Shadows.

.gitignore

@@ -22,3 +22,4 @@ docs/.asciidoctor/
 /Menu.sh
 /Menu.bat
 /data/
+screenshots/

CHANGELOG

@@ -21,7 +21,8 @@ API Changes
 - 3D scenes.  A game can draw a 3D scene between the disc video and the
   overlay:  primitive meshes and script-built geometry, materials with
   colour, textures from sprites, the disc or a loaded video, metallic and
-  roughness, up to eight lights, any node as the camera, and glTF 2.0
+  roughness, up to eight lights each able to cast shadows (cube maps for
+  point lights), any node as the camera, and glTF 2.0
   models (.glb, self-contained) with node animation and skinning, placed
   any number of times.  Everything is a node in one tree.  New calls:
   scene*, node*, mesh*, material*, light*, camera*, model*, animation*,

docs/Manual.adoc

@@ -759,7 +759,10 @@ Up to eight lights shine at once: `LIGHT_DIRECTIONAL` (a sun, position
 irrelevant), `LIGHT_POINT` (a bulb, fading with distance, optionally out to
 a range), and `LIGHT_SPOT` (a cone with inner and outer angles). Lights are
 nodes, so they parent and animate like anything else. `sceneSetAmbient`
-lights everything a little from everywhere. Any node can be the camera
+lights everything a little from everywhere. Any light can cast soft-edged
+shadows from everything opaque (`lightSetShadow`): directional and spot
+lights through a shadow map, point lights through a cube map that reaches
+every direction; `sceneSetShadowSize` trades sharpness for speed. Any node can be the camera
 (`cameraSet`); without one, the scene is viewed from `(0, 0, 5)` looking at
 the origin. Perspective is the default; `cameraSetOrthographic` is there
 for diagrams and HUD-like scenes.
@@ -2143,6 +2146,31 @@ How far a point or spot light reaches before fading to nothing, in world units;
 *Since:* 3.00.
 *See also:* <<lightsetintensity,lightSetIntensity>>
 
+[#lightsetshadow]
+==== lightSetShadow
+
+[source,text]
+----
+lightSetShadow(node, shadow)
+----
+
+Makes the light cast shadows from everything opaque in the scene, or stops it. Any of the eight lights may cast, in any mix. A directional or spot light gets one shadow map; a point light gets a cube map, six views, which covers every direction (a bulb inside a room) at six times the cost.
+
+*Notes:* Each shadow map costs one extra pass over the scene per frame, and each cube map six; the maps' size is set by `sceneSetShadowSize`. A modern integrated GPU manages eight casters comfortably; a Raspberry Pi wants one or two at `512`.
+
+*Since:* 3.00.
+*See also:* <<scenesetshadowsize,sceneSetShadowSize>>, <<lightnew,lightNew>>
+
+.Example
+[source,lua]
+----
+-- The sun casts shadows; the lamps only light.
+sun = lightNew(LIGHT_DIRECTIONAL)
+nodeSetPosition(sun, 3, 6, 4)
+nodeLookAt(sun, 0, 0, 0)
+lightSetShadow(sun, true)
+----
+
 [#material]
 === Material
 
@@ -3374,6 +3402,19 @@ The colour the layer clears to each frame. `a` below `255` lets the video show t
 *Since:* 3.00.
 *See also:* <<sceneenable,sceneEnable>>
 
+[#scenesetshadowsize]
+==== sceneSetShadowSize
+
+[source,text]
+----
+sceneSetShadowSize(size)
+----
+
+Texels per side of every shadow map and cube face, `256` to `4096` (default `1024`). Larger is sharper and slower; `512` is a fair choice on a Raspberry Pi.
+
+*Since:* 3.00.
+*See also:* <<lightsetshadow,lightSetShadow>>, <<scenesetantialias,sceneSetAntialias>>
+
 [#script]
 === Script
 

src/math3d.c

@@ -205,6 +205,22 @@ Mat4T mat4Orthographic(float width, float height, float near, float far) {
 }
 
 
+// An off-centre parallel projection (a shadow map fitted to what it covers).
+Mat4T mat4OrthographicBounds(float left, float right, float bottom, float top, float near, float far) {
+	Mat4T out;
+
+	memset(&out, 0, sizeof(out));
+	out.m[0]  = 2.0f / (right - left);
+	out.m[5]  = 2.0f / (top - bottom);
+	out.m[10] = -1.0f / (far - near);
+	out.m[12] = -(right + left) / (right - left);
+	out.m[13] = -(top + bottom) / (top - bottom);
+	out.m[14] = -near / (far - near);
+	out.m[15] = 1.0f;
+	return out;
+}
+
+
 Mat4T mat4Perspective(float fovDegrees, float aspect, float near, float far) {
 	Mat4T out;
 	float f = 1.0f / tanf(DEGREES_TO_RADIANS(fovDegrees) / 2.0f);

src/math3d.h

@@ -62,6 +62,7 @@ bool  mat4Invert(Mat4T a, Mat4T *out);
 Mat4T mat4LookAt(Vec3T eye, Vec3T target, Vec3T up);
 Mat4T mat4Multiply(Mat4T a, Mat4T b);
 Mat4T mat4Orthographic(float width, float height, float near, float far);
+Mat4T mat4OrthographicBounds(float left, float right, float bottom, float top, float near, float far);
 Mat4T mat4Perspective(float fovDegrees, float aspect, float near, float far);
 Vec3T mat4TransformPoint(Mat4T a, Vec3T p);
 Vec3T mat4TransformVector(Mat4T a, Vec3T v);

src/scene.c

@@ -52,6 +52,18 @@
 #define MIN_SEGMENTS       3
 #define PI                 3.14159265358979323846f
 #define NO_HANDLE          -1
+#define SHADOW_PIPELINES   4               // skinned x double sided
+#define MAX_SHADOWS        MAX_LIGHTS      // Every light may cast; the arrays are sized to the lights in use
+#define CUBE_FACES         6
+#define SHADOW_NONE        0
+#define SHADOW_MAP         1
+#define SHADOW_CUBE        2
+#define SHADOW_SIZE        1024
+#define SHADOW_SIZE_MIN    256
+#define SHADOW_SIZE_MAX    4096
+#define SHADOW_BIAS        0.0015f
+#define SHADOW_MARGIN      1.05f           // The fitted light frustum, a little larger than the scene
+#define SKIN_BOUNDS_GROW   1.5f            // A skinned mesh moves beyond its bind pose
 
 
 // Matches DrawUniforms in scene.hlsl.
@@ -76,13 +88,29 @@ typedef struct FragmentUniformsS {
 	float         emissive[4];
 	float         material[4];
 	float         counts[4];
+	float         shadowParams[4];
+	Mat4T         shadowMatrix[MAX_SHADOWS];
+	float         shadowInfo[MAX_SHADOWS][4];
 	LightUniformT lights[MAX_LIGHTS];
 } FragmentUniformsT;
 
