Skip to content

ggml-cpu: Add bounds checking in make_block_q4_0x4 function #1095

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

Open
ilhamsyahids wants to merge 2 commits into
base: master
Choose a base branch
from
+7 −4
Open

ggml-cpu: Add bounds checking in make_block_q4_0x4 function #1095

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9d5f18b
ggml-cpu: Add bounds checking in `make_block_q4_0x4` function
ilhamsyahids Feb 2, 2025
6bd5fb2
ggml-cpu: Simplify bounds checking in `make_block_q4_0x4`
ilhamsyahids Feb 4, 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
37 changes: 26 additions & 11 deletions src/ggml-cpu/ggml-cpu-aarch64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3592,13 +3592,16 @@ static void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs,
}

static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x4 out;
// Zero initialize the output structure
block_q4_0x4 out = {};

// Copy d values
for (int i = 0; i < 4; i++) {
out.d[i] = in[i].d;
}

const int end = QK4_0 * 2 / blck_size_interleave;
const size_t qs_size = sizeof(out.qs);

if (blck_size_interleave == 8) {
const uint64_t xor_mask = 0x8888888888888888ULL;
Expand All @@ -3607,11 +3610,17 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;

uint64_t elems;
// Using memcpy to avoid unaligned memory accesses
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
// Bounds checking
if (dst_offset + sizeof(uint64_t) <= qs_size &&
src_offset + sizeof(uint64_t) <= sizeof(in[src_id].qs)) {
uint64_t elems;
// Using memcpy to avoid unaligned memory accesses
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
} else {
GGML_ASSERT(false && "buffer overflow prevented in make_block_q4_0x4");
}
}
} else if (blck_size_interleave == 4) {
const uint32_t xor_mask = 0x88888888;
Expand All @@ -3620,13 +3629,19 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;

uint32_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
// Bounds checking
if (dst_offset + sizeof(uint32_t) <= qs_size &&
src_offset + sizeof(uint32_t) <= sizeof(in[src_id].qs)) {
uint32_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
} else {
GGML_ASSERT(false && "buffer overflow prevented in make_block_q4_0x4");
}
}
} else {
GGML_ASSERT(false);
GGML_ASSERT(false && "invalid block size interleave value");
}

return out;
Expand Down