RSX™ Best Practices
- 2. RSX™ Best Practices Mark Cerny, Cerny Games David Simpson, Naughty Dog Jon Olick, Naughty Dog
- 3. RSX™ Best Practices About libgcm Using the SPUs with the RSX™ Brief overview of GCM Replay
- 4. December 7th, 2004 Sony Computer Entertainment and NVidia announce joint development of RSX™
- 5. Vertex & Index Buffers Textures Fragment Programs
- 6. PLAYSTATION®2 Command List Command List Vertex Table Index Table
- 8. PLAYSTATION®3 Command Buffer Render Targets Textures Vertex & Fragment Programs Vertex & Index Buffers Command Buffer
- 10. More State in OpenGL glClearDepth(depth); glClearStencil(s); glClear(GL_DEPTH | GL_STENCIL) glClearDepth(depth); glClearStencil(s); glClear(GL_DEPTH | GL_STENCIL)
- 12. Goals with libgcm Support multiple buckets
- 13. Goals with libgcm Support multiple buckets Remove state
- 14. Goals with libgcm Support multiple buckets Remove state glClearDepth and glClearStencil become a single function
- 17. libgcm Context Structure Write Location Start End
- 18. libgcm Context Structure Write Location Start End Callback generated when out of space
- 21. Heap Virtual Buckets Write Location, Bucket 2 Write Location, Bucket 1
- 22. Heap Virtual Buckets Write Location, Bucket 1 Write Location, Bucket 2
- 23. Heap Virtual Buckets New allocation for bucket 1, connected by jump Write Location, Bucket 2 Write Location, Bucket 1
- 24. Heap Virtual Buckets New allocation for bucket 2, connected by jump Write Location, Bucket 2 Write Location, Bucket 1
- 25. Context Types Space Checking No Space Checking
- 26. Set Alpha Blend Space Checking 5.0x No Space Checking 1.1x
- 27. Setup Texture for Shader Space Checking 1.8x No Space Checking 1.8x
- 28. Space Checking Set up Alpha Blend 1.1x Set up Texture for Shader 1.05x
- 29. Patch Static Command Buffer Disable Draw Set Shader Constants Set Shader Constants Disable Draw Set Shader Constants Disable Draw
- 30. Concatenate Static Command Buffers Set Shader Constants Set Shader Constants Set Shader Constants Set Shader Constants Set Shader Constants
- 31. Using the RSX™ with the SPUs SPU 0 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 RSX™
- 32. Using the RSX™ with the SPUs SPUs can be used to supercharge vertex processing on the RSX™ SPUs can perform triangle and mesh operations that cannot be performed on the RSX™
- 33. Geometry Processing Pipeline Runs on SPUs Modular Need only to use some pieces Outputs index and vertex data which is directly read by the RSX™ Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 34. Geometry Processing Pipeline SPU processes one vertex set at a time One or more vertex sets are generated per mesh in an offline tools processing step called partitioning Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 35. The RSX™ can process vertices in large chunks
- 36. But a 50,000 vertex object won’t fit in an SPU! SPU
- 37. The object needs to be partitioned into smaller pieces, called vertex sets SPU
- 38. Culling is much better with smaller pieces too SPU
- 39. On the PLAYSTATION®2 we used vertex sets with about 64 vertices Many repeated vertices Data increase of about 30%
- 40. An SPU can handle vertex sets with between 500-1500 vertices Still some repeated vertices Data increase of about 7-10% Vertex data is ultimately smaller due to increased compression
- 41. Decompression SPUs free to use any type of compressed data – not restricted to 8, 16, 32 bit or the like Vertex data is decompressed into full floats, as they are easiest for the SPU to use Triangle index data can also be decompressed at this time Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 42. N-Bit Stream Decompression Each vertex attribute is an N-bit stream Each component of that attribute has its own number of bits, integer offset, scale, and bias Each component is decompressed with the following equation: out = float(in + intOffset) * scale + bias Scale and bias need to be constant across an entire object to prevent cracks The number of bits and integer offset need not be
- 43. Integer Offset Example The total range of this object is 21 units Requires 5 bits The range of the first list is only 15 units Requires only 4 bits The range of the second list is 8 units When intOffset is set to 13, entries in the second list require only 3 bits Object List 1 1 5 12 0 8 3 14 9 List 2 17 14 20 16 13 19 18