+// One light's shadow for this frame.
+typedef struct ShadowS {
+	int32_t node;
+	int32_t type;                // SHADOW_MAP or SHADOW_CUBE
+	int32_t layer;               // First layer in the shadow array (a point light uses six)
+	Mat4T   matrix;              // Map:  the light's view-projection
+	Mat4T   faces[CUBE_FACES];   // Cube:  one per face
+	float   near;
+	float   far;
+} ShadowT;
+
 typedef struct MeshS {
 	SDL_GPUBuffer *vertexBuffer;
 	SDL_GPUBuffer *indexBuffer;
 	uint32_t       indexCount;
+	Vec3T          boundsMin;     // Of the vertices, for fitting the shadow map
+	Vec3T          boundsMax;
 	bool           skinned;
 	bool           used;
 } MeshT;
@@ -130,6 +158,7 @@ typedef struct NodeS {
 	int32_t  mesh;
 	int32_t  material;
 	LightT   light;
+	bool     castsShadow;
 	int32_t *skinJoints;         // Nodes whose world matrices drive a skinned mesh
 	Mat4T   *skinInverseBind;
 	int32_t  skinCount;
@@ -163,8 +192,18 @@ typedef struct SceneS {
 	SDL_GPUShader            *vertexStatic;
 	SDL_GPUShader            *vertexSkinned;
 	SDL_GPUShader            *fragment;
+	SDL_GPUShader            *depthFragment;  // Empty; the shadow pass writes depth only
 	SDL_GPUGraphicsPipeline  *pipelines[PIPELINE_COUNT];
+	SDL_GPUGraphicsPipeline  *shadowPipelines[SHADOW_PIPELINES];
 	SDL_GPUSampler           *sampler;
+	SDL_GPUSampler           *shadowSampler;
+	SDL_GPUTexture           *shadowMaps;     // 2D array:  a layer per directional or spot shadow, six per point light
+	SDL_GPUTexture           *shadowMapsNone; // 1x1 stand-in bound when nothing casts, so the shader's slot is filled
+	int32_t                   shadowMapLayers;
+	SDL_GPUTextureFormat      shadowFormat;
+	int32_t                   shadowSize;
+	ShadowT                   shadows[MAX_SHADOWS];
+	int32_t                   shadowCount;
 	SDL_GPUTexture           *white;          // 1x1 stand-in for untextured materials
 	SDL_FColor                background;
 	Vec3T                     ambient;
@@ -198,13 +237,20 @@ static int32_t          _allocNode(void);
 static void             _attach(int32_t node, int32_t parent);
 static int32_t          _compareDraws(const void *a, const void *b);
 static bool             _createPipeline(int32_t variant);
+static SDL_GPUTexture  *_createShadowArray(SDL_GPUTextureType type, int32_t layers, int32_t size);
+static bool             _createShadowMaps(int32_t layers);
+static bool             _createShadowPipeline(int32_t variant);
 static SDL_GPUShader   *_createShader(const SceneShaderT *shader, SDL_GPUShaderStage stage, uint32_t samplers, uint32_t uniforms);
 static bool             _createShaders(void);
 static SDL_GPUTextureFormat _depthFormat(void);
+static SDL_GPUTextureFormat _shadowFormat(void);
 static void             _destroyTargets(void);
 static void             _detach(int32_t node);
 static void             _destroyPipelines(void);
+static void             _destroyShadowMaps(void);
+static void             _drawList(SDL_GPUCommandBuffer *commands, SDL_GPURenderPass *pass, int32_t drawCount, const Mat4T *viewProjection, bool shadowPass, FragmentUniformsT *fragmentUniforms);
 static void             _fillLights(FragmentUniformsT *uniforms);
+static void             _fitShadows(int32_t drawCount);
 static void             _fillSkin(const NodeT *node, SkinUniformsT *uniforms);
 static void             _freeFeed(FeedT *feed);
 static void             _freeSkin(NodeT *node);
@@ -257,7 +303,24 @@ static int32_t _addMesh(const SceneVertexT *vertices, int32_t vertexCount, const
 	mesh->indexCount = (uint32_t)indexCount;
 	mesh->skinned    = skinned;
 	mesh->used       = true;
-	return x;
+	mesh->boundsMin  = vec3(vertices[0].position[0], vertices[0].position[1], vertices[0].position[2]);
+	mesh->boundsMax  = mesh->boundsMin;
+	for (x = 1; x < vertexCount; x++) {
+		mesh->boundsMin.x = SDL_min(mesh->boundsMin.x, vertices[x].position[0]);
+		mesh->boundsMin.y = SDL_min(mesh->boundsMin.y, vertices[x].position[1]);
+		mesh->boundsMin.z = SDL_min(mesh->boundsMin.z, vertices[x].position[2]);
+		mesh->boundsMax.x = SDL_max(mesh->boundsMax.x, vertices[x].position[0]);
+		mesh->boundsMax.y = SDL_max(mesh->boundsMax.y, vertices[x].position[1]);
+		mesh->boundsMax.z = SDL_max(mesh->boundsMax.z, vertices[x].position[2]);
+	}
+	if (skinned) {
+		Vec3T centre = vec3Scale(vec3Add(mesh->boundsMin, mesh->boundsMax), 0.5f);
+		Vec3T half   = vec3Scale(vec3Subtract(mesh->boundsMax, mesh->boundsMin), 0.5f * SKIN_BOUNDS_GROW);
+
+		mesh->boundsMin = vec3Subtract(centre, half);
+		mesh->boundsMax = vec3Add(centre, half);
+	}
+	return (int32_t)(mesh - _scene.meshes);
 }
 
 
@@ -490,8 +553,101 @@ static SDL_GPUShader *_createShader(const SceneShaderT *shader, SDL_GPUShaderSta
 static bool _createShaders(void) {
 	_scene.vertexStatic  = _createShader(&sceneShaderVertexStatic, SDL_GPU_SHADERSTAGE_VERTEX, 0, 1);
 	_scene.vertexSkinned = _createShader(&sceneShaderVertexSkinned, SDL_GPU_SHADERSTAGE_VERTEX, 0, 2);
-	_scene.fragment      = _createShader(&sceneShaderFragmentMain, SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
+	_scene.fragment      = _createShader(&sceneShaderFragmentMain, SDL_GPU_SHADERSTAGE_FRAGMENT, 2, 1);
-	return (_scene.vertexStatic != NULL) && (_scene.vertexSkinned != NULL) && (_scene.fragment != NULL);
+	_scene.depthFragment = _createShader(&sceneShaderDepthMain, SDL_GPU_SHADERSTAGE_FRAGMENT, 0, 0);
+	return (_scene.vertexStatic != NULL) && (_scene.vertexSkinned != NULL) && (_scene.fragment != NULL) && (_scene.depthFragment != NULL);
+}
+
+
+// A depth texture array the shadow passes render into and the main pass samples.
+static SDL_GPUTexture *_createShadowArray(SDL_GPUTextureType type, int32_t layers, int32_t size) {
+	SDL_GPUTextureCreateInfo info;
+	SDL_GPUTexture          *texture;
+
+	memset(&info, 0, sizeof(info));
+	info.type                 = type;
+	info.format               = _scene.shadowFormat;
+	info.usage                = SDL_GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET | SDL_GPU_TEXTUREUSAGE_SAMPLER;
+	info.width                = (Uint32)size;
+	info.height               = (Uint32)size;
+	info.layer_count_or_depth = (Uint32)layers;
+	info.num_levels           = 1;
+	info.sample_count         = SDL_GPU_SAMPLECOUNT_1;
+	texture = SDL_CreateGPUTexture(_scene.device, &info);
+	if (texture == NULL) {
+		utilTrace("Scene:  shadow map:  %s", SDL_GetError());
+	}
+	return texture;
+}
+
+
+// Makes sure the array has room for this frame's shadows; it grows to the most used so far.
+static bool _createShadowMaps(int32_t layers) {
+	if (layers > _scene.shadowMapLayers) {
+		if (_scene.shadowMaps != NULL) {
+			SDL_ReleaseGPUTexture(_scene.device, _scene.shadowMaps);
+		}
+		_scene.shadowMaps      = _createShadowArray(SDL_GPU_TEXTURETYPE_2D_ARRAY, layers, _scene.shadowSize);
+		_scene.shadowMapLayers = (_scene.shadowMaps != NULL) ? layers : 0;
+	}
+	return _scene.shadowMaps != NULL;
+}
+
+
+// Depth-only pipelines for the shadow pass:  the scene's vertex shaders with an empty fragment
+// shader, no colour target, and a depth bias against self-shadowing.
+static bool _createShadowPipeline(int32_t variant) {
+	SDL_GPUGraphicsPipelineCreateInfo info;
+	SDL_GPUVertexBufferDescription    buffers[1];
+	SDL_GPUVertexAttribute            attributes[5];
+
+	memset(&info, 0, sizeof(info));
+	memset(buffers, 0, sizeof(buffers));
+	memset(attributes, 0, sizeof(attributes));
+	buffers[0].slot               = 0;
+	buffers[0].pitch              = sizeof(SceneVertexT);
+	buffers[0].input_rate         = SDL_GPU_VERTEXINPUTRATE_VERTEX;
+	attributes[0].location        = 0;
+	attributes[0].format          = SDL_GPU_VERTEXELEMENTFORMAT_FLOAT3;
+	attributes[0].offset          = offsetof(SceneVertexT, position);
+	attributes[1].location        = 1;
+	attributes[1].format          = SDL_GPU_VERTEXELEMENTFORMAT_FLOAT3;
+	attributes[1].offset          = offsetof(SceneVertexT, normal);
+	attributes[2].location        = 2;
+	attributes[2].format          = SDL_GPU_VERTEXELEMENTFORMAT_FLOAT2;
+	attributes[2].offset          = offsetof(SceneVertexT, uv);
+	attributes[3].location        = 3;
+	attributes[3].format          = SDL_GPU_VERTEXELEMENTFORMAT_UBYTE4;
+	attributes[3].offset          = offsetof(SceneVertexT, joints);
+	attributes[4].location        = 4;
+	attributes[4].format          = SDL_GPU_VERTEXELEMENTFORMAT_FLOAT4;
+	attributes[4].offset          = offsetof(SceneVertexT, weights);
+	info.vertex_shader                                   = (variant & PIPELINE_SKINNED) ? _scene.vertexSkinned : _scene.vertexStatic;
+	info.fragment_shader                                 = _scene.depthFragment;
+	info.vertex_input_state.vertex_buffer_descriptions   = buffers;
+	info.vertex_input_state.num_vertex_buffers           = 1;
+	info.vertex_input_state.vertex_attributes            = attributes;
+	info.vertex_input_state.num_vertex_attributes        = 5;
+	info.primitive_type                                  = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
+	info.rasterizer_state.fill_mode                      = SDL_GPU_FILLMODE_FILL;
+	info.rasterizer_state.cull_mode                      = (variant & PIPELINE_TWO_SIDED) ? SDL_GPU_CULLMODE_NONE : SDL_GPU_CULLMODE_BACK;
+	info.rasterizer_state.front_face                     = SDL_GPU_FRONTFACE_COUNTER_CLOCKWISE;
+	info.rasterizer_state.enable_depth_bias              = true;
+	info.rasterizer_state.depth_bias_constant_factor     = 2.0f;
+	info.rasterizer_state.depth_bias_slope_factor        = 2.0f;
+	info.multisample_state.sample_count                  = SDL_GPU_SAMPLECOUNT_1;
+	info.depth_stencil_state.compare_op                  = SDL_GPU_COMPAREOP_LESS_OR_EQUAL;
+	info.depth_stencil_state.enable_depth_test           = true;
+	info.depth_stencil_state.enable_depth_write          = true;
+	info.target_info.num_color_targets                   = 0;
+	info.target_info.depth_stencil_format                = _scene.shadowFormat;
+	info.target_info.has_depth_stencil_target            = true;
+	_scene.shadowPipelines[variant] = SDL_CreateGPUGraphicsPipeline(_scene.device, &info);
+	if (_scene.shadowPipelines[variant] == NULL) {
+		utilTrace("Scene:  shadow pipeline %d:  %s", variant, SDL_GetError());
+		return false;
+	}
+	return true;
 }
 
