Skip to content

perf: ARM CPU conv kernels for DocTR detection #152

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
olyasir wants to merge 4 commits into
base: temp-8828
Choose a base branch
from
Open

perf: ARM CPU conv kernels for DocTR detection #152

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
7426e7b
perf: ARM CPU conv kernels for DocTR detection
olyasir Jun 15, 2026
47ba155
chore: drop OCR_CPU_PROF dev profiler from CPU graph compute
olyasir Jun 16, 2026
69ff143
chore: drop dead llamafile re-enable + F16-src1 mul_mat path
olyasir Jun 16, 2026
b64c886
Restore F16-src1 mul_mat handling (conservative, LLM safety)
olyasir Jun 17, 2026
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
106 changes: 101 additions & 5 deletions ggml/src/ggml-cpu/ggml-cpu.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1384,7 +1384,7 @@ UseGgmlGemm1:;
const size_t nbw3 = nbw2*ne12;

assert(params->wsize >= ne13*nbw3);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);

#if 0
for (int64_t i13 = 0; i13 < ne13; ++i13) {
Expand All @@ -1403,9 +1403,28 @@ UseGgmlGemm1:;
size_t bs = ggml_blck_size(vec_dot_type);
int64_t ne10_block_start = (ith * ne10/bs) / nth;
int64_t ne10_block_end = ((ith + 1) * ne10/bs) / nth;
from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0),
(ne10_block_end - ne10_block_start) * bs);
if (src1->type == GGML_TYPE_F32) {
from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0),
(ne10_block_end - ne10_block_start) * bs);
} else {
// F16 src1 feeding a quantized src0 (e.g. an f16
// im2col against Q8_0 conv weights): convert to f32
// in cache-sized chunks, then quantize into wdata.
const ggml_fp16_t * x16 = (const ggml_fp16_t *)((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11) + ne10_block_start*bs;
char * wrow = wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0;
const int64_t n = (ne10_block_end - ne10_block_start) * bs;
float tmp_f32[1024]; // multiple of every block size
int64_t done = 0;
while (done < n) {
const int64_t chunk = MIN(n - done, (int64_t) 1024);
for (int64_t k = 0; k < chunk; ++k) {
tmp_f32[k] = GGML_CPU_FP16_TO_FP32(x16[done + k]);
}
from_float(tmp_f32, (void *)(wrow + (done/bs)*nbw0), chunk);
done += chunk;
}
}
}
}
}
Expand Down Expand Up @@ -3103,7 +3122,84 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
continue;
}

ggml_compute_forward(&params, node);
// Collapse CONV_2D -> ADD(per-channel bias) [-> UNARY(relu|hardswish)]
// chains into one fused kernel call: the bias add and activation are
// applied while the GEMM result is still hot instead of as separate
// full-tensor passes. The decision depends only on the (shared) graph,
// so every thread takes the same branch and the per-node barriers stay
// in sync; the skipped nodes' outputs are never read (single use,
// checked by ggml_can_fuse).
int fuse_extra = 0;
int32_t fuse_act = -1;
const struct ggml_tensor * fuse_bias = NULL;
struct ggml_tensor * fuse_out = NULL;
if (node->op == GGML_OP_CONV_2D) {
static const enum ggml_op fuse_ops3[3] = {GGML_OP_CONV_2D, GGML_OP_ADD, GGML_OP_UNARY};
static const enum ggml_op fuse_ops2[2] = {GGML_OP_CONV_2D, GGML_OP_ADD};
if (ggml_can_fuse(cgraph, node_n, fuse_ops3, 3)) {
struct ggml_tensor * unary_node = cgraph->nodes[node_n + 2];
const enum ggml_unary_op uop = ggml_get_unary_op(unary_node);
if ((uop == GGML_UNARY_OP_RELU || uop == GGML_UNARY_OP_HARDSWISH) &&
unary_node->type == GGML_TYPE_F32) {
fuse_extra = 2;
fuse_act = (int32_t) uop;
fuse_out = unary_node;
}
}
if (fuse_extra == 0 && ggml_can_fuse(cgraph, node_n, fuse_ops2, 2)) {
fuse_extra = 1;
fuse_out = cgraph->nodes[node_n + 1];
}
if (fuse_extra > 0) {
const struct ggml_tensor * add_node = cgraph->nodes[node_n + 1];
const struct ggml_tensor * bias =
add_node->src[0] == node ? add_node->src[1] : add_node->src[0];
// require ADD to broadcast a contiguous F32 [1,1,OC,1] row over
// the conv output (the conv must be the full-shape operand)
if (add_node->src[0] == node && add_node->type == GGML_TYPE_F32 &&
bias->type == GGML_TYPE_F32 && ggml_is_contiguous(bias) &&
bias->ne[0] == 1 && bias->ne[1] == 1 &&
bias->ne[2] == node->ne[2] && bias->ne[3] == 1) {
fuse_bias = bias;
} else {
fuse_extra = 0;
fuse_act = -1;
fuse_out = NULL;
}
}
}

// Standalone ADD(per-channel bias) -> UNARY(relu|hardswish) pairs (e.g.
// after the explicit im2col+mul_mat conv lowering) fuse into one pass.
if (fuse_extra == 0 && node->op == GGML_OP_ADD) {
static const enum ggml_op fuse_au[2] = {GGML_OP_ADD, GGML_OP_UNARY};
if (ggml_can_fuse(cgraph, node_n, fuse_au, 2)) {
struct ggml_tensor * unary_node = cgraph->nodes[node_n + 1];
const enum ggml_unary_op uop = ggml_get_unary_op(unary_node);
const struct ggml_tensor * a0 = node->src[0];
const struct ggml_tensor * a1 = node->src[1];
if ((uop == GGML_UNARY_OP_RELU || uop == GGML_UNARY_OP_HARDSWISH) &&
node->type == GGML_TYPE_F32 && unary_node->type == GGML_TYPE_F32 &&
a0->type == GGML_TYPE_F32 && a1->type == GGML_TYPE_F32 &&
ggml_are_same_shape(node, a0) &&
ggml_is_contiguous(a0) && ggml_is_contiguous(a1) &&
ggml_is_contiguous(unary_node) &&
a1->ne[0] == 1 && a1->ne[1] == 1 &&
a1->ne[2] == a0->ne[2] && a1->ne[3] == 1) {
fuse_extra = 1;
fuse_act = (int32_t) uop;
fuse_out = unary_node;
ggml_compute_forward_add_unary_fused(&params, node, fuse_act, fuse_out);
node_n += fuse_extra;
}
}
} else if (fuse_extra > 0) {
ggml_compute_forward_conv_2d_fused(&params, node, fuse_bias, fuse_act, fuse_out);
node_n += fuse_extra;
}
if (fuse_extra == 0) {
ggml_compute_forward(&params, node);
}

if (state->ith == 0 && cplan->abort_callback &&
cplan->abort_callback(cplan->abort_callback_data)) {
Expand Down
Loading
Loading