- 44. Blend Shapes Not really possible to do on a GPU SPU can blend any number of shapes and any number of vertex attributes Large data savings Store only deltas Use highly compressed data formats, like N-bit compression Only store data for changing vertices Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 45. Blend Shape Use in MLB 07: The Show
- 46. Skinning Huge offload of vertex processing from the RSX™ No need to set large number of vertex program constants with matrix data SPU can handle vertices with an arbitrary number of influences Number of influences can vary across the mesh, resulting in data and computational savings Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 47. LOD Systems Reduces processing time as an object moves into the distance Many LOD systems are not a good match for a GPU Often operate upon an entire mesh at a time Good match for the SPUs Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 48. Discrete Progressive Mesh Smoothly reduces the triangle count as a model moves into the distance With discrete progressive mesh, the LOD calculation is done once for an entire object An entire object is processed at once by the tools to avoid cracks between vertex sets Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 49. At an LOD there are two types of vertices LOD = 0.0 Parent Vertex Child Vertex
- 50. As the LOD level decreases, the children “slide” towards their parents LOD = 0.2 Parent Vertex Child Vertex
- 51. The children continue to move towards their parents LOD = 0.7 Parent Vertex Child Vertex
- 52. At the next integral LOD, all child vertices disappear as do the triangles LOD = 1.0 Parent Vertex Child Vertex
- 53. Vertices are arranged from lowest LOD to highest LOD At LOD 0 all vertices are needed. At LOD 1, child vertices from LOD 0 are no longer needed At LOD 2, child vertices from LOD 1 are also removed This saves bandwidth to the SPUsLOD 0 LOD 1 LOD 2 Vertex Table
- 54. Every LOD has its own index table of triangles and parent table The parent table contains an index to the parent for every child vertex Indexes for LOD 0 Indexes for LOD 1 Indexes for LOD 2 Parent Table for LOD 2 Parent Table for LOD 1 Parent Table for LOD 0
- 55. LODs are arranged in LOD groups to avoid small vertex sets LOD 0 LOD 1 LOD 2 LOD 0 LOD 1 LOD 2 LOD Group 0 LOD 0 LOD 1 LOD 2 LOD 3 LOD 4 LOD 5 LOD Group 1
- 57. Continuous Progressive Mesh Like discrete progressive mesh, child vertices move smoothly toward their parents However, the LOD is calculated for each vertex instead of just once for the object Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 58. Vertex set about to undergo continuous progressive mesh Parent Vertex Child Vertex, LOD 1 Child Vertex, LOD 0
- 59. A single vertex set can straddle several LOD ranges LOD = 1.0 Parent Vertex Child Vertex, LOD 1 Child Vertex, LOD 0 LOD = 0.0
- 60. Vertices move depending on their distance Parent Vertex Child Vertex, LOD 1 Child Vertex, LOD 0 LOD = 1.0 LOD = 0.0
- 62. Stencil Shadows Edge Table T14 T15 V1 T6 T23 V0 T10 T18 V2 V0T10T2 V0T8T0 V1T6T5 V2T26T5 V0T13T4 V2T17T15 V1T25T11 T14 T15 V1
- 63. Also need adjoining triangles and vertices from neighboring vertex sets Edge Table T14 T15 V1 T6 T23 V0 T10 T18 V2 V0T10T2 V0T8T0 V1T6T5 V2T26T5 V0T13T4 V2T17T15 V1T25T11
- 64. Find the profile edges and generate a new vertex table of extruded edges Extruded Edges T14 T15 V1 T6 T23 V0 T10 T18 V2 V0T10T2 V0T8T0 V1T6T5 V2T26T5 V0T13T4 V2T17T15 V1T25T11 Edge Table
- 65. Output the new vertex data and draw commands to a shadow context Command Context for Light 0 Shadow Draw 13 Shadow Draw 19 Shadow Draw 18 SPU
- 66. May as well do multiple lights at the same time Command Context for Light 0 Shadow Draw 13 Shadow Draw 19 Shadow Draw 18 SPU Shadow Draw 11 Shadow Draw 19 Shadow Draw 12 Shadow Draw 13 Shadow Draw 19 Shadow Draw 18 Shadow Draw 15 Shadow Draw 19 Shadow Draw 16 Command Context for Light 1 Command Context for Light 2 Command Context for Light 3
- 67. Normal and Tangent Calculation Typically having normals and tangents included in the vertex data is a good thing However, some operations can move the positions so much that the included normals and tangents are no longer correct Solution: Recalculate the normals and tangents on the SPU!
