Adapt to CUDA.jl profile changes

[thomasfaingnaert]

ref JuliaGPU/CUDA.jl#2139

Probably should add methods to CUDA.jl to access the kernel times instead of grabbing them directly from the dataframe.

[maleadt]

Benchmark results for commit a427f5e (comparing to 93ef046):

test	master	PR	Δmin
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (4096×4096) · (4096×4096) (NN) OP (16, 16, 16)	480.0 μs ± 3.85% (456.0 … 506.0 μs) 255 regs	432.0 μs ± 3.83% (414.0 … 462.0 μs) 136 regs	-9.2% ✅
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (TN) OP (8, 16, 2), base shape (4, 8, 1)	14.6 μs ± 0.908% (14.5 … 15.0 μs) 80 regs	124.0 μs ± 0.153% (123.0 … 124.0 μs) 116 regs	+747.5% ❌
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (4096×4096) · (4096×4096) (NT) OP (16, 16, 16)	474.0 μs ± 3.52% (451.0 … 495.0 μs) 255 regs	430.0 μs ± 4.26% (405.0 … 460.0 μs) 138 regs	-10.2% ✅
Tropical GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1)	9.17 μs ± 1.65% (8.82 … 9.54 μs) 80 regs	63.7 μs ± 0.27% (63.4 … 64.1 μs) 118 regs	+618.9% ❌
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (TT) OP (8, 16, 2), base shape (4, 8, 1)	9.54 μs ± 1.6% (9.3 … 9.78 μs) 82 regs	64.0 μs ± 0.263% (63.7 … 64.4 μs) 123 regs	+584.6% ❌
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1)	16.3 μs ± 0.985% (16.0 … 16.5 μs) 82 regs	110.0 μs ± 0.144% (110.0 … 111.0 μs) 123 regs	+589.6% ❌
WMMA GEMM Float16*Float16=Float32 (128×128) · (128×128) (NN) OP (16, 16, 16)	9.15 μs ± 1.8% (8.82 … 9.3 μs)	8.06 μs ± 2.14% (7.87 … 8.34 μs)	-10.8% ✅
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (128×128) · (128×128) (NN) OP (16, 16, 16)	12.8 μs ± 1.13% (12.4 … 12.9 μs) 154 regs	9.93 μs ± 1.29% (9.54 … 10.0 μs) 119 regs	-23.1% ✅
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (NT) OP (8, 16, 2), base shape (4, 8, 1)	18.3 μs ± 0.86% (18.1 … 18.6 μs) 80 regs	127.0 μs ± 0.144% (127.0 … 128.0 μs) 111 regs	+598.7% ❌
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (TT) OP (8, 16, 2), base shape (4, 8, 1)	16.1 μs ± 0.872% (15.7 … 16.5 μs) 82 regs	125.0 μs ± 0.123% (125.0 … 126.0 μs) 123 regs	+692.4% ❌
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (256×256) · (256×256) (NT) OP (16, 16, 16)	16.9 μs ± 0.677% (16.7 … 17.2 μs) 255 regs	14.3 μs ± 0.972% (14.1 … 14.5 μs) 121 regs	-15.7% ✅
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (128×128) · (128×128) (NT) OP (16, 16, 16)	11.3 μs ± 1.04% (11.0 … 11.7 μs) 154 regs	9.89 μs ± 1.49% (9.54 … 10.0 μs) 120 regs	-13.0% ✅
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1)	18.4 μs ± 0.705% (18.1 … 18.8 μs) 80 regs	127.0 μs ± 0.134% (127.0 … 128.0 μs) 111 regs	+600.0% ❌
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (NT) OP (8, 16, 2), base shape (4, 8, 1)	10.7 μs ± 1.3% (10.5 … 11.0 μs) 80 regs	64.9 μs ± 0.267% (64.4 … 65.3 μs) 111 regs	+513.6% ❌
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1)	14.8 μs ± 1.06% (14.5 … 15.0 μs) 80 regs	109.0 μs ± 0.161% (109.0 … 110.0 μs) 116 regs	+650.8% ❌
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (TN) OP (8, 16, 2), base shape (4, 8, 1)	8.76 μs ± 1.77% (8.58 … 9.06 μs) 80 regs	63.2 μs ± 0.23% (62.9 … 63.7 μs) 116 regs	+633.3% ❌
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (NN) OP (8, 16, 2), base shape (4, 8, 1)	15.7 μs ± 1.09% (15.5 … 16.0 μs) 80 regs	125.0 μs ± 0.118% (124.0 … 125.0 μs) 118 regs	+703.1% ❌
Tropical GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1)	9.42 μs ± 1.59% (9.06 … 9.78 μs) 82 regs	63.8 μs ± 0.27% (63.4 … 64.1 μs) 123 regs	+600.0% ❌
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (256×256) · (256×256) (NN) OP (16, 16, 16)	17.0 μs ± 0.757% (16.7 … 17.2 μs) 255 regs	14.2 μs ± 1.0% (14.1 … 14.5 μs) 121 regs	-15.7% ✅
Tropical GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1)	8.65 μs ± 1.82% (8.34 … 8.82 μs) 80 regs	55.6 μs ± 0.214% (55.1 … 56.0 μs) 116 regs	+560.0% ❌
Tropical GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1)	10.5 μs ± 0.723% (10.3 … 11.0 μs) 80 regs	57.0 μs ± 0.303% (56.7 … 57.5 μs) 111 regs	+453.5% ❌
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1)	15.9 μs ± 0.93% (15.5 … 16.2 μs) 80 regs	110.0 μs ± 0.153% (110.0 … 111.0 μs) 118 regs	+607.7% ❌
WMMA GEMM Float16*Float16=Float16 (2048×2048) · (2048×2048) (NT) OP (16, 16, 16)	277.0 μs ± 6.14% (256.0 … 298.0 μs)	311.0 μs ± 4.91% (278.0 … 324.0 μs)	+8.7% ❌
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (NN) OP (8, 16, 2), base shape (4, 8, 1)	9.32 μs ± 1.81% (9.06 … 9.54 μs) 80 regs	63.8 μs ± 0.276% (63.4 … 64.4 μs) 118 regs	+600.0% ❌