Support not writing top mip for HiZ.

It's very unlikely that it'll matter in practice and it saves a lot of
GPU time to not have to write out full-res.

Author: Themaister

assets/shaders/inc/meshlet_render.h

@@ -57,6 +57,7 @@ layout(set = MESHLET_RENDER_DESCRIPTOR_SET, binding = MESHLET_RENDER_FRUSTUM_BIN
 	mat4 view;
 	vec4 viewport_scale_bias;
 	ivec2 hiz_resolution;
+	int hiz_min_lod;
 	int hiz_max_lod;
 } frustum;
 
@@ -114,13 +115,16 @@ bool hiz_cull(vec2 view_range_x, vec2 view_range_y, float closest_z)
 	// We need to sample from a LOD where where there is at most one texel delta
 	// between lo/hi values.
 	int max_delta = max(ix.y - ix.x, iy.y - iy.x);
-	int lod = min(findMSB(max_delta - 1) + 1, frustum.hiz_max_lod);
+	int lod = clamp(findMSB(max_delta - 1) + 1, frustum.hiz_min_lod, frustum.hiz_max_lod);
 	ivec2 lod_max_coord = max(frustum.hiz_resolution >> lod, ivec2(1)) - 1;
 	ix = min(ix >> lod, lod_max_coord.xx);
 	iy = min(iy >> lod, lod_max_coord.yy);
 
 	ivec2 hiz_coord = ivec2(ix.x, iy.x);
 
+	// We didn't write the top LOD.
+	lod -= frustum.hiz_min_lod;
+
 	float d = texelFetch(uHiZDepth, hiz_coord, lod).x;
 	bool nx = ix.y != ix.x;
 	bool ny = iy.y != iy.x;

assets/shaders/post/hiz.comp

@@ -1,13 +1,14 @@
 #version 450
 #extension GL_KHR_shader_subgroup_basic : require
-#extension GL_KHR_shader_subgroup_shuffle : require
 #extension GL_KHR_shader_subgroup_quad : require
 
 // A rewrite of SPD to support HiZ correctly and moar wave ops for good measure.
 
 layout(local_size_x = 256) in;
 
-layout(set = 0, binding = 0, r32f) uniform writeonly image2D uImageTop;
+#if defined(WRITE_TOP_LEVEL) && WRITE_TOP_LEVEL
+layout(set = 0, binding = 0, r32f) coherent writeonly uniform image2D uImageTop;
+#endif
 layout(set = 0, binding = 1, r32f) coherent uniform image2D uImages[12];
 layout(set = 1, binding = 0) uniform sampler2D uTexture;
 layout(set = 1, binding = 1) buffer Counter
@@ -66,13 +67,12 @@ void write_image(ivec2 coord, int mip, float v)
     imageStore(uImages[mip - 1], coord, vec4(v));
 }
 
-void write_image4_top(ivec2 coord, int mip, vec4 v)
+#if defined(WRITE_TOP_LEVEL) && WRITE_TOP_LEVEL
+void write_image_top(ivec2 coord, float v)
 {
-    imageStore(uImageTop, coord + ivec2(0, 0), v.xxxx);
-    imageStore(uImageTop, coord + ivec2(1, 0), v.yyyy);
-    imageStore(uImageTop, coord + ivec2(0, 1), v.zzzz);
-    imageStore(uImageTop, coord + ivec2(1, 1), v.wwww);
+    imageStore(uImageTop, coord, vec4(v, 0, 0, 0));
 }
+#endif
 
 const int SHARED_WIDTH = 32;
 const int SHARED_HEIGHT = 32;
@@ -120,11 +120,35 @@ float fetch_image_mip6(ivec2 coord)
     return imageLoad(uImages[5], coord).x;
 }
 
