vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Work around q5_0 OpenCL issue

Browse source
This commit is contained in:
0cc4m 2023-04-29 10:26:58 +02:00
parent 9439da6f95
commit 1560c10f24
2 changed files with 33 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-opencl-dequant.cl
View file

@ -63,7 +63,7 @@ __kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global f
struct block_q5_0
{
ushort d;
float d;
uint qh;
uchar qs[16];
};
@ -72,7 +72,7 @@ __kernel void dequantize_row_q5_0(__global struct block_q5_0* blocks, __global f
const uint i = get_global_id(0) / 32;
const uint l = get_local_id(0);
const float d = vload_half(0, (__global half*) &blocks[i].d);
const float d = blocks[i].d;
const uchar vi = blocks[i].qs[l];

34
ggml-opencl.c
View file

@ -19,6 +19,20 @@
} \
} while (0)
#define QK5_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
typedef struct {
float d; // delta
uint32_t qh; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} cl_block_q5_0;
static cl_platform_id platform;
static cl_device_id device;
static cl_context context;
@ -131,6 +145,7 @@ void ggml_cl_sgemm_wrapper(
cl_kernel kernel;
size_t global = n * k, local, size_qb;
bool dequant;
cl_block_q5_0* cl_host_b;
switch (btype) {
case GGML_TYPE_F32:
@ -152,19 +167,29 @@ void ggml_cl_sgemm_wrapper(
dequant = true;
kernel = kernel_q4_2;
local = 8;
size_qb = global * (sizeof(short) + local) / 16;
size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
break;
case GGML_TYPE_Q5_0:
dequant = true;
kernel = kernel_q5_0;
local = 16;
size_qb = global * (sizeof(short) + 4 + local) / 32;
// For some reason OpenCL seems to be incapable of working with structs of size 22.
// 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU...
// TODO Find the reason, fix and remove workaround.
const block_q5_0* b = (const block_q5_0*) host_b;
cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32);
for (size_t i = 0; i < global / 32; i++) {
cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d);
memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t) + QK5_0 / 2);
}
host_b = (const float*) cl_host_b;
size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32;
break;
case GGML_TYPE_Q5_1:
dequant = true;
kernel = kernel_q5_1;
local = 16;
size_qb = global * (sizeof(short) * 2 + 4 + local) / 32;
size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
break;
case GGML_TYPE_Q8_0:
dequant = true;
@ -237,4 +262,7 @@ void ggml_cl_sgemm_wrapper(
clWaitForEvents(1, &ev_c);
clReleaseEvent(ev_sgemm);
clReleaseEvent(ev_c);
if (btype == GGML_TYPE_Q5_0) {
free((void*) cl_host_b);
}
}