--- title: "GPU Programming Tip Line Thread /" type: archive source: twitter source_url: "https://x.com/NOTimothyLottes/status/1588902081502797826" author: "NOTimothyLottes" handle: NOTimothyLottes post_id: "1588902081502797826" date: 2022-11-05 archived: 2026-08-23 draft: false tags: - archive - twitter - NOTimothyLottes description: "GPU Programming Tip Line Thread /" in_reply_to: "" --- ## Source - URL: https://x.com/NOTimothyLottes/status/1588902081502797826 - Author: NOTimothyLottes (@NOTimothyLottes) - Posted: 2022-11-05 14:32:31 ## Thread **1/** **@NOTimothyLottes** ^1588902081502797826 GPU Programming Tip Line Thread / **2/** **@NOTimothyLottes** ^1588902190034620417 [0] Normalization fail case is 'rsq(0)=INF*0=NaN', trim out intermediate INF to fix, 'normalize_safe(x){return x*min(MAX_FLOAT,rsq(dot(x,x)));}' **3/** **@NOTimothyLottes** ^1588902600686329857 [1] PC FP16 1/denormals generates INFs which can easily eventually result in NaNs, fix positives with 'rcp(max(x,SMALLEST_NORMAL))' **4/** **@NOTimothyLottes** ^1588903127549607936 [2] 'spirv-opt -Os' (optimize for size) is your weapon against "back-before-its-done--getting-groceries" IHV compile times **5/** **@NOTimothyLottes** ^1588903378952007680 [3] Driver ignoring your `[[dont_unroll]]`, 'int eatThisBuddy=1+/*hidden-zero*/constantBuffer.zero[0];for(i=0;i<1024;i+=eatThisBuddy){...' **6/** **@NOTimothyLottes** ^1588903586721042432 [4] 'Ship-One-Shader' = One SPIR-V binary, use specialization constants to select shader at PSO generation time, minimizes released shader binary size **7/** **@NOTimothyLottes** ^1588903663468437508 [5] 'Ship-One-Shader' requires 'spirv-opt -Os' **8/** **@NOTimothyLottes** ^1588903725783220224 [6] Use signed 'mask=bitfieldExtract(int(v),bit,1)' to turn bit into all 0's or 1's mask **9/** **@NOTimothyLottes** ^1588903926510014467 [7] On PC for AMD's native V_BFI_B32 (to select bits based on mask) use 'Bfi(int src,int ins,int mask){return (ins&mask)|(src&(~mask));}' **10/** **@NOTimothyLottes** ^1588904013239844870 [8] Use 'bitfieldInsert(,,0,compileTimeImmediate)' hits fast V_BFI_B32 on AMD, and portable to other vendors **11/** **@NOTimothyLottes** ^1588904166269005824 [9] AMD ONLY: clamp(a,b,c) is implemented as med3(a,b,c), so can get V_MED3_* without an extension (in a non-portable way) **12/** **@NOTimothyLottes** ^1588904259265114112 [10] Bools as 0|1 floats '(a&b)|c' can be done via 'saturate(a*b+c)' **13/** **@NOTimothyLottes** ^1588904325543493633 [11] Bools as 0|1 floats '!(a&b)' can be done via '(-a)*b+1.0' **14/** **@NOTimothyLottes** ^1588904434268262400 [12] Convert INFs to NaNs via 'x*0.0+x' **15/** **@NOTimothyLottes** ^1588904891707424768 [13] Semi-persistent workgroup opt = reusing the workgroup for more work before exit (ie processing four 8x8 tiles in a 16x16 footprint using a 64-wide group) **16/** **@NOTimothyLottes** ^1588905179449282560 [14] Semi-persistent workgroups can be good for up to 10% perf on AMD (YMMV) ... assuming compiler doesn't fail VGPR allocation, check your disassembly **17/** **@NOTimothyLottes** ^1588905286697639936 [15] Semi-persistent workgroups gain by keeping more local work on the same L0, by factoring out wait for store on wave exit, and better scheduling **18/** **@NOTimothyLottes** ^1588905339122257921 [16] Merge passes to avoid round trip through DRAM, often huge wins there **19/** **@NOTimothyLottes** ^1588905491027341312 [17] Sometimes serial dependent passes can be merged into one shader to keep work in L2 for >10% gains, requires "unsafe you-shouldnt-do-that" logic that works **20/** **@NOTimothyLottes** ^1588905764005216258 [18] Proper double rate "packed" 16-bit can provides gains up to 30% depending on workload/platform (except NV) **21/** **@NOTimothyLottes** ^1588906002212323328 [19] Even without double rate 16-bit, 16-bit is the most important tool for managing register pressure problems, esp with smaller HW register limits or compiler troubles