Cycles: Improve denoising speed on GPUs with small tile sizes

Previously, the NLM kernels would be launched once per offset with one thread per pixel.
However, with the smaller tile sizes that are now feasible, there wasn't enough work to fully occupy GPUs which results in a significant slowdown.

Therefore, the kernels are now launched in a single call that handles all offsets at once.
This has two downsides: Memory accesses to accumulating buffers are now atomic, and more importantly, the temporary memory now has to be allocated for every shift at once, increasing the required memory.
On the other hand, of course, the smaller tiles significantly reduce the size of the memory.

The main bottleneck right now is the construction of the transformation - there is nothing to be parallelized there, one thread per pixel is the maximum.
I tried to parallelize the SVD implementation by storing the matrix in shared memory and launching one block per pixel, but that wasn't really going anywhere.

To make the new code somewhat readable, the handling of rectangular regions was cleaned up a bit and commented, it should be easier to understand what's going on now.
Also, some variables have been renamed to make the difference between buffer width and stride more apparent, in addition to some general style cleanup.

2 files changed, 120 insertions, 109 deletions

diff --git a/intern/cycles/device/opencl/opencl.h b/intern/cycles/device/opencl/opencl.h
index c02f8ffafe6..f38c2f65c1e 100644
--- a/intern/cycles/device/opencl/opencl.h
+++ b/intern/cycles/device/opencl/opencl.h
@@ -353,7 +353,9 @@ public:
 	void tex_free(device_memory& mem);
 
 	size_t global_size_round_up(int group_size, int global_size);
-	void enqueue_kernel(cl_kernel kernel, size_t w, size_t h, size_t max_workgroup_size = -1);
+	void enqueue_kernel(cl_kernel kernel, size_t w, size_t h,
+	                    bool x_workgroups = false,
+	                    size_t max_workgroup_size = -1);
 	void set_kernel_arg_mem(cl_kernel kernel, cl_uint *narg, const char *name);
 	void set_kernel_arg_buffers(cl_kernel kernel, cl_uint *narg);
 
diff --git a/intern/cycles/device/opencl/opencl_base.cpp b/intern/cycles/device/opencl/opencl_base.cpp
index f43177247ef..fe084edc90e 100644
--- a/intern/cycles/device/opencl/opencl_base.cpp
+++ b/intern/cycles/device/opencl/opencl_base.cpp
@@ -560,7 +560,7 @@ size_t OpenCLDeviceBase::global_size_round_up(int group_size, int global_size)
 	return global_size + ((r == 0)? 0: group_size - r);
 }
 
-void OpenCLDeviceBase::enqueue_kernel(cl_kernel kernel, size_t w, size_t h, size_t max_workgroup_size)
+void OpenCLDeviceBase::enqueue_kernel(cl_kernel kernel, size_t w, size_t h, bool x_workgroups, size_t max_workgroup_size)
 {
 	size_t workgroup_size, max_work_items[3];
 
@@ -574,8 +574,15 @@ void OpenCLDeviceBase::enqueue_kernel(cl_kernel kernel, size_t w, size_t h, size
 	}
 
 	/* Try to divide evenly over 2 dimensions. */
-	size_t sqrt_workgroup_size = max((size_t)sqrt((double)workgroup_size), 1);
-	size_t local_size[2] = {sqrt_workgroup_size, sqrt_workgroup_size};
+	size_t local_size[2];
+	if(x_workgroups) {
+		local_size[0] = workgroup_size;
+		local_size[1] = 1;
+	}
+	else {
+		size_t sqrt_workgroup_size = max((size_t)sqrt((double)workgroup_size), 1);
+		local_size[0] = local_size[1] = sqrt_workgroup_size;
+	}
 
 	/* Some implementations have max size 1 on 2nd dimension. */
 	if(local_size[1] > max_work_items[1]) {
@@ -731,17 +738,25 @@ bool OpenCLDeviceBase::denoising_non_local_means(device_ptr image_ptr,
                                                  device_ptr out_ptr,
                                                  DenoisingTask *task)
 {
-	int4 rect = task->rect;
-	int w = rect.z-rect.x;
-	int h = rect.w-rect.y;
+
+	int stride = task->buffer.stride;
+	int w = task->buffer.width;
+	int h = task->buffer.h;
 	int r = task->nlm_state.r;
 	int f = task->nlm_state.f;
 	float a = task->nlm_state.a;
 	float k_2 = task->nlm_state.k_2;
 
