2nd/10_Wiki/Topics/AI_and_ML/Depth Pre-Pass.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-depth-pre-pass	Depth Pre-Pass	10_Wiki/Topics	verified	self	Z-prepass depth prepass early-Z Z-only pass	none	A	0.9	applied	graphics rendering gpu optimization depth-buffer overdraw		2026-05-10	pending	language framework HLSL / GLSL / Shader Unreal / Unity / Custom

Depth Pre-Pass

매 한 줄

"매 main shading 전 의 depth-only render". 매 overdraw 의 reduce — 매 expensive pixel shader 의 hidden surface 의 skip. 매 modern GPU 의 early-Z + Hi-Z + 매 deferred / forward+. 매 cost: 매 vertex 2x.

매 핵심

매 motivation

• Overdraw: 매 same pixel 의 shade 여러 번.
• Expensive shader: 매 PBR + IBL + many light → 매 pixel cost ↑.
• Solution: 매 depth 만 의 first → 매 main pass 의 occluded fragment 의 reject.

매 mechanism

1. Pass 1: 매 depth-only (vertex + null pixel shader).
2. Pass 2: 매 full shading + EQUAL depth test.
3. 매 GPU 의 early-Z 의 pixel shader 전 의 reject.

매 trade-off

• Win: 매 expensive pixel shader 의 occluded fragment 의 skip.
• Loss: 매 vertex stage 2x.
• Net: 매 shader complexity 의 high 의 win.

매 modern variant

• Hi-Z: 매 hierarchical depth 의 GPU 의 cull tile.
• Forward+: 매 light culling + Z-prepass.
• Deferred: 매 G-buffer 의 prepass-like.
• Visibility buffer: 매 modern alternative (Unreal 5 Nanite).

매 응용

1. Open world: 매 dense vegetation overdraw.
2. Particle: 매 alpha sort cost.
3. PBR-heavy: 매 expensive shader.
4. VR: 매 fill rate critical.
5. Mobile: 매 tile-based 의 different approach (defer to TBDR).

💻 패턴

Z-prepass (Unreal-style HLSL)

// Pass 1 — depth only
struct VSInput { float3 pos : POSITION; };

float4 VS_DepthOnly(VSInput v) : SV_POSITION {
    return mul(float4(v.pos, 1.0), MVP);
}
// 매 no pixel shader (or null)

// Pass 2 — full shading with EQUAL test
DepthStencilState : DepthFunc = EQUAL; DepthWrite = OFF;

float4 PS_Main(VSOutput v) : SV_TARGET {
    return PBR_Shade(v);  // 매 expensive
}

OpenGL setup

// Pass 1
glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
glDepthMask(GL_TRUE);
glDepthFunc(GL_LESS);
DrawScene(depth_only_program);

// Pass 2
glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
glDepthMask(GL_FALSE);
glDepthFunc(GL_EQUAL);
DrawScene(main_program);

Vulkan render pass

VkAttachmentDescription depthAttach = {
    .format = VK_FORMAT_D32_SFLOAT,
    .loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR,
    .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
    .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
    .finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL,
};
// Subpass 1: depth only
// Subpass 2: shading w/ EQUAL test, read-only depth

Cost-benefit decision

bool should_use_depth_prepass(SceneStats stats) {
    // 매 high overdraw + expensive shader → 매 win
    float avg_overdraw = stats.pixels_shaded / stats.unique_pixels;
    float shader_cost = stats.shader_alu_count;
    return avg_overdraw > 2.0 && shader_cost > 100;
}

Hi-Z chain (compute)

// 매 each level 의 max of 4 source pixels
[numthreads(8, 8, 1)]
void GenerateHiZ(uint3 tid : SV_DispatchThreadID) {
    float4 d;
    d.x = SrcDepth[tid.xy * 2 + uint2(0, 0)];
    d.y = SrcDepth[tid.xy * 2 + uint2(1, 0)];
    d.z = SrcDepth[tid.xy * 2 + uint2(0, 1)];
    d.w = SrcDepth[tid.xy * 2 + uint2(1, 1)];
    DstDepth[tid.xy] = max(max(d.x, d.y), max(d.z, d.w));
}

Occlusion query (validation)

GLuint q;
glGenQueries(1, &q);
glBeginQuery(GL_SAMPLES_PASSED, q);
DrawObject(obj);
glEndQuery(GL_SAMPLES_PASSED);
GLuint samples;
glGetQueryObjectuiv(q, GL_QUERY_RESULT, &samples);
// 매 samples == 0 → fully occluded

Early-Z disqualification

// 매 pixel shader 의 discard / depth-write → 매 early-Z 의 disabled
float4 PS_Bad(VSOut v) : SV_TARGET {
    if (alpha < 0.5) discard;  // ❌ early-Z 의 break
    return shade(v);
}

// 매 ✅ early-Z friendly: 매 alpha test 의 discard 의 separate pass
[earlydepthstencil]
float4 PS_Good(VSOut v) : SV_TARGET {
    return shade(v);
}

TBDR mobile (no prepass needed)

// 매 mobile (Adreno, Mali, Apple) 의 tile-based deferred
// 매 already 의 hidden surface 의 reject before pixel shader
// 매 explicit Z-prepass 의 redundant

매 결정 기준

상황	Approach
Desktop + heavy PBR	Z-prepass
Mobile (TBDR)	NO (HW already does it)
Forward+	Z-prepass + light cull
Deferred	G-buffer pass = prepass
Particle / transparent	After opaque (no prepass)
Modern Unreal 5	Visibility buffer / Nanite

기본값: 매 desktop + 매 heavy shader → Z-prepass. 매 mobile → skip. 매 modern engine → visibility buffer.

🔗 Graph

• 부모: Rendering-Pipeline
• 변형: Depth Pre-Pass · Hi-Z
• 응용: Deferred-Rendering
• Adjacent: Overdraw · Compute Shader (WebGPU)

🤖 LLM 활용

언제: 매 PBR-heavy desktop. 매 dense scene. 매 fill-rate-bound. 언제 X: 매 mobile TBDR. 매 vertex-bound. 매 simple shader.

❌ 안티패턴

• Always prepass: 매 vertex-bound 의 lose.
• Discard in main: 매 early-Z 의 break.
• Mobile prepass: 매 redundant + cost.
• No EQUAL test in main: 매 prepass 의 useless.
• Animated geometry mismatch: 매 prepass 와 main 의 different vertex.

🧪 검증 / 중복

• Verified (Real-Time Rendering 4ed, Unreal docs, Vulkan spec).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-04-20	Auto-reinforced
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Z-prepass + 매 HLSL / GL / Vulkan / Hi-Z code