- 68. Blend shapes can move the positions quite a lot!
- 69. Recalculate the normals! Like stencil shadows, calculating normals and tangents requires information about adjoining triangles and vertices from neighboring vertex sets Only worth the cost in limited situations
- 70. Triangle Culling Many triangles in a scene will ultimately have no renderable area Culling these triangles on the SPU removes the burden of the RSX™ processing triangles which do not contribute to the final image This leaves the RSX™ with more time to process relevant triangles Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 71. What types of triangles can be culled?
- 72. Off screen triangles can be culled
- 73. Back facing triangles can be culled, but be sure to use error bars
- 74. Degenerate triangles can be culled
- 75. Some triangles are very small
- 76. Some triangles are so small that they do not cover a pixel center
- 77. These triangles can be culled
- 79. But these triangles can still be culled
- 80. The SPU starts with the input triangle index table Tri 0 Tri 1 Tri 2 Tri 3 Tri 4 Tri 5 Tri 6 Tri 7 Tri 8 Tri 9 Tri 10 Tri 11 Tri 12 Tri 13 Tri 14 Original Index Table
- 81. The culling algorithm determines which triangles are to be kept Tri 0 Tri 1 Tri 2 Tri 3 Tri 4 Tri 5 Tri 6 Tri 7 Tri 8 Tri 9 Tri 10 Tri 11 Tri 12 Tri 13 Tri 14 Original Index Table
- 82. And a new index table is created from these triangles Tri 0 Tri 1 Tri 2 Tri 3 Tri 4 Tri 5 Tri 6 Tri 7 Tri 8 Tri 9 Tri 10 Tri 11 Tri 12 Tri 13 Tri 14 Original Index Table Tri 1 Tri 4 Tri 6 Tri 7 Tri 11 Tri 14 Culled Index Table Culling can remove 60-70% of all triangles!