 
@@ -520,6 +676,21 @@ static void _destroyPipelines(void) {
 			_scene.pipelines[x] = NULL;
 		}
 	}
+	for (x = 0; x < SHADOW_PIPELINES; x++) {
+		if (_scene.shadowPipelines[x] != NULL) {
+			SDL_ReleaseGPUGraphicsPipeline(_scene.device, _scene.shadowPipelines[x]);
+			_scene.shadowPipelines[x] = NULL;
+		}
+	}
+}
+
+
+static void _destroyShadowMaps(void) {
+	if (_scene.shadowMaps != NULL) {
+		SDL_ReleaseGPUTexture(_scene.device, _scene.shadowMaps);
+		_scene.shadowMaps = NULL;
+	}
+	_scene.shadowMapLayers = 0;
 }
 
 
@@ -546,6 +717,108 @@ static void _destroyTargets(void) {
 }
 
 
+// Issues the collected draws into a pass:  the shadow pass with the light's view-projection and
+// depth-only pipelines (blended meshes cast nothing), or the main pass with the camera's and the
+// full material.
+static void _drawList(SDL_GPUCommandBuffer *commands, SDL_GPURenderPass *pass, int32_t drawCount, const Mat4T *viewProjection, bool shadowPass, FragmentUniformsT *fragmentUniforms) {
+	SDL_GPUBufferBinding         binding;
+	SDL_GPUTextureSamplerBinding samplerBindings[2];
+	DrawUniformsT                drawUniforms;
+	SkinUniformsT               *skinUniforms = NULL;
+	Mat4T                        identity     = mat4Identity();
+	Mat4T                        inverse;
+	int32_t                      x;
+	int32_t                      lastPipeline = NO_HANDLE;
+	int32_t                      variant;
+	NodeT                       *node;
+	MeshT                       *mesh;
+	MaterialT                   *material;
+	MaterialT                    defaultMaterial;
+	SDL_GPUGraphicsPipeline     *pipeline;
+
+	memset(&defaultMaterial, 0, sizeof(defaultMaterial));
+	defaultMaterial.baseColor.x = 1.0f;
+	defaultMaterial.baseColor.y = 1.0f;
+	defaultMaterial.baseColor.z = 1.0f;
+	defaultMaterial.baseColor.w = 1.0f;
+	defaultMaterial.roughness   = 0.5f;
+	defaultMaterial.feed        = NO_HANDLE;
+	for (x = 0; x < drawCount; x++) {
+		node     = &_scene.nodes[_scene.draws[x].node];
+		mesh     = &_scene.meshes[node->mesh];
+		material = (node->material != NO_HANDLE) ? &_scene.materials[node->material] : &defaultMaterial;
+		variant  = _lookupPipeline(_scene.draws[x].node);
+		if (variant == NO_HANDLE) {
+			continue;
+		}
+		if (shadowPass) {
+			if (material->blend) {
+				continue;
+			}
+			variant &= PIPELINE_SKINNED | PIPELINE_TWO_SIDED;
+			variant  = (variant & PIPELINE_SKINNED ? 1 : 0) | (variant & PIPELINE_TWO_SIDED ? 2 : 0);
+			if ((_scene.shadowPipelines[variant] == NULL) && !_createShadowPipeline(variant)) {
+				continue;
+			}
+			pipeline = _scene.shadowPipelines[variant];
+		} else {
+			pipeline = _scene.pipelines[variant];
+		}
+		if (variant != lastPipeline) {
+			SDL_BindGPUGraphicsPipeline(pass, pipeline);
+			lastPipeline = variant;
+		}
+		drawUniforms.modelViewProjection = mat4Multiply(*viewProjection, node->world);
+		drawUniforms.model               = node->world;
+		if (mat4Invert(node->world, &inverse)) {
+			drawUniforms.normalMatrix = mat4Transpose(inverse);
+		} else {
+			drawUniforms.normalMatrix = identity;
+		}
+		SDL_PushGPUVertexUniformData(commands, 0, &drawUniforms, sizeof(drawUniforms));
+		if (mesh->skinned && (node->skinCount > 0)) {
+			// 8 KB per skinned draw; allocated once per pass that needs it.
+			if (skinUniforms == NULL) {
+				skinUniforms = SDL_malloc(sizeof(SkinUniformsT));
+				if (skinUniforms == NULL) {
+					utilDie("Out of memory posing a skin.");
+				}
+			}
+			_fillSkin(node, skinUniforms);
+			SDL_PushGPUVertexUniformData(commands, 1, skinUniforms, sizeof(SkinUniformsT));
+		}
+		if (!shadowPass) {
+			fragmentUniforms->baseColor[0] = material->baseColor.x;
+			fragmentUniforms->baseColor[1] = material->baseColor.y;
+			fragmentUniforms->baseColor[2] = material->baseColor.z;
+			fragmentUniforms->baseColor[3] = material->baseColor.w;
+			fragmentUniforms->emissive[0]  = material->emissive.x;
+			fragmentUniforms->emissive[1]  = material->emissive.y;
+			fragmentUniforms->emissive[2]  = material->emissive.z;
+			fragmentUniforms->emissive[3]  = 1.0f;
+			fragmentUniforms->material[0]  = material->metallic;
+			fragmentUniforms->material[1]  = material->roughness;
+			fragmentUniforms->material[2]  = material->unlit ? 1.0f : 0.0f;
+			fragmentUniforms->material[3]  = (_materialTexture(material) != NULL) ? 1.0f : 0.0f;
+			SDL_PushGPUFragmentUniformData(commands, 0, fragmentUniforms, sizeof(FragmentUniformsT));
+			memset(samplerBindings, 0, sizeof(samplerBindings));
+			samplerBindings[0].texture = (_materialTexture(material) != NULL) ? _materialTexture(material) : _scene.white;
+			samplerBindings[0].sampler = _scene.sampler;
+			samplerBindings[1].texture = (_scene.shadowMaps != NULL) ? _scene.shadowMaps : _scene.shadowMapsNone;
+			samplerBindings[1].sampler = _scene.shadowSampler;
+			SDL_BindGPUFragmentSamplers(pass, 0, samplerBindings, 2);
+		}
+		memset(&binding, 0, sizeof(binding));
+		binding.buffer = mesh->vertexBuffer;
+		SDL_BindGPUVertexBuffers(pass, 0, &binding, 1);
+		binding.buffer = mesh->indexBuffer;
+		SDL_BindGPUIndexBuffer(pass, &binding, SDL_GPU_INDEXELEMENTSIZE_32BIT);
+		SDL_DrawGPUIndexedPrimitives(pass, mesh->indexCount, 1, 0, 0, 0);
+	}
+	SDL_free(skinUniforms);
+}
+
+
 // Unlinks node from its parent's child list.
 static void _detach(int32_t node) {
 	int32_t parent = _scene.nodes[node].parent;
@@ -570,10 +843,12 @@ static void _detach(int32_t node) {
 }
 
