CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-15 11:17:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

opencl: split ggml-opencl.cl into multiple files and cleanup (#12886)

Browse Source 
* opencl: refactor - split the kernel files

---------

Co-authored-by: Shangqing Gu <quic_shawngu@quicinc.com>

* opencl: split more kernels into separate files

* opencl: specify subgroup size instead of querying it

* opencl: refine Adreno cl compiler version parsing

* opencl: skip some kernels not used by Adreno on old compilers

* opencl: refine logic for selecting Adreno kernels

* opencl: refine Adreno cl compiler version

* opencl: cleanup preprocessor for kernels

* opencl: consider Adreno CL compiler on Windows

* opencl: add final newline for `mul_mv_f16_f16.cl`

---------

Co-authored-by: Shangqing Gu <quic_shawngu@quicinc.com>
This commit is contained in:
lhez
2025-04-15 12:26:00 -07:00
committed by GitHub
parent f8f820cc4d
commit 80f19b4186
43 changed files with 5021 additions and 4996 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
ggml/src/ggml-opencl/kernels/cvt.cl Normal file
View File

@@ -0,0 +1,118 @@
//------------------------------------------------------------------------------
// This file is contains kernels for data conversion.
// These kernels are used when loading the model, so its performance is less
// important.
//------------------------------------------------------------------------------
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#ifdef cl_intel_required_subgroup_size
#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
#define INTEL_GPU 1
#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
#elif defined(cl_qcom_reqd_sub_group_size)
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
#endif
#define QK4_0 32
#define QR4_0 2
#define QK4_1 32
#define QR4_1 2
#define QK5_0 32
#define QR5_0 2
#define QK5_1 32
#define QR5_1 2
#define QK8_0 32
#define QR8_0 1
#define QK_K 256
#define K_QUANTS_PER_ITERATION 2
typedef char int8_t;
typedef uchar uint8_t;
typedef short int16_t;
typedef ushort uint16_t;
typedef int int32_t;
typedef uint uint32_t;
//------------------------------------------------------------------------------
// block_q4_0
//------------------------------------------------------------------------------
struct block_q4_0
{
half d;
uint8_t qs[QK4_0 / 2];
};
//------------------------------------------------------------------------------
// kernel_convert_block_q4_0
// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
// This kernel does not deshuffle the bits.
//------------------------------------------------------------------------------
kernel void kernel_convert_block_q4_0(
global struct block_q4_0 * src0,
global uchar * dst_q,
global half * dst_d
) {
global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
global half * d = (global half *) dst_d + get_global_id(0);
*d = b->d;
for (int i = 0; i < QK4_0/2; ++i) {
q[i] = b->qs[i];
}
}
kernel void kernel_restore_block_q4_0(
global uchar * src_q,
global half * src_d,
global struct block_q4_0 * dst
) {
global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
global half * d = (global half *) src_d + get_global_id(0);
b->d = *d;
for (int i = 0; i < QK4_0/2; ++i) {
b->qs[i] = q[i];
}
}
//------------------------------------------------------------------------------
// kernel_convert_block_q4_0_noshuffle
// Flatten q4_0 weights and unshuffle the bits
//------------------------------------------------------------------------------
kernel void kernel_convert_block_q4_0_noshuffle(
global struct block_q4_0 * src0,
global uchar * dst_q,
global half * dst_d
) {
global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
global half * d = (global half *) dst_d + get_global_id(0);
*d = b->d;
for (int i = 0; i < QK4_0/4; ++i) {
uchar x0 = b->qs[2*i + 0];
uchar x1 = b->qs[2*i + 1];
q[i + 0 ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
q[i + QK4_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
#ifdef ADRENO_GPU
// Workaround for adreno - must have the following printf statement for
// the kernel to work properly. Otherwise it produces incorrect result.
// convert_uchar above also seems necessary.
// Compare against a large number so that it does not print anything.
// get_sub_group_local_id() also works.
if (get_global_id(0) == 65536*4096) {
printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
}
#endif
}
}