- 83. Vertex Culling After triangles are culled, some vertices are no longer used in any triangle These vertices can be removed from the vertex table This is done by first building a vertex renaming table which contains the new vertex index for each vertex Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 84. Start with an empty vertex renaming table 4 5 6 7 8 9 10 11 12 13 14 15 3 2 1 0 Index Table Vertex Table Renaming Table -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 0 2 5 1 4 10 1 8 13 2 5 7 5 7 8 7 8 9 8 10 13 10 13 14
- 85. Add a new index for each used vertex in the index table 4 5 6 7 8 9 10 11 12 13 14 15 3 2 1 00 2 5 1 4 10 1 8 13 2 5 7 5 7 8 7 8 9 8 10 13 10 13 14 Index Table Vertex Table Renaming Table 0 3 1 -1 4 2 -1 8 6 9 5 -1 -1 7 10 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 0 3 1 4 2 8 6 9 5 7 10
- 86. Using the renaming table, build a new vertex table with only used vertices 4 5 6 7 8 9 10 11 12 13 14 15 3 2 1 0 Vertex Table Renaming Table 0 3 1 -1 4 2 -1 8 6 9 5 -1 -1 7 10 -1 4 10 8 13 7 9 14 1 5 2 0 New Vertex TableIndex Table 0 2 5 1 4 10 1 8 13 2 5 7 5 7 8 7 8 9 8 10 13 10 13 14
- 87. Finally, replace the old indices in the index table with the new indices New Vertex TableIndex Table Renaming Table 0 3 1 -1 4 2 -1 8 6 9 5 -1 -1 7 10 -1 4 10 8 13 7 9 14 1 5 2 0 4 5 6 7 8 9 10 11 12 13 14 15 3 2 1 0 Vertex Table 0 2 5 1 4 10 1 8 13 2 5 7 5 7 8 7 8 9 8 10 13 10 13 14 0 1 2 3 4 5 3 6 7 1 2 8 5 7 8 8 6 9 6 5 7 5 7 10
- 88. Only minor performance gains on the RSX™, if any Removes about 30% of the vertex data Better use of the pre-transform cache, but not much else
- 89. Vertex Stream Combining A vertex stream is an interleaved set of vertex attributes, which is used natively by the RSX™ Fewer vertex streams results in better performance Easy to combine streams while compressing vertex attributes into RSX™ formats Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 90. Vertex attributes can be input into the SPUs in multiple streams Input Vertex Stream 0 Input Vertex Stream 1
- 91. The vertex streams are decompressed into tables of floats Input Vertex Stream 0 Input Vertex Stream 1 Float Tables
- 92. When done, the vertex attributes are compressed into one output stream Input Vertex Stream 0 Output Vertex Stream Float Tables Input Vertex Stream 1
- 93. Output Buffering Schemes Vertex and index data constructed by the SPUs is output from SPU local store Holes in the command buffer are patched with pointers to the vertex and index data as well as the draw commands Decompression Blend Shapes Skinning Progressive Mesh Culling Compression Output
- 94. Double Buffer Each buffer stores vertex and index data for an entire frame SPUs atomically access a mutex which is used to allocate memory from a buffer Easy synchronization with the RSX™ once a frame Uses lots of memory Vertex and Index Data for Frame 0 Vertex and Index Data for Frame 1
- 95. It is possible to completely fill a buffer Can use a callback to allocate new memory (which you may not have) Don’t draw geometry that doesn’t fit (difficult to pick which geometry not to draw) SPUData Vertex and Index Data
- 96. Double buffer requires an extra frame of lag in the rendering pipeline Build Jobs on PPU Process Jobs on SPU Render on RSX™ Scan Out Build Jobs on PPU Process Jobs on SPU Render on RSX™ Scan Out Build Jobs on PPU Process Jobs on SPU Render on RSX™ Scan Out
- 97. Single Buffer Uses only half the memory! Still possible to completely fill the buffer Vertex and Index Data for Single Frame
- 98. Single buffer uses a shorter rendering pipeline Vertex and index data is created just-in-time for the RSX™ Draw commands are inserted into the command buffer while the RSX™ is rendering Requires tight SPU↔RSX™ synchronization Build Jobs on PPU SPU Processing/ RSX™ Rendering Scan Out Build Jobs on PPU SPU Processing/ RSX™ Rendering Scan Out Build Jobs on PPU SPU Processing/ RSX™ Rendering Scan Out