-vec4 write_mip0_transformed(vec4 v, ivec2 base_coord)
+vec4 write_mip0_transformed(vec4 v, ivec2 base_coord, ivec2 local_coord)
 {
     v = transform_z(v);
+
+#if defined(WRITE_TOP_LEVEL) && WRITE_TOP_LEVEL
+    // Ensure that top-level image is written with full cache lines per warp.
+    // Writing in the strided 2x2 pattern is noticably bad for L2 performance.
+    // Taking extra time on the shader cores to reshuffle data is actually beneficial since we're fully bandwidth bound
+    // in these shaders, so we should give the memory system all the help it can get.
+    store_shared(2 * local_coord + ivec2(0, 0), v.x);
+    store_shared(2 * local_coord + ivec2(1, 0), v.y);
+    store_shared(2 * local_coord + ivec2(0, 1), v.z);
+    store_shared(2 * local_coord + ivec2(1, 1), v.w);
+
+    barrier();
+
     // Write out transformed LOD 0
-    write_image4_top(base_coord, 0, v);
+    for (int y = 0; y < 2; y++)
+    {
+        for (int x = 0; x < 2; x++)
+        {
+            ivec2 tile_offset = ivec2(x, y) * 16;
+            write_image_top(base_coord + tile_offset + local_coord, load_shared(local_coord + tile_offset));
+        }
+    }
+
+    barrier();
+#endif
+
     return v;
 }
 
@@ -241,10 +265,14 @@ void main()
     // It seems like we need to be super careful about memory access patterns to get optimal bandwidth.
 
     // LOD 0 feedback with transform.
-    vec4 tile00 = write_mip0_transformed(fetch_2x2_texture(base_coord_00), base_coord_00);
-    vec4 tile10 = write_mip0_transformed(fetch_2x2_texture(base_coord_10), base_coord_10);
-    vec4 tile01 = write_mip0_transformed(fetch_2x2_texture(base_coord_01), base_coord_01);
-    vec4 tile11 = write_mip0_transformed(fetch_2x2_texture(base_coord_11), base_coord_11);
+    vec4 tile00 = write_mip0_transformed(
+        fetch_2x2_texture(base_coord_00), ivec2(gl_WorkGroupID.xy * 64u) + ivec2(0, 0), ivec2(local_coord));
+    vec4 tile10 = write_mip0_transformed(
+        fetch_2x2_texture(base_coord_10), ivec2(gl_WorkGroupID.xy * 64u) + ivec2(32, 0), ivec2(local_coord));
+    vec4 tile01 = write_mip0_transformed(
+        fetch_2x2_texture(base_coord_01), ivec2(gl_WorkGroupID.xy * 64u) + ivec2(0, 32), ivec2(local_coord));
+    vec4 tile11 = write_mip0_transformed(
+        fetch_2x2_texture(base_coord_11), ivec2(gl_WorkGroupID.xy * 64u) + ivec2(32, 32), ivec2(local_coord));
     if (registers.mips <= 1)
         return;
 
@@ -275,6 +303,7 @@ void main()
         store_shared(local_coord_shared + ivec2(0, 8), reduced01);
         store_shared(local_coord_shared + ivec2(8, 8), reduced11);
     }
+
     barrier();
 
     // Write LOD 3

tests/hiz.cpp

@@ -97,14 +97,14 @@ int main()
 		dev.begin_renderdoc_capture();
 
 	auto cmd = dev.request_command_buffer();
-	cmd->set_program("builtin://shaders/post/hiz.comp");
+	cmd->set_program("builtin://shaders/post/hiz.comp", {{ "WRITE_TOP_LEVEL", 1 }});
 	for (unsigned i = 0; i < 13; i++)
 		cmd->set_storage_texture(0, i, *views[i < push.mips ? i : (push.mips - 1)]);
 	cmd->set_texture(1, 0, img->get_view(), StockSampler::NearestClamp);
 	cmd->set_storage_buffer(1, 1, *counter_buffer);
 	cmd->push_constants(&push, 0, sizeof(push));
 	cmd->enable_subgroup_size_control(true);
-	cmd->set_subgroup_size_log2(true, 4, 7);
+	cmd->set_subgroup_size_log2(true, 2, 7);
 	cmd->dispatch(wg_x, wg_y, 1);
 	dev.submit(cmd);
 

tests/meshlet_viewer.cpp

@@ -438,6 +438,7 @@ struct MeshletViewerApplication : Granite::Application, Granite::EventHandler //
 			mat4 view;
 			vec4 viewport_scale_bias;
 			uvec2 hiz_resolution;
+			uint hiz_min_lod;
 			uint hiz_max_lod;
 		};
 