 
-// The first MAX_LIGHTS visible lights, in world space.
+// The first MAX_LIGHTS visible lights, in world space, and the shadow slots:  every one of them
+// flagged to cast gets a slot (a map for directional and spot lights, a cube for point lights).
 static void _fillLights(FragmentUniformsT *uniforms) {
 	int32_t        x;
 	int32_t        count  = 0;
+	int32_t        layers = 0;
 	NodeT         *node;
 	LightUniformT *light;
 	Vec3T          position;
@@ -603,6 +878,22 @@ static void _fillLights(FragmentUniformsT *uniforms) {
 		light->cone[1]           = cosf(node->light.outerDegrees * PI / 180.0f);
 		light->cone[2]           = 0.0f;
 		light->cone[3]           = 0.0f;
+		if (node->castsShadow && (_scene.shadowCount < MAX_SHADOWS)) {
+			ShadowT *shadow = &_scene.shadows[_scene.shadowCount];
+
+			memset(shadow, 0, sizeof(*shadow));
+			shadow->node  = x;
+			shadow->layer = layers;
+			if (node->light.type == LIGHT_POINT) {
+				shadow->type = SHADOW_CUBE;
+				layers      += CUBE_FACES;
+			} else {
+				shadow->type = SHADOW_MAP;
+				layers++;
+			}
+			light->cone[2] = (float)(_scene.shadowCount + 1);
+			_scene.shadowCount++;
+		}
 		count++;
 	}
 	uniforms->counts[0] = (float)count;
@@ -633,6 +924,102 @@ static void _fillSkin(const NodeT *node, SkinUniformsT *uniforms) {
 }
 
 
+// Every shadow's projection, fitted to what is drawn:  a directional light gets a parallel box
+// round the scene's bounds, a spot light its own cone, a point light six 90 degree faces out to
+// its range or the far edge of the scene.
+static void _fitShadows(int32_t drawCount) {
+	Vec3T   corners[8];
+	Vec3T   boundsMin = vec3(0.0f, 0.0f, 0.0f);
+	Vec3T   boundsMax = vec3(0.0f, 0.0f, 0.0f);
+	Vec3T   centre;
+	float   radius;
+	int32_t x;
+	int32_t c;
+	int32_t slot;
+	bool    any = false;
+	// The six faces of a point light's shadow:  the direction each looks and its up.  The fragment
+	// shader rebuilds the same frames to look them up, so these need only match it.
+	Vec3T   faceForward[CUBE_FACES] = { { 1.0f, 0.0f, 0.0f }, { -1.0f, 0.0f, 0.0f }, { 0.0f, 1.0f, 0.0f }, { 0.0f, -1.0f, 0.0f }, { 0.0f, 0.0f, 1.0f }, { 0.0f, 0.0f, -1.0f } };
+	Vec3T   faceUp[CUBE_FACES]      = { { 0.0f, 1.0f, 0.0f }, { 0.0f, 1.0f, 0.0f }, { 0.0f, 0.0f, -1.0f }, { 0.0f, 0.0f, 1.0f }, { 0.0f, 1.0f, 0.0f }, { 0.0f, 1.0f, 0.0f } };
+
+	// World bounds of everything drawn.
+	for (x = 0; x < drawCount; x++) {
+		NodeT *node = &_scene.nodes[_scene.draws[x].node];
+		MeshT *mesh = &_scene.meshes[node->mesh];
+
+		for (c = 0; c < 8; c++) {
+			Vec3T corner = vec3((c & 1) ? mesh->boundsMax.x : mesh->boundsMin.x, (c & 2) ? mesh->boundsMax.y : mesh->boundsMin.y, (c & 4) ? mesh->boundsMax.z : mesh->boundsMin.z);
+
+			corner = mat4TransformPoint(node->world, corner);
+			if (!any) {
+				boundsMin = corner;
+				boundsMax = corner;
+				any       = true;
+			} else {
+				boundsMin = vec3(SDL_min(boundsMin.x, corner.x), SDL_min(boundsMin.y, corner.y), SDL_min(boundsMin.z, corner.z));
+				boundsMax = vec3(SDL_max(boundsMax.x, corner.x), SDL_max(boundsMax.y, corner.y), SDL_max(boundsMax.z, corner.z));
+			}
+		}
+	}
+	centre = vec3Scale(vec3Add(boundsMin, boundsMax), 0.5f);
+	radius = vec3Length(vec3Subtract(boundsMax, centre)) * SHADOW_MARGIN + 0.001f;
+	for (c = 0; c < 8; c++) {
+		corners[c] = vec3((c & 1) ? boundsMax.x : boundsMin.x, (c & 2) ? boundsMax.y : boundsMin.y, (c & 4) ? boundsMax.z : boundsMin.z);
+	}
+	for (slot = 0; slot < _scene.shadowCount; slot++) {
+		ShadowT *shadow    = &_scene.shadows[slot];
+		NodeT   *light     = &_scene.nodes[shadow->node];
+		Vec3T    direction = vec3Normalize(mat4TransformVector(light->world, vec3(0.0f, 0.0f, -1.0f)));
+		Vec3T    position  = mat4TransformPoint(light->world, vec3(0.0f, 0.0f, 0.0f));
+		Vec3T    up        = (fabsf(direction.y) < 0.99f) ? vec3(0.0f, 1.0f, 0.0f) : vec3(0.0f, 0.0f, 1.0f);
+		Mat4T    view;
+		float    far;
+		float    minX = 0.0f;
+		float    maxX = 0.0f;
+		float    minY = 0.0f;
+		float    maxY = 0.0f;
+		float    minZ = 0.0f;
+		float    maxZ = 0.0f;
+
+		if (light->light.type == LIGHT_DIRECTIONAL) {
+			view = mat4LookAt(vec3Subtract(centre, vec3Scale(direction, radius * 2.0f)), centre, up);
+			for (c = 0; c < 8; c++) {
+				Vec3T v = mat4TransformPoint(view, corners[c]);
+
+				if (c == 0) {
+					minX = v.x;
+					maxX = v.x;
+					minY = v.y;
+					maxY = v.y;
+					minZ = v.z;
+					maxZ = v.z;
+				} else {
+					minX = SDL_min(minX, v.x);
+					maxX = SDL_max(maxX, v.x);
+					minY = SDL_min(minY, v.y);
+					maxY = SDL_max(maxY, v.y);
+					minZ = SDL_min(minZ, v.z);
+					maxZ = SDL_max(maxZ, v.z);
+				}
+			}
+			// View space looks down -Z:  the nearest point has the largest z.
+			shadow->matrix = mat4Multiply(mat4OrthographicBounds(minX * SHADOW_MARGIN, maxX * SHADOW_MARGIN, minY * SHADOW_MARGIN, maxY * SHADOW_MARGIN, SDL_max(-maxZ / SHADOW_MARGIN, 0.01f), -minZ * SHADOW_MARGIN), view);
+		} else if (light->light.type == LIGHT_SPOT) {
+			view = mat4LookAt(position, vec3Add(position, direction), up);
+			far  = (light->light.range > 0.0f) ? light->light.range : vec3Length(vec3Subtract(centre, position)) + radius;
+			shadow->matrix = mat4Multiply(mat4Perspective(SDL_min(light->light.outerDegrees * 2.0f * SHADOW_MARGIN, 170.0f), 1.0f, SDL_max(far * 0.005f, 0.01f), far), view);
+		} else {
+			far          = (light->light.range > 0.0f) ? light->light.range : vec3Length(vec3Subtract(centre, position)) + radius;
+			shadow->near = SDL_max(far * 0.005f, 0.01f);
+			shadow->far  = far;
+			for (c = 0; c < CUBE_FACES; c++) {
+				shadow->faces[c] = mat4Multiply(mat4Perspective(90.0f, 1.0f, shadow->near, shadow->far), mat4LookAt(position, vec3Add(position, faceForward[c]), faceUp[c]));
+			}
+		}
+	}
+}
+
+
 static void _freeFeed(FeedT *feed) {
 	if (feed->target != NULL) {
 		SDL_DestroyTexture(feed->target);
@@ -772,6 +1159,20 @@ static Mat4T _projection(void) {
 }
 