-	cl_mem difference     = CL_MEM_PTR(task->nlm_state.temporary_1_ptr);
-	cl_mem blurDifference = CL_MEM_PTR(task->nlm_state.temporary_2_ptr);
-	cl_mem weightAccum    = CL_MEM_PTR(task->nlm_state.temporary_3_ptr);
+	int shift_stride = stride*h;
+	int num_shifts = (2*r+1)*(2*r+1);
+	int mem_size = sizeof(float)*shift_stride*num_shifts;
+
+	cl_mem weightAccum = CL_MEM_PTR(task->nlm_state.temporary_3_ptr);
+
+	cl_mem difference = clCreateBuffer(cxContext, CL_MEM_READ_WRITE, mem_size, NULL, &ciErr);
+	opencl_assert_err(ciErr, "clCreateBuffer denoising_non_local_means");
+	cl_mem blurDifference = clCreateBuffer(cxContext, CL_MEM_READ_WRITE, mem_size, NULL, &ciErr);
+	opencl_assert_err(ciErr, "clCreateBuffer denoising_non_local_means");
 
 	cl_mem image_mem = CL_MEM_PTR(image_ptr);
 	cl_mem guide_mem = CL_MEM_PTR(guide_ptr);
@@ -757,31 +772,45 @@ bool OpenCLDeviceBase::denoising_non_local_means(device_ptr image_ptr,
 	cl_kernel ckNLMUpdateOutput   = denoising_program(ustring("filter_nlm_update_output"));
 	cl_kernel ckNLMNormalize      = denoising_program(ustring("filter_nlm_normalize"));
 
-	for(int i = 0; i < (2*r+1)*(2*r+1); i++) {
-		int dy = i / (2*r+1) - r;
-		int dx = i % (2*r+1) - r;
-		int4 local_rect = make_int4(max(0, -dx), max(0, -dy), rect.z-rect.x - max(0, dx), rect.w-rect.y - max(0, dy));
-		kernel_set_args(ckNLMCalcDifference, 0,
-		                dx, dy, guide_mem, variance_mem,
-		                difference, local_rect, w, 0, a, k_2);
-		kernel_set_args(ckNLMBlur, 0,
-		                difference, blurDifference, local_rect, w, f);
-		kernel_set_args(ckNLMCalcWeight, 0,
-		                blurDifference, difference, local_rect, w, f);
-		kernel_set_args(ckNLMUpdateOutput, 0,
-		                dx, dy, blurDifference, image_mem,
-		                out_mem, weightAccum, local_rect, w, f);
-
-		enqueue_kernel(ckNLMCalcDifference, w, h);
-		enqueue_kernel(ckNLMBlur,           w, h);
-		enqueue_kernel(ckNLMCalcWeight,     w, h);
-		enqueue_kernel(ckNLMBlur,           w, h);
-		enqueue_kernel(ckNLMUpdateOutput,   w, h);
-	}
+	kernel_set_args(ckNLMCalcDifference, 0,
+	                guide_mem,
+	                variance_mem,
+	                difference,
+	                w, h, stride,
+	                shift_stride,
+	                r, 0, a, k_2);
+	kernel_set_args(ckNLMBlur, 0,
+	                difference,
+	                blurDifference,
+	                w, h, stride,
+	                shift_stride,
+	                r, f);
+	kernel_set_args(ckNLMCalcWeight, 0,
+	                blurDifference,
+	                difference,
+	                w, h, stride,
+	                shift_stride,
+	                r, f);
+	kernel_set_args(ckNLMUpdateOutput, 0,
+	                blurDifference,
+	                image_mem,
+	                out_mem,
+	                weightAccum,
+	                w, h, stride,
+	                shift_stride,
+	                r, f);
+
+	enqueue_kernel(ckNLMCalcDifference, w*h, num_shifts, true);
+	enqueue_kernel(ckNLMBlur,           w*h, num_shifts, true);
+	enqueue_kernel(ckNLMCalcWeight,     w*h, num_shifts, true);
+	enqueue_kernel(ckNLMBlur,           w*h, num_shifts, true);
+	enqueue_kernel(ckNLMUpdateOutput,   w*h, num_shifts, true);
+
+	opencl_assert(clReleaseMemObject(difference));
+	opencl_assert(clReleaseMemObject(blurDifference));
 
-	int4 local_rect = make_int4(0, 0, w, h);
 	kernel_set_args(ckNLMNormalize, 0,
-	                out_mem, weightAccum, local_rect, w);
+	                out_mem, weightAccum, w, h, stride);
 	enqueue_kernel(ckNLMNormalize, w, h);
 
 	return true;
@@ -837,81 +866,63 @@ bool OpenCLDeviceBase::denoising_reconstruct(device_ptr color_ptr,
 	cl_kernel ckNLMConstructGramian = denoising_program(ustring("filter_nlm_construct_gramian"));
 	cl_kernel ckFinalize            = denoising_program(ustring("filter_finalize"));
 
