Advanced Metal Shader Optimization

More documents

Recommendations

Info

Address Spaces Example: projection matrix Variable Number of items Amount of reuse Address space transform one all constant vertex float4 transformedVertex(uint vid [[ vertex_id ]]), const device float4 *positions [[ buffer(0) ]], constant matrix_float4x4 &transform [[ buffer(1) ]]) { return transform * positions[vid]; }
Address Spaces Example: skinning matrices Variable Number of items Amount of reuse Address space skinningMatrices fixed number of bones all vertices using bone constant struct SkinningMatrices { matrix_float4x4 position_transforms[MAXBONES]; }; vertex float4 skinnedVertex(uint vid [[ vertex_id ]]), const device Vertex *vertices [[ buffer(0) ]], constant SkinningMatrices &skinningMatrices [[ buffer(1) ]] { … for (ushort i = 0; i < NBONES; ++i) { skinnedPosition += (skinningMatrices.position_transforms[vertices[vid].boneIndices[i]] * … }
Page 1 and 2: Graphics and Games #WWDC16 Advanced
Page 3 and 4: Metal at WWDC This Year A look at t
Page 5 and 6: Optimizing Shaders An overview Ther
Page 7 and 8: Metal Pipeline Buffer Buffer Textur
Page 9 and 10: Shader Performance Fundamentals
Page 11 and 12: Address Spaces Comparison GPUs have
Page 13 and 14: Address Spaces Constant memory Read
Page 15 and 16: Address Spaces Example: vertex data
Page 17: Address Spaces Example: projection
Page 21 and 22: Address Spaces Example: per-instanc
Page 23 and 24: Buffer Preloading Buffer loads can
Page 25 and 26: Constant Buffer Preloading Direct l
Page 27 and 28: Constant Buffer Preloading Use cons
Page 33 and 34: Constant Buffer Preloading A practi
Page 39 and 40: Vertex Buffer Preloading Fixed-func
Page 41 and 42: Fragment Function Resource Writes N
Page 43 and 44: Compute Kernel Organization Per-thr
Page 45 and 46: kernel void sobel_1_1(/* ... */ ush
Page 53 and 54: continue to pixel 2... dstCoord.x++
Page 55 and 56: continue to pixel 2... dstCoord.x++
Page 57 and 58: Compute Kernel Organization NEW Con
Page 59 and 60: Shader Performance Fundamentals Con
Page 65 and 66: GPU Architecture Focus on the bottl
Page 67 and 68: Optimization Opportunities Data typ
Page 69 and 70:
Data Types Using half and short ari
Page 71 and 72:
Page 73 and 74:
Page 75 and 76:
Page 77 and 78:
Arithmetic Built-ins Use built-ins
Page 79 and 80:
Arithmetic A8 and later GPUs are sc
Page 81 and 82:
Arithmetic ILP (Instruction Level P
Page 83 and 84:
Arithmetic A8 and later GPUs have v
Page 85 and 86:
Arithmetic A8 and later GPUs have v
Page 87 and 88:
Arithmetic Integer divisions Avoid
Page 89 and 90:
Arithmetic Fast-math If you absolut
Page 91 and 92:
Control Flow Control flow uniform a
Page 93 and 94:
Memory Access Stack access Avoid dy
Page 95 and 96:
Memory Access Stack access Avoid dy
Page 97 and 98:
Memory Access Loads and stores One
Page 99 and 100:
Memory Access Loads and stores One
Page 101 and 102:
Memory Access Loads and stores Use
Page 103 and 104:
Memory Access Loads and stores Use
Page 105 and 106:
Latency/Occupancy The more latency,
Page 107 and 108:
Memory Access Latency-hiding: False
Page 109 and 110:
Memory Access Latency-hiding: False
Page 111 and 112:
Summary Pick correct address spaces
Page 113 and 114:
Summary Pick correct address spaces
Page 115 and 116:
More Information https://developer.
Page 117:
Labs Xcode Open Hours Metal Lab Xco
show all

Advanced Metal Shader Optimization

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?