 
+// A depth format the shadow map can be both rendered into and sampled from.
+static SDL_GPUTextureFormat _shadowFormat(void) {
+	SDL_GPUTextureFormat wanted[] = { SDL_GPU_TEXTUREFORMAT_D32_FLOAT, SDL_GPU_TEXTUREFORMAT_D16_UNORM, SDL_GPU_TEXTUREFORMAT_D24_UNORM };
+	int32_t              x;
+
+	for (x = 0; x < (int32_t)SDL_arraysize(wanted); x++) {
+		if (SDL_GPUTextureSupportsFormat(_scene.device, wanted[x], SDL_GPU_TEXTURETYPE_2D, SDL_GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET | SDL_GPU_TEXTUREUSAGE_SAMPLER)) {
+			return wanted[x];
+		}
+	}
+	return SDL_GPU_TEXTUREFORMAT_D16_UNORM;
+}
+
+
 // World matrices and inherited visibility, depth first.
 static void _updateWorld(int32_t node, const Mat4T *parentWorld, bool parentVisible) {
 	NodeT  *n = &_scene.nodes[node];
@@ -1040,6 +1441,17 @@ bool lightSetRange(int32_t node, float range) {
 }
 
 
+// Whether this light casts shadows.  Every light may; a point light's cost six passes to a
+// directional or spot light's one.
+bool lightSetShadow(int32_t node, bool shadow) {
+	if (!nodeValid(node) || !_scene.nodes[node].hasLight) {
+		return false;
+	}
+	_scene.nodes[node].castsShadow = shadow;
+	return true;
+}
+
+
 // ===== Materials =====
 
 bool materialDelete(int32_t material) {
@@ -1801,10 +2213,12 @@ bool sceneInit(SDL_GPUDevice *device, SDL_Renderer *renderer) {
 	_scene.ambient     = vec3(0.1f, 0.1f, 0.1f);
 	_scene.antialias   = true;
 	_scene.sampleCount = SDL_GPU_SAMPLECOUNT_1;
+	_scene.shadowSize  = SHADOW_SIZE;
 	if (device == NULL) {
 		return false;
 	}
 	_scene.depthFormat  = _depthFormat();
+	_scene.shadowFormat = _shadowFormat();
 	_allocNode();
 	nodeSetName(SCENE_ROOT_NODE, "root");
 	if (!_createShaders()) {
@@ -1819,13 +2233,22 @@ bool sceneInit(SDL_GPUDevice *device, SDL_Renderer *renderer) {
 	samplerInfo.address_mode_v = SDL_GPU_SAMPLERADDRESSMODE_REPEAT;
 	samplerInfo.address_mode_w = SDL_GPU_SAMPLERADDRESSMODE_REPEAT;
 	_scene.sampler = SDL_CreateGPUSampler(device, &samplerInfo);
+	// The shadow map is compared texel by texel, so no filtering, and clamped so its edge holds.
+	samplerInfo.min_filter     = SDL_GPU_FILTER_NEAREST;
+	samplerInfo.mag_filter     = SDL_GPU_FILTER_NEAREST;
+	samplerInfo.mipmap_mode    = SDL_GPU_SAMPLERMIPMAPMODE_NEAREST;
+	samplerInfo.address_mode_u = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
+	samplerInfo.address_mode_v = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
+	samplerInfo.address_mode_w = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
+	_scene.shadowSampler = SDL_CreateGPUSampler(device, &samplerInfo);
 	pixel = SDL_CreateSurface(1, 1, SDL_PIXELFORMAT_RGBA32);
 	if (pixel != NULL) {
 		SDL_FillSurfaceRect(pixel, NULL, 0xFFFFFFFFu);
 		_scene.white = _uploadTexture(pixel);
 		SDL_DestroySurface(pixel);
 	}
-	if ((_scene.sampler == NULL) || (_scene.white == NULL)) {
+	_scene.shadowMapsNone = _createShadowArray(SDL_GPU_TEXTURETYPE_2D_ARRAY, 1, 1);
+	if ((_scene.sampler == NULL) || (_scene.shadowSampler == NULL) || (_scene.white == NULL) || (_scene.shadowMapsNone == NULL)) {
 		utilTrace("Scene:  %s", SDL_GetError());
 		sceneQuit();
 		return false;
@@ -1880,9 +2303,19 @@ void sceneQuit(void) {
 		if (_scene.fragment != NULL) {
 			SDL_ReleaseGPUShader(_scene.device, _scene.fragment);
 		}
+		if (_scene.depthFragment != NULL) {
+			SDL_ReleaseGPUShader(_scene.device, _scene.depthFragment);
+		}
 		if (_scene.sampler != NULL) {
 			SDL_ReleaseGPUSampler(_scene.device, _scene.sampler);
 		}
+		if (_scene.shadowSampler != NULL) {
+			SDL_ReleaseGPUSampler(_scene.device, _scene.shadowSampler);
+		}
+		_destroyShadowMaps();
+		if (_scene.shadowMapsNone != NULL) {
+			SDL_ReleaseGPUTexture(_scene.device, _scene.shadowMapsNone);
+		}
 		if (_scene.white != NULL) {
 			SDL_ReleaseGPUTexture(_scene.device, _scene.white);
 		}
@@ -1909,24 +2342,17 @@ SDL_Texture *sceneRender(void) {
 	SDL_GPUDepthStencilTargetInfo depth;
 	SDL_GPUCommandBuffer         *commands;
 	SDL_GPURenderPass            *pass;
-	SDL_GPUBufferBinding          binding;
-	SDL_GPUTextureSamplerBinding  samplerBinding;
-	DrawUniformsT                 drawUniforms;
-	SkinUniformsT                *skinUniforms = NULL;
 	FragmentUniformsT             fragmentUniforms;
 	Mat4T                         identity = mat4Identity();
 	Mat4T                         view;
-	Mat4T                         inverse;
 	Vec3T                         eye;
 	int32_t                       x;
 	int32_t                       drawCount   = 0;
 	int32_t                       opaqueCount = 0;
-	int32_t                       lastPipeline = NO_HANDLE;
+	int32_t                       slot;
-	int32_t                       variant;
+	int32_t                       face;
+	int32_t                       layers      = 0;
 	NodeT                        *node;
-	MeshT                        *mesh;
-	MaterialT                    *material;
-	MaterialT                     defaultMaterial;
 