@@ -463,9 +464,10 @@ struct MeshletViewerApplication : Granite::Application, Granite::EventHandler //
 
 				ubo->view = render_context.get_render_parameters().view;
 				ubo->viewport_scale_bias = viewport_scale_bias;
-				ubo->hiz_resolution.x = hiz->get_view_width();
-				ubo->hiz_resolution.y = hiz->get_view_height();
-				ubo->hiz_max_lod = hiz->get_create_info().levels - 1;
+				ubo->hiz_resolution.x = hiz->get_view_width() * 2;
+				ubo->hiz_resolution.y = hiz->get_view_height() * 2;
+				ubo->hiz_min_lod = 1;
+				ubo->hiz_max_lod = hiz->get_create_info().levels;
 			}
 		};
 
@@ -805,24 +807,15 @@ struct MeshletViewerApplication : Granite::Application, Granite::EventHandler //
 		    (depth_view.get_view_width() + 63u) & ~63u,
 		    (depth_view.get_view_height() + 63u) & ~63u,
 		    VK_FORMAT_R32_SFLOAT);
+		info.width /= 2;
+		info.height /= 2;
 		info.usage = VK_IMAGE_USAGE_STORAGE_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
 		info.initial_layout = VK_IMAGE_LAYOUT_UNDEFINED;
-		info.levels = Util::floor_log2(max(depth_view.get_view_width(), depth_view.get_view_height()));
+		info.levels = Util::floor_log2(max(depth_view.get_view_width(), depth_view.get_view_height())) - 1;
+		info.misc |= IMAGE_MISC_CREATE_PER_MIP_LEVEL_VIEWS_BIT;
 
 		auto hiz = device.create_image(info);
 
-		ImageViewHandle views[13];
-		for (unsigned i = 0; i < info.levels; i++)
-		{
-			ImageViewCreateInfo view = {};
-			view.base_level = i;
-			view.levels = 1;
-			view.image = hiz.get();
-			view.view_type = VK_IMAGE_VIEW_TYPE_2D;
-			view.aspect = VK_IMAGE_ASPECT_COLOR_BIT;
-			views[i] = device.create_image_view(view);
-		}
-
 		struct Push
 		{
 			mat2 z_transform;
@@ -846,22 +839,22 @@ struct MeshletViewerApplication : Granite::Application, Granite::EventHandler //
 
 		Push push = {};
 		push.z_transform = inv_z;
-		push.resolution = uvec2(info.width, info.height);
+		push.resolution = uvec2(info.width * 2, info.height * 2);
 		push.inv_resolution = vec2(1.0f / float(depth_view.get_view_width()), 1.0f / float(depth_view.get_view_height()));
-		push.mips = info.levels;
+		push.mips = info.levels + 1;
 
 		uint32_t wg_x = (push.resolution.x + 63) / 64;
 		uint32_t wg_y = (push.resolution.y + 63) / 64;
 		push.target_counter = wg_x * wg_y;
 
 		cmd->set_program("builtin://shaders/post/hiz.comp");
-		for (unsigned i = 0; i < 13; i++)
-			cmd->set_storage_texture(0, i, *views[i < push.mips ? i : (push.mips - 1)]);
+		for (unsigned i = 0; i < 12; i++)
+			cmd->set_storage_texture_level(0, i + 1, hiz->get_view(), i < info.levels ? i : (info.levels - 1));
 		cmd->set_texture(1, 0, depth_view, StockSampler::NearestClamp);
 		cmd->set_storage_buffer(1, 1, *counter);
 		cmd->push_constants(&push, 0, sizeof(push));
 		cmd->enable_subgroup_size_control(true);
-		cmd->set_subgroup_size_log2(true, 4, 7);
+		cmd->set_subgroup_size_log2(true, 2, 7);
 
 		auto start_ts = cmd->write_timestamp(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT);