray *LightRay_coop,          /* Required for direct lighting */
-        ccl_global BsdfEval *BSDFEval_coop,     /* Required for direct lighting */
-        ccl_global char *ray_state,             /* Denotes the state of each ray */
-        int ray_index)
+ccl_device void kernel_direct_lighting(KernelGlobals *kg)
 {
+	ccl_local unsigned int local_queue_atomics;
+	if(ccl_local_id(0) == 0 && ccl_local_id(1) == 0) {
+		local_queue_atomics = 0;
+	}
+	ccl_barrier(CCL_LOCAL_MEM_FENCE);
+
 	char enqueue_flag = 0;
-	if(IS_STATE(ray_state, ray_index, RAY_ACTIVE)) {
-		ccl_global PathState *state = &PathState_coop[ray_index];
+	int ray_index = ccl_global_id(1) * ccl_global_size(0) + ccl_global_id(0);
+	ray_index = get_ray_index(kg, ray_index,
+	                          QUEUE_ACTIVE_AND_REGENERATED_RAYS,
+	                          kernel_split_state.queue_data,
+	                          kernel_split_params.queue_size,
+	                          0);
+
+#ifdef __COMPUTE_DEVICE_GPU__
+	/* If we are executing on a GPU device, we exit all threads that are not
+	 * required.
+	 *
+	 * If we are executing on a CPU device, then we need to keep all threads
+	 * active since we have barrier() calls later in the kernel. CPU devices,
+	 * expect all threads to execute barrier statement.
+	 */
+	if(ray_index == QUEUE_EMPTY_SLOT) {
+		return;
+	}
+#endif
+
+#ifndef __COMPUTE_DEVICE_GPU__
+	if(ray_index != QUEUE_EMPTY_SLOT) {
+#endif
+
+	if(IS_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE)) {
+		ccl_global PathState *state = &kernel_split_state.path_state[ray_index];
+		ShaderData *sd = kernel_split_state.sd;
 
 		/* direct lighting */
 #ifdef __EMISSION__
@@ -68,7 +90,7 @@ ccl_device char kernel_direct_lighting(
 		    (ccl_fetch(sd, flag) & SD_BSDF_HAS_EVAL)))
 		{
 			/* Sample illumination from lights to find path contribution. */
-			ccl_global RNG* rng = &rng_coop[ray_index];
+			ccl_global RNG* rng = &kernel_split_state.rng[ray_index];
 			float light_t = path_state_rng_1D(kg, rng, state, PRNG_LIGHT);
 			float light_u, light_v;
 			path_state_rng_2D(kg, rng, state, PRNG_LIGHT_U, &light_u, &light_v);
@@ -89,20 +111,36 @@ ccl_device char kernel_direct_lighting(
 
 				BsdfEval L_light;
 				bool is_lamp;
-				if(direct_emission(kg, sd, kg->sd_input, &ls, state, &light_ray, &L_light, &is_lamp, terminate)) {
+				if(direct_emission(kg, sd, kernel_split_state.sd_DL_shadow, &ls, state, &light_ray, &L_light, &is_lamp, terminate)) {
 					/* Write intermediate data to global memory to access from
 					 * the next kernel.
 					 */
-					LightRay_coop[ray_index] = light_ray;
-					BSDFEval_coop[ray_index] = L_light;
-					ISLamp_coop[ray_index] = is_lamp;
+					kernel_split_state.light_ray[ray_index] = light_ray;
+					kernel_split_state.bsdf_eval[ray_index] = L_light;
+					kernel_split_state.is_lamp[ray_index] = is_lamp;
 					/* Mark ray state for next shadow kernel. */
-					ADD_RAY_FLAG(ray_state, ray_index, RAY_SHADOW_RAY_CAST_DL);
+					ADD_RAY_FLAG(kernel_split_state.ray_state, ray_index, RAY_SHADOW_RAY_CAST_DL);
 					enqueue_flag = 1;
 				}
 			}
 		}
 #endif  /* __EMISSION__ */
 	}
-	return enqueue_flag;
+
+#ifndef __COMPUTE_DEVICE_GPU__
+	}
+#endif
+
+#ifdef __EMISSION__
+	/* Enqueue RAY_SHADOW_RAY_CAST_DL rays. */
+	enqueue_ray_index_local(ray_index,
+	                        QUEUE_SHADOW_RAY_CAST_DL_RAYS,
+	                        enqueue_flag,
+	                        kernel_split_params.queue_size,
+	                        &local_queue_atomics,
+	                        kernel_split_state.queue_data,
+	                        kernel_split_params.queue_index);
+#endif
 }
+
+CCL_NAMESPACE_END