Skip to content

opencl: add fastdiv and use it in set_rows, ported from cuda #17090

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Nov 10, 2025
Merged

opencl: add fastdiv and use it in set_rows, ported from cuda #17090

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
69ddcdb
opencl: add fastdiv for mm q8_0
lhez Nov 4, 2025
792b353
opencl: use uint4 for fastdiv vals
lhez Nov 4, 2025
985cbe3
opencl: use fastdiv for set_rows
lhez Nov 5, 2025
c586dca
opencl: do not use fastdiv for q8_0 mm
lhez Nov 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 36 additions & 2 deletions ggml/src/ggml-opencl/ggml-opencl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,37 @@

bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor);

// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
// Precompute mp (m' in the paper) and L such that division
// can be computed using a multiply (high 32b of 64b result)
// and a shift:
//
// n/d = (mulhi(n, mp) + n) >> L;
struct fastdiv_vals {
uint32_t mp;
uint32_t L;
uint32_t d;
uint32_t pad;
};
static_assert(sizeof(fastdiv_vals) == 16, "fastdiv_vals size incorrect");

static fastdiv_vals init_fastdiv_values(uint64_t d_64) {
GGML_ASSERT(d_64 != 0);
GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());

uint32_t d = (uint32_t)d_64;

// compute L = ceil(log2(d));
uint32_t L = 0;
while (L < 32 && (uint32_t{ 1 } << L) < d) {
L++;
}

uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
// pack divisor as well to reduce error surface
return { mp, L, d, 0 };
}

enum GPU_FAMILY {
ADRENO,
INTEL,
Expand Down Expand Up @@ -4461,6 +4492,9 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c
GGML_ABORT("not implemented");
}

fastdiv_vals ne11_ = init_fastdiv_values(ne11);
fastdiv_vals ne12_ = init_fastdiv_values(ne12);

CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
Expand All @@ -4471,8 +4505,8 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb01));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb02));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb03));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne11));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(fastdiv_vals), &ne11_));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(fastdiv_vals), &ne12_));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
Expand Down
51 changes: 35 additions & 16 deletions ggml/src/ggml-opencl/kernels/set_rows.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,16 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable

// v = { mp, L, d }
inline uint fastdiv(uint n, uint4 v) {
uint msbs;
msbs = mul_hi(n, v.s0);
return (msbs + n) >> v.s1;
}
inline uint fastmod(uint n, uint4 v) {
uint q = fastdiv(n, v);
return n - q * v.s2;
}

kernel void kernel_set_rows_f32_i64(
global char * src0,
ulong offset0,
Expand All @@ -11,8 +22,8 @@ kernel void kernel_set_rows_f32_i64(
ulong nb01,
ulong nb02,
ulong nb03,
int ne11,
int ne12,
uint4 ne11,
uint4 ne12,
ulong nb10,
ulong nb11,
ulong nb12,
Expand All @@ -33,8 +44,10 @@ kernel void kernel_set_rows_f32_i64(
return;
}

int i12 = i03%ne12;
int i11 = i02%ne11;
//int i12 = i03%ne12;
//int i11 = i02%ne11;
int i12 = fastmod(i03, ne12);
int i11 = fastmod(i02, ne11);

int i10 = i01;
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
Expand All @@ -58,8 +71,8 @@ kernel void kernel_set_rows_f16_i64(
ulong nb01,
ulong nb02,
ulong nb03,
int ne11,
int ne12,
uint4 ne11,
uint4 ne12,
ulong nb10,
ulong nb11,
ulong nb12,
Expand All @@ -80,8 +93,10 @@ kernel void kernel_set_rows_f16_i64(
return;
}

int i12 = i03%ne12;
int i11 = i02%ne11;
//int i12 = i03%ne12;
//int i11 = i02%ne11;
int i12 = fastmod(i03, ne12);
int i11 = fastmod(i02, ne11);

int i10 = i01;
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
Expand All @@ -105,8 +120,8 @@ kernel void kernel_set_rows_f32_i32(
ulong nb01,
ulong nb02,
ulong nb03,
int ne11,
int ne12,
uint4 ne11,
uint4 ne12,
ulong nb10,
ulong nb11,
ulong nb12,
Expand All @@ -127,8 +142,10 @@ kernel void kernel_set_rows_f32_i32(
return;
}

int i12 = i03%ne12;
int i11 = i02%ne11;
//int i12 = i03%ne12;
//int i11 = i02%ne11;
int i12 = fastmod(i03, ne12);
int i11 = fastmod(i02, ne11);

int i10 = i01;
int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
Expand All @@ -152,8 +169,8 @@ kernel void kernel_set_rows_f16_i32(
ulong nb01,
ulong nb02,
ulong nb03,
int ne11,
int ne12,
uint4 ne11,
uint4 ne12,
ulong nb10,
ulong nb11,
ulong nb12,
Expand All @@ -174,8 +191,10 @@ kernel void kernel_set_rows_f16_i32(
return;
}

int i12 = i03%ne12;
int i11 = i02%ne11;
//int i12 = i03%ne12;
//int i11 = i02%ne11;
int i12 = fastmod(i03, ne12);
int i11 = fastmod(i02, ne11);

int i10 = i01;
int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
Expand Down