- 99. Command Buffer Holes SPU processing requires some setup by the PPU Some job data is required for each vertex set Static portions of the command buffer are built on the PPU Static vertex attribute pointers Index table pointer and draw commands when not performing triangle culling on the SPU “Holes” are left for the dynamic portion built by the SPU Static 18 Other State State Hole 18 Hole 19 Hole 20 Hole 21 Hole 22 Static 19 Static 20 Static 21 Static 22 Command Buffer
- 100. Filling the Holes Dynamic portions are built on the SPU Vertex attribute pointers for any attributes output by the SPU Draw commands when performing triangle culling Commands necessary for ring buffer synchronization For brevity, we will not show the PPU generated commands going forward Static 18 Other State State Static 19 Static 20 Static 21 Static 22 Command Buffer Hole 18 Hole 19 Hole 20 Hole 21 Hole 22 Draw 18 Draw 19 Draw 20 Draw 21 Draw 22
- 101. RSX™ Put Pointer RSX™ has an internal “Put Pointer” Commands in the command buffer are executed up to the Put Pointer Commands after the Put Pointer are not executed Command Buffer Put Pointer
- 102. Put Pointer SPU↔RSX™ Synchronization Using the Put Pointer Command Buffer Output Buffer SPU Draw 17 Data 17
- 103. Put Pointer SPU outputs vertex and index data Command Buffer Output Buffer SPU Draw 17 Data 17 Data 18
- 104. Put Pointer SPU outputs vertex attribute pointers and draw commands Command Buffer Output Buffer SPU Draw 17 Data 17 Data 18Draw 18
- 105. Put Pointer SPU updates the Put Pointer Command Buffer Output Buffer SPU Draw 17 Data 17 Data 18Draw 18
- 106. Put Pointer But there are six SPUs, so who updates the Put Pointer? Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 21Draw 18 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 19 Draw 20 Draw 21 Draw 22 Draw 23 Data 19 Data 18 Data 20 Data 23 Data 22 Other
- 107. Put Pointer SPUs are asynchronous, so they can finish in any order! Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 21 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Other
- 108. Put Pointer So, the SPUs must synchronize with each other! Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 19 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Draw 19 Data 21 Other
- 109. The Put Pointer is updated only when ALL previous jobs are done… Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 18 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Draw 18 Data 21 Draw 19 Data 19 Put Pointer Other
- 110. But can only be moved to the end of this job’s draw commands Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 20 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Draw 20 Data 21 Draw 19 Data 18 Put Pointer Draw 18 Data 19 Other
- 111. Remember to guarantee progress! Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 23 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Draw 23 Data 21 Draw 19 Data 18 Put Pointer Draw 18 Data 19 Draw 20 Data 20 Other
- 112. The last job finished moves the Put Pointer to the end of the buffer Command Buffer Output Buffer SPU 0Draw 17 Data 17 Data 22 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Draw 21 Draw 22 Data 21 Draw 19 Data 18 Put Pointer Draw 18 Data 19 Draw 20 Data 20 Other Draw 23 Data 23
- 113. SPU↔RSX™ Synchronization Using Local Stalls Command Buffer Draw 17 Put Pointer Local Stall Other Local Stall Local Stall Local Stall Local Stall Local Stall Easier and faster than Put Pointer synchronization Place local stalls in the command buffer where necessary RSX™ will stop processing at a local stall until it is overwritten by new commands SPUs will generally stay ahead of the RSX™, so stalls rarely occur Local Stall
- 114. SPU will overwrite local stalls when it outputs a set of new commands No SPU↔SPU synchronization required! Please see the document regarding this technique on the PS3 Developer’s Support website for crucial details SPU Command Buffer Draw 17 Put Pointer Local Stall Other Local Stall Local Stall Local Stall Local Stall Local Stall Local Stall New Commands
- 115. Ring Buffers Small memory footprint Will not run out of memory Can stall the SPUs if buffers become full Objects need to be processed in the same order the RSX™ renders them to prevent deadlock Vertex and Index Data End of Free Area Start of Free Area