 	if (!_scene.enabled || (_scene.colour == NULL)) {
 		return NULL;
@@ -1945,6 +2371,7 @@ SDL_Texture *sceneRender(void) {
 	fragmentUniforms.ambient[1]        = _scene.ambient.y;
 	fragmentUniforms.ambient[2]        = _scene.ambient.z;
 	fragmentUniforms.ambient[3]        = 1.0f;
+	_scene.shadowCount = 0;
 	_fillLights(&fragmentUniforms);
 	// Collect what to draw:  opaque first in node order, then blended back to front.
 	if (_scene.drawCapacity < _scene.nodeCount) {
@@ -1980,12 +2407,57 @@ SDL_Texture *sceneRender(void) {
 	if (drawCount > opaqueCount) {
 		qsort(&_scene.draws[opaqueCount], (size_t)(drawCount - opaqueCount), sizeof(DrawT), _compareDraws);
 	}
-	// Record the pass.
 	commands = SDL_AcquireGPUCommandBuffer(_scene.device);
 	if (commands == NULL) {
 		utilTrace("Scene:  %s", SDL_GetError());
 		return NULL;
 	}
+	// The shadow passes:  depth from every shadow light into its map layer, or its six cube faces.
+	for (slot = 0; slot < _scene.shadowCount; slot++) {
+		layers += (_scene.shadows[slot].type == SHADOW_CUBE) ? CUBE_FACES : 1;
+	}
+	if ((_scene.shadowCount > 0) && (drawCount > 0) && _createShadowMaps(layers)) {
+		_fitShadows(drawCount);
+		fragmentUniforms.shadowParams[0] = SHADOW_BIAS;
+		fragmentUniforms.shadowParams[1] = 1.0f / (float)_scene.shadowSize;
+		fragmentUniforms.shadowParams[2] = (float)_scene.shadowCount;
+		memset(&depth, 0, sizeof(depth));
+		depth.clear_depth      = 1.0f;
+		depth.load_op          = SDL_GPU_LOADOP_CLEAR;
+		depth.store_op         = SDL_GPU_STOREOP_STORE;
+		depth.stencil_load_op  = SDL_GPU_LOADOP_DONT_CARE;
+		depth.stencil_store_op = SDL_GPU_STOREOP_DONT_CARE;
+		for (slot = 0; slot < _scene.shadowCount; slot++) {
+			ShadowT *shadow = &_scene.shadows[slot];
+
+			fragmentUniforms.shadowInfo[slot][0] = (float)shadow->type;
+			fragmentUniforms.shadowInfo[slot][1] = (float)shadow->layer;
+			fragmentUniforms.shadowInfo[slot][2] = shadow->near;
+			fragmentUniforms.shadowInfo[slot][3] = shadow->far;
+			if (shadow->type == SHADOW_CUBE) {
+				for (face = 0; face < CUBE_FACES; face++) {
+					depth.texture = _scene.shadowMaps;
+					depth.layer   = (Uint8)(shadow->layer + face);
+					pass = SDL_BeginGPURenderPass(commands, NULL, 0, &depth);
+					_drawList(commands, pass, drawCount, &shadow->faces[face], true, NULL);
+					SDL_EndGPURenderPass(pass);
+				}
+			} else {
+				fragmentUniforms.shadowMatrix[slot] = shadow->matrix;
+				depth.texture = _scene.shadowMaps;
+				depth.layer   = (Uint8)shadow->layer;
+				pass = SDL_BeginGPURenderPass(commands, NULL, 0, &depth);
+				_drawList(commands, pass, drawCount, &shadow->matrix, true, NULL);
+				SDL_EndGPURenderPass(pass);
+			}
+		}
+	} else {
+		// Lights flagged to cast but nothing to draw into:  no slots this frame.
+		for (x = 0; x < MAX_LIGHTS; x++) {
+			fragmentUniforms.lights[x].cone[2] = 0.0f;
+		}
+	}
+	// The main pass.
 	memset(&colour, 0, sizeof(colour));
 	colour.clear_color = _scene.background;
 	colour.load_op     = SDL_GPU_LOADOP_CLEAR;
@@ -2005,71 +2477,9 @@ SDL_Texture *sceneRender(void) {
 	depth.stencil_load_op  = SDL_GPU_LOADOP_DONT_CARE;
 	depth.stencil_store_op = SDL_GPU_STOREOP_DONT_CARE;
 	pass = SDL_BeginGPURenderPass(commands, &colour, 1, &depth);
-	memset(&defaultMaterial, 0, sizeof(defaultMaterial));
+	_drawList(commands, pass, drawCount, &_scene.viewProjection, false, &fragmentUniforms);
-	defaultMaterial.feed        = NO_HANDLE;
-	defaultMaterial.baseColor.x = 1.0f;
-	defaultMaterial.baseColor.y = 1.0f;
-	defaultMaterial.baseColor.z = 1.0f;
-	defaultMaterial.baseColor.w = 1.0f;
-	defaultMaterial.roughness   = 0.5f;
-	for (x = 0; x < drawCount; x++) {
-		node     = &_scene.nodes[_scene.draws[x].node];
-		mesh     = &_scene.meshes[node->mesh];
-		material = (node->material != NO_HANDLE) ? &_scene.materials[node->material] : &defaultMaterial;
-		variant  = _lookupPipeline(_scene.draws[x].node);
-		if (variant == NO_HANDLE) {
-			continue;
-		}
-		if (variant != lastPipeline) {
-			SDL_BindGPUGraphicsPipeline(pass, _scene.pipelines[variant]);
-			lastPipeline = variant;
-		}
-		drawUniforms.modelViewProjection = mat4Multiply(_scene.viewProjection, node->world);
-		drawUniforms.model               = node->world;
-		if (mat4Invert(node->world, &inverse)) {
-			drawUniforms.normalMatrix = mat4Transpose(inverse);
-		} else {
-			drawUniforms.normalMatrix = identity;
-		}
-		SDL_PushGPUVertexUniformData(commands, 0, &drawUniforms, sizeof(drawUniforms));
-		if (variant & PIPELINE_SKINNED) {
-			// 8 KB per skinned draw; allocated once per frame that needs it.
-			if (skinUniforms == NULL) {
-				skinUniforms = SDL_malloc(sizeof(SkinUniformsT));
-				if (skinUniforms == NULL) {
-					utilDie("Out of memory posing a skin.");
-				}
-			}
-			_fillSkin(node, skinUniforms);
-			SDL_PushGPUVertexUniformData(commands, 1, skinUniforms, sizeof(SkinUniformsT));
-		}
-		fragmentUniforms.baseColor[0] = material->baseColor.x;
-		fragmentUniforms.baseColor[1] = material->baseColor.y;
-		fragmentUniforms.baseColor[2] = material->baseColor.z;
-		fragmentUniforms.baseColor[3] = material->baseColor.w;
-		fragmentUniforms.emissive[0]  = material->emissive.x;
-		fragmentUniforms.emissive[1]  = material->emissive.y;
-		fragmentUniforms.emissive[2]  = material->emissive.z;
-		fragmentUniforms.emissive[3]  = 1.0f;
-		fragmentUniforms.material[0]  = material->metallic;
-		fragmentUniforms.material[1]  = material->roughness;
-		fragmentUniforms.material[2]  = material->unlit ? 1.0f : 0.0f;
-		fragmentUniforms.material[3]  = (_materialTexture(material) != NULL) ? 1.0f : 0.0f;
-		SDL_PushGPUFragmentUniformData(commands, 0, &fragmentUniforms, sizeof(fragmentUniforms));
-		memset(&samplerBinding, 0, sizeof(samplerBinding));
-		samplerBinding.texture = (_materialTexture(material) != NULL) ? _materialTexture(material) : _scene.white;
-		samplerBinding.sampler = _scene.sampler;
-		SDL_BindGPUFragmentSamplers(pass, 0, &samplerBinding, 1);
-		memset(&binding, 0, sizeof(binding));
-		binding.buffer = mesh->vertexBuffer;
-		SDL_BindGPUVertexBuffers(pass, 0, &binding, 1);
-		binding.buffer = mesh->indexBuffer;
-		SDL_BindGPUIndexBuffer(pass, &binding, SDL_GPU_INDEXELEMENTSIZE_32BIT);
-		SDL_DrawGPUIndexedPrimitives(pass, mesh->indexCount, 1, 0, 0, 0);
-	}
 	SDL_EndGPURenderPass(pass);
 	SDL_SubmitGPUCommandBuffer(commands);
-	SDL_free(skinUniforms);
 	return _scene.composite;
 }
 
