Skip to content

Repository navigation

hexagon: extend conv2d to conv3d #12

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

Closed
Mark Todorovich (marktodorovich) wants to merge 3 commits into from
Closed

hexagon: extend conv2d to conv3d #12

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e6b0448
hexagon: Generalization of CONV_2D implementation for CONV_3D
marktodorovich Oct 29, 2026
bb21a4a
hexagon: modify checking to eliminate divide-by-zero hazards
marktodorovich Oct 31, 2026
fdad23a
hexagon: conv2d tiled: compute the conv_3d unfold divisors on the host
marktodorovich Nov 4, 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
92 changes: 83 additions & 9 deletions ggml/src/ggml-hexagon/htp-kparams.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@ htp_op_code htp_op_remap(const struct ggml_tensor * t) {
case GGML_OP_IM2COL_3D: return HTP_OP_IM2COL_3D;
case GGML_OP_CONV_2D: return HTP_OP_CONV_2D;
case GGML_OP_CONV_2D_DW: return HTP_OP_CONV_2D_DW;
case GGML_OP_CONV_3D: return HTP_OP_CONV_3D;
case GGML_OP_ROLL: return HTP_OP_ROLL;
case GGML_OP_UPSCALE: return HTP_OP_UPSCALE;
case GGML_OP_LIGHTNING_INDEXER: return HTP_OP_LIGHTNING_INDEXER;
Expand Down Expand Up @@ -3994,16 +3995,30 @@ static bool htp_kparams_conv2d_tiled(

const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
const bool is3d = dst->op == GGML_OP_CONV_3D;
const int32_t s0 = ggml_get_op_params_i32(dst, 0);
const int32_t d0 = ggml_get_op_params_i32(dst, 4);
const int32_t d0 = ggml_get_op_params_i32(dst, is3d ? 6 : 4);

// range checks in 64 bits before narrowing, conv_3d folds KD into KH and OD into OH
const int64_t kh_w = is3d ? src0->ne[2] * src0->ne[1] : src0->ne[1];
const int64_t cin_w = is3d ? ggml_get_op_params_i32(dst, 9) : src0->ne[2];
const int64_t oh_w = is3d ? dst->ne[2] * dst->ne[1] : dst->ne[1];
const int64_t n_w = is3d ? ggml_get_op_params_i32(dst, 10) : dst->ne[3];
const int64_t k_w = src0->ne[0] * kh_w * cin_w;
const int64_t ct_w = n_w * oh_w * ((dst->ne[0] + 31) / 32);

// each stage buffer keeps KH input DMAs in flight on a 256-entry queue (main.c), which holds at most 255
if (kh_w < 1 || kh_w > 255 || k_w < 1 || k_w > (1 << 24) || ct_w < 1 || ct_w > UINT32_MAX) {
return false;
}

const uint32_t KW = src0->ne[0];
const uint32_t KH = src0->ne[1];
const uint32_t Cin = src0->ne[2];
const uint32_t Cout = src0->ne[3];
const uint32_t KH = kh_w;
const uint32_t Cin = cin_w;
const uint32_t Cout = is3d ? ggml_get_op_params_i32(dst, 11) : src0->ne[3];
const uint32_t OW = dst->ne[0];
const uint32_t OH = dst->ne[1];
const uint32_t N = dst->ne[3];
const uint32_t OH = oh_w;
const uint32_t N = n_w;
const size_t es = ggml_type_size(src0->type);

if (s0 != 1 || dst->nb[0] != sizeof(float)) {
Expand All @@ -4012,10 +4027,10 @@ static bool htp_kparams_conv2d_tiled(

