Forge Cascaded Shadow Maps

Add directional-light cascaded shadow maps (CSM) with 3x3 PCF soft shadows to an SDL3 GPU application. Based on Lesson 15.

When to use

• You need directional light shadows covering a large area
• Single shadow maps lack resolution for your scene
• You want soft shadow edges without expensive techniques

Key API calls

• SDL_CreateGPUTexture — shadow map textures (D32_FLOAT, DEPTH_STENCIL_TARGET | SAMPLER)
• SDL_CreateGPUSampler — NEAREST filter, CLAMP_TO_EDGE for shadow map sampling
• SDL_CreateGPUGraphicsPipeline — shadow pipeline (depth-only, front-face cull, depth bias)
• SDL_BeginGPURenderPass / SDL_EndGPURenderPass — one pass per cascade (depth only)
• SDL_BindGPUFragmentSamplers — bind shadow maps + sampler for scene pass
• SDL_PushGPUVertexUniformData / SDL_PushGPUFragmentUniformData — per-draw uniforms

Correct order

1. Create shadow map textures + sampler (in SDL_AppInit)
2. Create shadow pipeline — depth-only, front-face culling, depth bias
3. Create scene pipeline — binds shadow maps as fragment samplers
4. Each frame: a. Compute cascade split depths (logarithmic-linear blend) b. Compute light VP matrix per cascade (frustum corners → light-space AABB) c. Shadow pass: render all casters into each cascade's depth map d. Scene pass: render with shadow maps bound, fragment shader selects cascade by view-space depth

Key concepts

1. Shadow map textures need DEPTH_STENCIL_TARGET | SAMPLER usage flags
2. Depth-only render pass: no color targets, only depth attachment
3. Cascade splits: logarithmic-linear blend for practical depth ranges
4. Light VP matrices: tight orthographic projection from frustum corners
5. PCF sampling: 3x3 grid averages shadow/lit for soft edges
6. Store op MUST be STORE (not DONT_CARE) for shadow map depth passes

Pipeline setup

Shadow pipeline (depth-only)

/* No color targets — depth only */
pipe.target_info.num_color_targets        = 0;
pipe.target_info.has_depth_stencil_target = true;
pipe.target_info.depth_stencil_format     = SDL_GPU_TEXTUREFORMAT_D32_FLOAT;

/* Front-face culling reduces peter-panning */
pipe.rasterizer_state.cull_mode = SDL_GPU_CULLMODE_FRONT;

/* Depth bias reduces shadow acne */
pipe.rasterizer_state.depth_bias_constant_factor = 2;
pipe.rasterizer_state.depth_bias_slope_factor    = 2.0f;

Shadow map texture

SDL_GPUTextureCreateInfo info;
SDL_zero(info);
info.type   = SDL_GPU_TEXTURETYPE_2D;
info.format = SDL_GPU_TEXTUREFORMAT_D32_FLOAT;
info.usage  = SDL_GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET |
              SDL_GPU_TEXTUREUSAGE_SAMPLER;  /* key: both flags */
info.width  = 2048;
info.height = 2048;
info.layer_count_or_depth = 1;
info.num_levels           = 1;

Shadow sampler

SDL_GPUSamplerCreateInfo smp;
SDL_zero(smp);
smp.min_filter     = SDL_GPU_FILTER_NEAREST;
smp.mag_filter     = SDL_GPU_FILTER_NEAREST;
smp.address_mode_u = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
smp.address_mode_v = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;

Cascade split algorithm

/* Lengyel's logarithmic-linear blend */
for (int i = 0; i < NUM_CASCADES; i++) {
    float p = (float)(i + 1) / (float)NUM_CASCADES;
    float log_split = near * powf(far / near, p);
    float lin_split = near + (far - near) * p;
    splits[i] = lambda * log_split + (1.0f - lambda) * lin_split;
}

Light VP computation per cascade

1. Unproject 8 NDC corners via inverse camera VP
2. Interpolate near/far to get cascade slice corners
3. Compute cascade center, build light view matrix
4. Transform corners to light space, find tight AABB
5. Build orthographic projection from AABB

mat4 light_view = mat4_look_at(light_pos, center, up);
/* Transform corners to light space → compute AABB */
mat4 light_proj = mat4_orthographic(min_x, max_x, min_y, max_y,
                                     -max_z, -min_z);
mat4 light_vp = mat4_multiply(light_proj, light_view);

Shadow pass render loop

for (int ci = 0; ci < NUM_CASCADES; ci++) {
    SDL_GPUDepthStencilTargetInfo depth;
    SDL_zero(depth);
    depth.texture  = shadow_maps[ci];
    depth.load_op  = SDL_GPU_LOADOP_CLEAR;
    depth.store_op = SDL_GPU_STOREOP_STORE;  /* MUST store */
    depth.clear_depth = 1.0f;

    SDL_GPURenderPass *pass = SDL_BeginGPURenderPass(cmd, NULL, 0, &depth);
    SDL_BindGPUGraphicsPipeline(pass, shadow_pipeline);

    /* Draw all shadow casters with light_vp[ci] * model */
    /* ... */

    SDL_EndGPURenderPass(pass);
}

HLSL shadow sampling (fragment shader)

float sample_shadow_pcf(Texture2D shadow_map, SamplerState smp,
                         float2 shadow_uv, float current_depth) {
    float shadow = 0.0;
    [unroll]
    for (int y = -1; y <= 1; y++) {
        [unroll]
        for (int x = -1; x <= 1; x++) {
            float2 offset = float2(x, y) * texel_size;
            float map_depth = shadow_map.Sample(smp, shadow_uv + offset).r;
            shadow += (map_depth >= current_depth - bias) ? 1.0 : 0.0;
        }
    }
    return shadow / 9.0;
}

Common mistakes

• Forgetting SAMPLER usage flag on shadow map texture — it will not be readable in fragment shaders without it
• Using DONT_CARE for shadow pass store_op — depth data is discarded and shadow maps will be empty
• Not flipping Y when converting NDC to UV: shadow_uv.y = 1.0 - shadow_uv.y
• Missing depth bias — causes shadow acne (moire self-shadowing pattern)
• Culling back faces in shadow pass — should cull FRONT faces instead to reduce peter-panning
• Not expanding Z range of the light AABB — shadow casters behind the frustum slice won't cast shadows into it

HLSL register map

Vertex uniform slot 0 → register(b0, space1)
Vertex uniform slot 1 → register(b1, space1)
Fragment sampler slot N → register(tN, space2) + register(sN, space2)
Fragment uniform slot 0 → register(b0, space3)

Reference

See: Lesson 15 — Cascaded Shadow Maps