@@ -2206,6 +2616,14 @@ void sceneSetBackground(uint8_t r, uint8_t g, uint8_t b, uint8_t a) {
 }
 
 
+// The shadow map's size in texels per side (default 1024, clamped to 256..4096); larger is
+// sharper and slower.  The map is rebuilt on the next frame that needs it.
+void sceneSetShadowSize(int32_t size) {
+	_scene.shadowSize = SDL_clamp(size, SHADOW_SIZE_MIN, SHADOW_SIZE_MAX);
+	_destroyShadowMaps();
+}
+
+
 // Overlay coordinates back to a world point:  the point on the ray through that pixel at the given
 // distance from the camera's near plane, using the last rendered frame's camera.  Two distances
 // give a ray for picking.

src/scene.h

@@ -63,6 +63,7 @@ void         sceneSetAntialias(bool antialias);
 void         sceneSetAmbient(uint8_t r, uint8_t g, uint8_t b);
 void         sceneComputeNormals(SceneVertexT *vertices, int32_t vertexCount, const uint32_t *indices, int32_t indexCount);
 void         sceneSetBackground(uint8_t r, uint8_t g, uint8_t b, uint8_t a);
+void         sceneSetShadowSize(int32_t size);
 Vec3T        sceneUnproject(float x, float y, float distance);
 void         sceneUpdateVideo(SceneVideoSourceFn source);
 
@@ -76,6 +77,7 @@ bool         lightSetColor(int32_t node, uint8_t r, uint8_t g, uint8_t b);
 bool         lightSetCone(int32_t node, float innerDegrees, float outerDegrees);
 bool         lightSetIntensity(int32_t node, float intensity);
 bool         lightSetRange(int32_t node, float range);
+bool         lightSetShadow(int32_t node, bool shadow);
 
 bool         materialDelete(int32_t material);
 int32_t      materialNew(void);

src/shaders/build.sh

@@ -14,7 +14,7 @@ OUT=sceneShaders.h
 TMP=$(mktemp -d)
 trap 'rm -rf "$TMP"' EXIT
 
-ENTRIES="vertexStatic:vertex vertexSkinned:vertex fragmentMain:fragment"
+ENTRIES="vertexStatic:vertex vertexSkinned:vertex fragmentMain:fragment depthMain:fragment"
 
 emit() {
 	# emit NAME FILE:  a C array from a binary file

src/shaders/scene.hlsl

@@ -1,5 +1,6 @@
 // Singe 3 scene shaders:  one vertex shader for static meshes, one for skinned meshes, one
-// fragment shader for both.  Compiled offline by src/shaders/build.sh (SDL_shadercross) into the
+// fragment shader for both, and an empty fragment shader for the depth-only shadow pass (which
+// reuses the vertex shaders with the light's view-projection).  Compiled offline by src/shaders/build.sh (SDL_shadercross) into the
 // SPIR-V, DXIL and MSL blobs in sceneShaders.h; the engine build never compiles shaders.
 //
 // Resource bindings follow SDL_GPU's HLSL convention:  vertex uniforms in space1, fragment
@@ -7,6 +8,11 @@
 
 #define MAX_LIGHTS  8
 #define MAX_JOINTS  128
+#define MAX_SHADOWS 8
+
+#define SHADOW_NONE 0
+#define SHADOW_MAP  1
+#define SHADOW_CUBE 2
 
 #define LIGHT_DIRECTIONAL 0
 #define LIGHT_POINT       1
@@ -81,7 +87,7 @@ struct Light {
 	float4 positionType;     // xyz position (point, spot) or unused; w = type
 	float4 directionRange;   // xyz direction the light shines in (directional, spot); w = range (0 = infinite)
 	float4 color;            // rgb already multiplied by intensity
-	float4 cone;             // x = cos(inner), y = cos(outer)
+	float4 cone;             // x = cos(inner), y = cos(outer), z = shadow slot + 1 (0 = casts none)
 };
 
 cbuffer FragmentUniforms : register(b0, space3) {
@@ -91,11 +97,102 @@ cbuffer FragmentUniforms : register(b0, space3) {
 	float4   emissive;
 	float4   material;       // x = metallic, y = roughness, z = unlit (1/0), w = textured (1/0)
 	float4   counts;         // x = light count
+	float4   shadowParams;   // x = depth bias, y = 1 / map size, z = shadow count
+	float4x4 shadowMatrix[MAX_SHADOWS];   // Light view-projection per slot (maps)
+	float4   shadowInfo[MAX_SHADOWS];     // x = SHADOW_MAP / SHADOW_CUBE, y = layer, z = near, w = far (cubes)
 	Light    lights[MAX_LIGHTS];
 };
 
