| |
| """Round 6: which matmul operand each math fidelity truncates (ttnn.matmul, bf16 x bf16, fp32 DST, fp32 out). |
| Pre-rounding an operand to b explicit mantissa bits makes the truncation exact: HiFi2 is unbiased once in0 has |
| <= 6 bits (it truncates the ACTIVATION to 6 bits: -0.28% magnitude bias), HiFi3 / HiFi4 are unbiased.""" |
| import sys, torch |
| sys.path.insert(0,'/home/ttuser/experiments/tt-models/tools') |
| from open_device_12x10 import open_device, close_device |
| import ttnn |
| def rnd(x, bits): |
| m, e = torch.frexp(x.double()) |
| s = 2.0 ** (bits + 1) |
| return torch.ldexp(torch.round(m * s) / s, e).to(torch.bfloat16) |
| dev=open_device(l1_small_size=0, trace_region_size=0) |
| try: |
| torch.manual_seed(0) |
| M,K,N=192,1024,1024 |
| a0=torch.randn(M,K).bfloat16(); w0=(torch.randn(K,N)/32).bfloat16() |
| for tag,a,w in (("plain",a0,w0),("a6",rnd(a0,6),w0),("w6",a0,rnd(w0,6)),("a5",rnd(a0,5),w0),("w5",a0,rnd(w0,5)),("a4",rnd(a0,4),w0),("w4",a0,rnd(w0,4))): |
| refq=a.double()@w.double() |
| ta=ttnn.from_torch(a,dtype=ttnn.bfloat16,layout=ttnn.TILE_LAYOUT,device=dev) |
| tw=ttnn.from_torch(w,dtype=ttnn.bfloat16,layout=ttnn.TILE_LAYOUT,device=dev) |
| for fid in ("LoFi","HiFi2","HiFi3"): |
| cc=ttnn.init_device_compute_kernel_config(dev.arch(),math_fidelity=getattr(ttnn.MathFidelity,fid),fp32_dest_acc_en=True,packer_l1_acc=True) |
| o=ttnn.to_torch(ttnn.matmul(ta,tw,compute_kernel_config=cc,dtype=ttnn.float32)).double() |
| e=o-refq |
| print(f"{tag:6s} {fid}: mag-bias {float((e*refq.sign()).mean()/refq.abs().mean()):+.2e} rel|e| {float(e.abs().mean()/refq.abs().mean()):.2e}") |
| finally: |
| close_device(dev) |
|
|
