Commits
Sushanth Rajasankar authored and GitHub committed 271c509d592
DP4AMatMul perf refinements (#23539)
In this change
1. Vectorization of k is updated to 4.
2. Tile_A, Tile_B are stored transposed in shared memory. This makes it
so that memory locality is improved for our access pattern.
3. Lane output is switched to being individual vectors and its loop
unrolled, this solves the problem where laneoutput was not on registers
before.
Perf improvements are not very consistent with this change. On Tigerlake
GPU with 32.0.101.6460 (latest intel drivers)
```
Baseline
model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web\ -l 1000
Batch size: 1, prompt tokens: 1001, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 7.36557e+06 <<<<
avg (tokens/s): 135.903
p50 (us): 7.35498e+06
stddev (us): 27599
n: 5 * 1001 token(s)
With Change
model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web\ -l 1000
Batch size: 1, prompt tokens: 1001, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 6.52302e+06 <<<<
avg (tokens/s): 153.457
p50 (us): 6.52224e+06
stddev (us): 10407.3
n: 5 * 1001 token(s)
```
However, using the Intel GPA comparing before and after profile, one can
clearly see straight runs of ALU work without being interspersed by
writebacks to local memory that contained lane_output before.