 Texture2D<float4>     baseTexture   : register(t0, space2);
 SamplerState          baseSampler   : register(s0, space2);
+Texture2DArray<float> shadowMaps    : register(t1, space2);   // A layer per directional or spot shadow, six per point light
+SamplerState          shadowSampler : register(s1, space2);
+
+
+// How lit a point is by a directional or spot light's shadow map:  its depth from the light against
+// the map, averaged over a 3x3 block of texels so edges soften.  Points outside the map are lit.
+float shadowFactorMap(int slot, float3 worldPosition, float3 normal, float3 toLight) {
+	float4 lightSpace = mul(shadowMatrix[slot], float4(worldPosition + normal * shadowParams.x * 8.0, 1.0));
+	float  layer      = shadowInfo[slot].y;
+	float2 uv;
+	float  depth;
+	float  texel = shadowParams.y;
+	float  lit;
+
+	lightSpace.xyz /= lightSpace.w;
+	uv    = float2(lightSpace.x * 0.5 + 0.5, 0.5 - lightSpace.y * 0.5);
+	depth = lightSpace.z - shadowParams.x * (1.0 + 2.0 * (1.0 - saturate(dot(normal, toLight))));
+	if ((uv.x < 0.0) || (uv.x > 1.0) || (uv.y < 0.0) || (uv.y > 1.0) || (depth > 1.0)) {
+		return 1.0;
+	}
+	// Written out because DXC crashes on loops inside a function called from a loop.
+	lit  = (shadowMaps.Sample(shadowSampler, float3(uv + float2(-texel, -texel), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(0.0, -texel), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(texel, -texel), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(-texel, 0.0), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv, layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(texel, 0.0), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(-texel, texel), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(0.0, texel), layer)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(texel, texel), layer)).r >= depth) ? 1.0 : 0.0;
+	return lit / 9.0;
+}
+
+
+// How lit a point is by a point light's six shadow faces.  The face is picked by the direction's
+// major axis and looked up with the same frame the engine rendered it with (forward, up, right =
+// forward x up, a 90 degree perspective), so no cube-map convention is involved; the point's own
+// projected depth is compared over five taps.
+float shadowFactorCube(int slot, float3 worldPosition, float3 normal, float3 lightPosition) {
+	float  base  = shadowInfo[slot].y;
+	float  near  = shadowInfo[slot].z;
+	float  far   = shadowInfo[slot].w;
+	float3 dir   = (worldPosition + normal * shadowParams.x * 8.0) - lightPosition;
+	float3 ad    = abs(dir);
+	float  texel = shadowParams.y;
+	float  ma;
+	float  face;
+	float3 forward;
+	float3 up;
+	float3 right;
+	float2 uv;
+	float  depth;
+	float  lit;
+
+	if ((ad.x >= ad.y) && (ad.x >= ad.z)) {
+		ma      = ad.x;
+		face    = (dir.x > 0.0) ? 0.0 : 1.0;
+		forward = float3((dir.x > 0.0) ? 1.0 : -1.0, 0.0, 0.0);
+		up      = float3(0.0, 1.0, 0.0);
+	} else if (ad.y >= ad.z) {
+		ma      = ad.y;
+		face    = (dir.y > 0.0) ? 2.0 : 3.0;
+		forward = float3(0.0, (dir.y > 0.0) ? 1.0 : -1.0, 0.0);
+		up      = float3(0.0, 0.0, (dir.y > 0.0) ? -1.0 : 1.0);
+	} else {
+		ma      = ad.z;
+		face    = (dir.z > 0.0) ? 4.0 : 5.0;
+		forward = float3(0.0, 0.0, (dir.z > 0.0) ? 1.0 : -1.0);
+		up      = float3(0.0, 1.0, 0.0);
+	}
+	right = cross(forward, up);
+	uv    = float2(0.5 + 0.5 * dot(dir, right) / ma, 0.5 - 0.5 * dot(dir, up) / ma);
+	depth = far / (far - near) - near * far / ((far - near) * ma) - shadowParams.x * 2.0;
+	if (depth > 1.0) {
+		return 1.0;
+	}
+	lit  = (shadowMaps.Sample(shadowSampler, float3(uv, base + face)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(texel, 0.0), base + face)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv - float2(texel, 0.0), base + face)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv + float2(0.0, texel), base + face)).r >= depth) ? 1.0 : 0.0;
+	lit += (shadowMaps.Sample(shadowSampler, float3(uv - float2(0.0, texel), base + face)).r >= depth) ? 1.0 : 0.0;
+	return lit / 5.0;
+}
+
+
+// The shadow pass draws depth only.
+void depthMain(VertexOutput input) {
+}
 
 
 // Named fragmentMain because "fragment" is a keyword in Metal and the MSL entry point keeps this name.
@@ -157,6 +254,15 @@ float4 fragmentMain(VertexOutput input) : SV_Target {
 				attenuation *= smoothstep(lights[x].cone.y, lights[x].cone.x, cosAngle);
 			}
 		}
+		if (lights[x].cone.z > 0.5) {
+			int slot = (int)lights[x].cone.z - 1;
+
+			if (shadowInfo[slot].x == SHADOW_CUBE) {
+				attenuation *= shadowFactorCube(slot, input.worldPosition, normal, lights[x].positionType.xyz);
+			} else {
+				attenuation *= shadowFactorMap(slot, input.worldPosition, normal, toLight);
+			}
+		}
 		diffuse  = saturate(dot(normal, toLight));
 		halfway  = normalize(toLight + view);
 		specular = pow(saturate(dot(normal, halfway)), shininess) * (1.0 - roughness * 0.5);

src/singe.c

@@ -573,6 +573,7 @@ static int32_t      apiLightSetColor(lua_State *L);
 static int32_t      apiLightSetCone(lua_State *L);
 static int32_t      apiLightSetIntensity(lua_State *L);
 static int32_t      apiLightSetRange(lua_State *L);
+static int32_t      apiLightSetShadow(lua_State *L);
 static int32_t      apiMaterialDelete(lua_State *L);
 static int32_t      apiMaterialNew(lua_State *L);
 static int32_t      apiMaterialSetBlend(lua_State *L);
@@ -640,6 +641,7 @@ static int32_t      apiSceneProject(lua_State *L);
 static int32_t      apiSceneSetAmbient(lua_State *L);
 static int32_t      apiSceneSetAntialias(lua_State *L);
 static int32_t      apiSceneSetBackground(lua_State *L);
+static int32_t      apiSceneSetShadowSize(lua_State *L);
 static int32_t      apiSceneUnproject(lua_State *L);
 static int32_t      apiScriptExecute(lua_State *L);
 static int32_t      apiScriptPush(lua_State *L);
@@ -3401,6 +3403,19 @@ static int32_t apiLightSetRange(lua_State *L) {
 }
 
 
+// lightSetShadow(node, bool):  this light casts shadows (a point light's cost six passes)
+static int32_t apiLightSetShadow(lua_State *L) {
+	int32_t node;
+
+	_argCheck(L, "lightSetShadow", 2, 2);
+	node = _argNode(L, "lightSetShadow", 1);
+	if (!lightSetShadow(node, _argBoolean(L, "lightSetShadow", 2))) {
+		_luaDie(L, "lightSetShadow", "Node %d is not a light.", node);
+	}
+	return 0;
+}
+
+
 // materialDelete(material)
 static int32_t apiMaterialDelete(lua_State *L) {
 	_argCheck(L, "materialDelete", 1, 1);
@@ -4491,6 +4506,14 @@ static int32_t apiSceneSetBackground(lua_State *L) {
 }
 
 
+// sceneSetShadowSize(size):  shadow map texels per side, 256 to 4096 (default 1024)
+static int32_t apiSceneSetShadowSize(lua_State *L) {
+	_argCheck(L, "sceneSetShadowSize", 1, 1);
+	sceneSetShadowSize(_argInteger(L, "sceneSetShadowSize", 1));
+	return 0;
+}
+
+
 // x, y, z = sceneUnproject(sx, sy, distance):  the world point that far along the ray through an overlay point
 static int32_t apiSceneUnproject(lua_State *L) {
 	_argCheck(L, "sceneUnproject", 3, 3);
@@ -5973,6 +5996,7 @@ void singe(SDL_Window *window, SDL_Renderer *renderer, SDL_GPUDevice *device, Co
 	lua_register(_global.luaContext, "lightSetCone",                apiLightSetCone);                 // 3.00
 	lua_register(_global.luaContext, "lightSetIntensity",           apiLightSetIntensity);            // 3.00
 	lua_register(_global.luaContext, "lightSetRange",               apiLightSetRange);                // 3.00
+	lua_register(_global.luaContext, "lightSetShadow",              apiLightSetShadow);               // 3.00
 	lua_register(_global.luaContext, "materialDelete",              apiMaterialDelete);               // 3.00
 	lua_register(_global.luaContext, "materialNew",                 apiMaterialNew);                  // 3.00
 	lua_register(_global.luaContext, "materialSetBlend",            apiMaterialSetBlend);             // 3.00
@@ -6041,6 +6065,7 @@ void singe(SDL_Window *window, SDL_Renderer *renderer, SDL_GPUDevice *device, Co
 	lua_register(_global.luaContext, "sceneSetAmbient",             apiSceneSetAmbient);              // 3.00
 	lua_register(_global.luaContext, "sceneSetAntialias",           apiSceneSetAntialias);            // 3.00
 	lua_register(_global.luaContext, "sceneSetBackground",          apiSceneSetBackground);           // 3.00
+	lua_register(_global.luaContext, "sceneSetShadowSize",          apiSceneSetShadowSize);           // 3.00
 	lua_register(_global.luaContext, "sceneUnproject",              apiSceneUnproject);               // 3.00
 	lua_register(_global.luaContext, "scriptExecute",               apiScriptExecute);                // 2.00
 	lua_register(_global.luaContext, "scriptPush",                  apiScriptPush);                   // 2.00