sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 31 Packages Projects Releases Wiki Activity

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

Browse Source 
* opencl: add fastdiv for mm q8_0

* opencl: use uint4 for fastdiv vals

* opencl: use fastdiv for set_rows

* opencl: do not use fastdiv for q8_0 mm
This commit is contained in:
lhez
2025-11-10 15:00:13 -08:00
committed by GitHub
parent 7bef684118
commit ece0f5c177
2 changed files with 71 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/ggml-opencl.cpp
+36 -2
View File
@@ -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,
@@ -4464,6 +4495,9 @@ 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));
@@ -4474,8 +4508,8 @@ 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));