- 116. Each SPU has its own buffer to prevent deadlock Data 23 Data 17 Data 10 Data 22 Data 11 Data 16 Data 19 Data 6 Data 14 Data 20 Data 7 Data 15 Data 21 Data 12 Data 9 Data 18 Data 8 Data 13 SPU 0 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 Buffer 0 Buffer 1 Buffer 3 Buffer 2 Buffer 4 Buffer 5
- 117. RSX™ writes a semaphore once a chunk of data has been consumed A command to write a semaphore needs to be added to the command buffer after all commands that use the data The value of the semaphore to be written is the new end of free area pointer Data 19 Data 6 Data 14 Start of Free Area Draw 6 Draw 5 Draw 7 Semaphore Semaphore Semaphore Current RSX™ Execution New End of Free Area
- 118. SPU 0 SPU 1 SPU 2 SPU 3 SPU 4 SPU 5 RSX™
- 119. Future Work Cg compiler for SPUs Complicated vertex programs could be run on the SPUs instead of the RSX™ Can’t have too many outputs otherwise the RSX™ will take longer loading them than it would have to run the program
- 120. Future Work Shadow map generation on SPUs Large load removed from RSX™ Very doable Much more complicated if you have alpha cutouts in your textures
- 121. GCM Replay
- 122. GCM Replay New tool for use with the RSX™ Analysis Debugging Profiling Will be released soon to all licensed developers as part of PLAYSTATION®Edge Main tool runs on the PC Integration into your title is simple and easy
- 124. Capture a Snapshot
- 125. Render Targets Textures Vertex & Fragment Programs Vertex & Index Buffers Snapshot Contents Command Buffer
- 127. Performance Analysis
- 128. Find Bottlenecks
- 129. “What Ifs”
- 130. Q: What If I Had Efficient Triangle Culling? What would the performance gain be? GCM Replay can remove all draw calls to triangles which never write a pixel Once this is done, GCM Replay can reprofile the snapshot and compute the speed increase
- 131. A: Cull Triangles Using an SPU! Triangle culling techniques shown earlier can dramatically increase performance
- 132. Q: What If My Setting of Fragment Program Constants Was Done Externally to the Command Buffer? One copy of each fragment program is kept in memory Individual fragment program constants are patched by placing draw commands in the command buffer in the appropriate locations Fragment Program Patched Constants Conventional Patch Technique
- 133. A: Patch Using the PPU or SPU! Multiple copies of fragment programs can be patched with the appropriate constants either on the PPU or an SPU Removes 100% of the RSX™ load for patching fragment programs If done as part of SPU processing of a vertex set, synchronization will be already be taken care of Patched Copy 3 Patched Copy 2 Patched Copy 4 Patched Copy 1
- 134. Q: What If I Had More Optimal Indexed Triangle Lists?
- 135. A: Optimize for the Four Vertex Mini-cache! GCM Replay contains an optimizer for indexed triangle list ordering Corresponding offline indexed triangle list optimizer available as part of PLAYSTATION®Edge Vertex 1 Vertex 0 Vertex 2 Vertex 3 Four Vertex Mini-cache
- 136. Strip Example 3 new vertices 3
- 137. Strip Example 3 1 1 new vertex
- 138. Strip Example 3 1 1 1 new vertex
- 139. Strip Example 3 1 1 1 1 new vertex… 2 vertices + 1 per triangle in total
- 140. Free Form Example 3 new vertices 3
- 141. Free Form Example 1 new vertex 3 1
- 142. Free Form Example 1 new vertex 3 1 1
- 143. Free Form Example 1 new vertex 3 1 1 1
- 144. Free Form Example 1 new vertex 3 1 1 1 1
- 145. Free Form Example 0 new vertices! 3 1 1 1 1 0
- 146. Free Form Example 1 new vertex 3 1 1 1 1 0 1
- 147. Free Form Example 1 new vertex 3 1 1 1 1 0 1 1
- 148. Free Form Example 1 new vertex 3 1 1 1 1 0 1 1 1
- 149. Free Form Example 0 new vertices… 3 1 1 1 1 0 1 1 1 0 2 vertices + 3 per 4 triangles in total
- 150. Q: What If I Had Perfect Object Z-Culling? Some objects will not contribute to the final scene because they are entirely blocked by other objects GCM Replay will soon be able to show the performance difference if good object Z-culling was performed
- 151. A: Object Z-Culling on SPUs Write an SPU rasterizer Render the depth values of a low polygon version of the environment Rasterize and check bounding volumes of objects
- 152. Warhawk