yum-archive/TaSTT-Whisper

High-performance GPGPU inference of OpenAI's Whisper automatic speech recognition (ASR) model

git clone https://git.yummers.dev/yum-archive/TaSTT-Whisper

KonstantinSource codes8c4603c

master
929 B39 linesraw
1#ifndef THREADS
2#define THREADS 512
3#endif
4
5Buffer<float> arg0: register( t0 );
6RWBuffer<float> result: register( u0 );
7
8cbuffer Constants: register( b0 )
9{
10	uint4 size: packoffset( c0 );
11	uint4 strides: packoffset( c1 );
12	uint4 argStrides: packoffset( c3 );
13}
14
15inline uint rowOffset( uint3 idx, uint4 strides )
16{
17	return idx[ 0 ] * strides[ 1 ] + idx[ 1 ] * strides[ 2 ] + idx[ 2 ] * strides[ 3 ];
18}
19
20[ numthreads( THREADS, 1, 1 ) ]
21void main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )
22{
23	uint rdi = rowOffset( group, strides );
24	uint rsi = rowOffset( group, argStrides );
25
26	const uint rdiEnd = rdi + size[ 0 ] * strides[ 0 ];
27	rdi += thread * strides[ 0 ];
28	rsi += thread * argStrides[ 0 ];
29
30	const uint rdiInc = THREADS * strides[ 0 ];
31	const uint rsiInc = THREADS * argStrides[ 0 ];
32
33	for( ; rdi < rdiEnd; rdi += rdiInc, rsi += rsiInc )
34	{
35		float f = result[ rdi ];
36		f += arg0[ rsi ];
37		result[ rdi ] = f;
38	}
39}