-	cl_mem difference     = CL_MEM_PTR(task->reconstruction_state.temporary_1_ptr);
-	cl_mem blurDifference = CL_MEM_PTR(task->reconstruction_state.temporary_2_ptr);
-
-	int r = task->radius;
-	int f = 4;
-	float a = 1.0f;
-	for(int i = 0; i < (2*r+1)*(2*r+1); i++) {
-		int dy = i / (2*r+1) - r;
-		int dx = i % (2*r+1) - r;
-
-		int local_rect[4] = {max(0, -dx), max(0, -dy),
-		                     task->reconstruction_state.source_w - max(0, dx),
-		                     task->reconstruction_state.source_h - max(0, dy)};
-
-		kernel_set_args(ckNLMCalcDifference, 0,
-		                dx, dy,
-		                color_mem,
-		                color_variance_mem,
-		                difference,
-		                local_rect,
-		                task->buffer.w,
-		                task->buffer.pass_stride,
-		                a, task->nlm_k_2);
-		enqueue_kernel(ckNLMCalcDifference,
-		               task->reconstruction_state.source_w,
-		               task->reconstruction_state.source_h);
-
-		kernel_set_args(ckNLMBlur, 0,
-		                difference,
-		                blurDifference,
-		                local_rect,
-		                task->buffer.w,
-		                f);
-		enqueue_kernel(ckNLMBlur,
-		               task->reconstruction_state.source_w,
-		               task->reconstruction_state.source_h);
-
-		kernel_set_args(ckNLMCalcWeight, 0,
-		                blurDifference,
-		                difference,
-		                local_rect,
-		                task->buffer.w,
-		                f);
-		enqueue_kernel(ckNLMCalcWeight,
-		               task->reconstruction_state.source_w,
-		               task->reconstruction_state.source_h);
-
-		/* Reuse previous arguments. */
-		enqueue_kernel(ckNLMBlur,
-		               task->reconstruction_state.source_w,
-		               task->reconstruction_state.source_h);
-
-		kernel_set_args(ckNLMConstructGramian, 0,
-		                dx, dy,
-		                blurDifference,
-		                buffer_mem,
-		                transform_mem,
-		                rank_mem,
-		                XtWX_mem,
-		                XtWY_mem,
-		                local_rect,
-		                task->reconstruction_state.filter_rect,
-		                task->buffer.w,
-		                task->buffer.h,
-		                f,
-	                    task->buffer.pass_stride);
-		enqueue_kernel(ckNLMConstructGramian,
-		               task->reconstruction_state.source_w,
-		               task->reconstruction_state.source_h,
-		               256);
-	}
+	int w = task->reconstruction_state.source_w;
+	int h = task->reconstruction_state.source_h;
+	int stride = task->buffer.stride;
+
+	int shift_stride = stride*h;
+	int num_shifts = (2*task->radius + 1)*(2*task->radius + 1);
+	int mem_size = sizeof(float)*shift_stride*num_shifts;
+
+	cl_mem difference = clCreateBuffer(cxContext, CL_MEM_READ_WRITE, mem_size, NULL, &ciErr);
+	opencl_assert_err(ciErr, "clCreateBuffer denoising_reconstruct");
+	cl_mem blurDifference = clCreateBuffer(cxContext, CL_MEM_READ_WRITE, mem_size, NULL, &ciErr);
+	opencl_assert_err(ciErr, "clCreateBuffer denoising_reconstruct");
+
+	kernel_set_args(ckNLMCalcDifference, 0,
+	                color_mem,
+	                color_variance_mem,
+	                difference,
+	                w, h, stride,
+	                shift_stride,
+	                task->radius,
+	                task->buffer.pass_stride,
+	                1.0f, task->nlm_k_2);
+	kernel_set_args(ckNLMBlur, 0,
+	                difference,
+	                blurDifference,
+	                w, h, stride,
+	                shift_stride,
+	                task->radius, 4);
+	kernel_set_args(ckNLMCalcWeight, 0,
+	                blurDifference,
+	                difference,
+	                w, h, stride,
+	                shift_stride,
+	                task->radius, 4);
+	kernel_set_args(ckNLMConstructGramian, 0,
+	                blurDifference,
+	                buffer_mem,
+	                transform_mem,
+	                rank_mem,
+	                XtWX_mem,
+	                XtWY_mem,
+	                task->reconstruction_state.filter_window,
+	                w, h, stride,
+	                shift_stride,
+	                task->radius, 4,
+	                task->buffer.pass_stride);
+
+	enqueue_kernel(ckNLMCalcDifference,   w*h, num_shifts, true);
+	enqueue_kernel(ckNLMBlur,             w*h, num_shifts, true);
+	enqueue_kernel(ckNLMCalcWeight,       w*h, num_shifts, true);
+	enqueue_kernel(ckNLMBlur,             w*h, num_shifts, true);
+	enqueue_kernel(ckNLMConstructGramian, w*h, num_shifts, true, 256);
+
+	opencl_assert(clReleaseMemObject(difference));
+	opencl_assert(clReleaseMemObject(blurDifference));
 
 	kernel_set_args(ckFinalize, 0,
-	                task->buffer.w,
-	                task->buffer.h,
 	                output_mem,
 	                rank_mem,
 	                XtWX_mem,
@@ -919,9 +930,7 @@ bool OpenCLDeviceBase::denoising_reconstruct(device_ptr color_ptr,
 	                task->filter_area,
 	                task->reconstruction_state.buffer_params,
 	                task->render_buffer.samples);
-	enqueue_kernel(ckFinalize,
-	               task->reconstruction_state.source_w,
-	               task->reconstruction_state.source_h);
+	enqueue_kernel(ckFinalize, w, h);
 
 	return true;
 }