llama.cpp/kompute/op_mul_mat_mat_q4_0.comp
2023-11-03 17:22:22 -04:00

77 lines
2.5 KiB
Plaintext

/**
* Copyright (c) 2023 Nomic, Inc. All rights reserved.
*
* This software is licensed under the terms of the Software for Open Models
* License (SOM), version 1.0, as detailed in the LICENSE_SOM.txt file. A copy
* of this license should accompany this software. Except as expressly granted
* in the SOM license, all rights are reserved by Nomic, Inc.
*/
#version 450
#include "common.comp"
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
layout (local_size_x_id = 0) in;
layout (local_size_y_id = 1) in;
layout (constant_id = 1) const uint nsg = 2;
layout(binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout(binding = 1) readonly buffer tensorInB { float inB[]; };
layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
layout(push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
int ne11;
int ne12;
uint nb01;
uint nb02;
uint nb11;
uint nb12;
uint nb1;
uint nb2;
}
pcs;
const uint els_per_block = 32;
const uint qs_offset = 2;
const uint block_size = (els_per_block / 2) + qs_offset;
void main() {
uvec3 gid = gl_WorkGroupID;
uvec3 lid = gl_LocalInvocationID;
gid.y = gid.y * nsg + lid.y;
uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) + pcs.inAOff; // Based from inA
const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
float sum = 0.0f;
for (uint i = gl_SubgroupInvocationID * 2; i < pcs.ne00; i+=gl_SubgroupSize * 2) {
const uint block_number = i / els_per_block;
const uint block_offset = block_number * block_size;
const float d = u8BufToFloat16(inA, x + block_offset);
const uint j = (i % els_per_block) / 2;
const uint byte_position_in_block = j;
const int q0 = (inA[x+block_offset+qs_offset+byte_position_in_block] & 0x0F) - 8;
const int q1 = (inA[x+block_offset+qs_offset+byte_position_in_block] >> 4) - 8;
const float dq0 = d * q0;
const float dq1 = d * q1;
const uint block_base = block_number * els_per_block;
sum += (dq0 * float(inB[y+block_base+j])) + \
(dq1 * float(inB[y+block_base+j+(els_per_block/2)]));
}
const float all_sum = subgroupAdd(sum);
if (subgroupElect())
out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
}