// kernel k order must match the staged input: (cin, kh, kw) for WHCN input, (kh, kw, cin) for CWHN input.
// Kernel is either k-contiguous rows per cout (WHCN only) or cout-contiguous with a uniform k stride.
const bool whcn = src1->nb[0] == sizeof(float);
const bool whcn = is3d || src1->nb[0] == sizeof(float);
const bool cwhn = !whcn && src1->nb[2] == sizeof(float) && src1->nb[0] == Cin * sizeof(float);
const bool co_c = Cout == 1 || src0->nb[3] == es;
const bool w_rows = whcn && src0->nb[0] == es && (KH == 1 || src0->nb[1] == KW * es) && (Cin == 1 || src0->nb[2] == KW * KH * es);
const bool w_rows = is3d || (whcn && src0->nb[0] == es && (KH == 1 || src0->nb[1] == KW * es) && (Cin == 1 || src0->nb[2] == KW * KH * es));
const size_t ks = whcn ? src0->nb[0] : src0->nb[2];
bool kmajor = false;
if (whcn) {
Expand Down Expand Up @@ -4085,7 +4100,7 @@ static bool htp_kparams_conv2d_tiled(
const size_t thr2 = hex_round_up((std::max)(2 * stage + ostage + prow, wpack), tile);
struct htp_conv2d_vtcm_layout L2;
htp_conv2d_vtcm_layout_build(&L2, rt_block, w_size, act_size, out_size, thr2, nth);
if (stage <= 128 * 1024 && L2.total_bytes <= htp_session_vtcm_size(sess)) {
if (stage <= 128 * 1024 && 2 * KH <= 255 && L2.total_bytes <= htp_session_vtcm_size(sess)) {
n_stage = 2;
thr_size = thr2;
total = L2.total_bytes;
Expand Down Expand Up @@ -4115,6 +4130,8 @@ static bool htp_kparams_conv2d_tiled(
kp->tiled.wk_chunk = wk;
kp->tiled.div_ow_tiles = init_fastdiv_values(n_owt);
kp->tiled.div_oh = init_fastdiv_values(OH);
kp->tiled.div_khr = init_fastdiv_values(src0->ne[1]);
kp->tiled.div_ohr = init_fastdiv_values(dst->ne[1]);
return true;
}

Expand Down Expand Up @@ -4337,6 +4354,60 @@ static bool htp_kparams_solve_conv2d_dw(
return true;
}

static bool htp_kparams_solve_conv3d(
const struct ggml_hexagon_session * sess,
const struct ggml_tensor * op,
struct htp_conv2d_kernel_params * kparams
) {
memset(kparams, 0, sizeof(*kparams));

const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const struct ggml_tensor * dst = op;

if ((src0->type != GGML_TYPE_F16 && src0->type != GGML_TYPE_F32) ||
src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
return false;
}

const int32_t IC = ggml_get_op_params_i32(op, 9);
const int32_t N = ggml_get_op_params_i32(op, 10);
const int32_t OC = ggml_get_op_params_i32(op, 11);
if (IC <= 0 || N <= 0 || OC <= 0 || src0->ne[3] != (int64_t) IC * OC || src1->ne[3] != (int64_t) IC * N ||
dst->ne[3] != (int64_t) OC * N) {
return false;
}

for (int i = 0; i < 3; i++) {
const int32_t s = ggml_get_op_params_i32(op, i);
const int32_t p = ggml_get_op_params_i32(op, 3 + i);
const int32_t d = ggml_get_op_params_i32(op, 6 + i);
if (s <= 0 || d <= 0 || p < 0 || src0->ne[i] <= 0) {
return false;
}
const int64_t o = (src1->ne[i] + 2 * p - d * (src0->ne[i] - 1) - 1) / s + 1;
if (o != dst->ne[i]) {
return false;
}
}

const size_t es = ggml_type_size(src0->type);
const bool w_rows = src0->nb[0] == es && src0->nb[1] == src0->ne[0] * es && src0->nb[2] == src0->ne[1] * src0->nb[1] &&
((int64_t) IC * OC == 1 || src0->nb[3] == src0->ne[2] * src0->nb[2]);
if (!w_rows || src1->nb[0] != sizeof(float) || dst->nb[2] != dst->ne[1] * dst->nb[1]) {
return false;
}

const int64_t total_rows = dst->ne[1] * dst->ne[2] * dst->ne[3];
if (total_rows > UINT32_MAX) {
return false;
}
kparams->total_rows = (uint32_t) total_rows;
kparams->n_threads = total_rows == 0 ? 1 : (std::min)((uint32_t) htp_session_n_threads(sess), (uint32_t) total_rows);

return htp_kparams_conv2d_tiled(sess, op, kparams);
}

static bool htp_dsv4_hc_pre_can_collapse(
const struct ggml_tensor * x,
const struct ggml_tensor * weights,
Expand Down Expand Up @@ -5241,6 +5312,9 @@ bool htp_kparams_solve(
case HTP_OP_CONV_2D_DW:
return htp_kparams_invoke<htp_kparams_solve_conv2d_dw>(sess, op, kparams);

case HTP_OP_CONV_3D:
return htp_kparams_invoke<htp_kparams_solve_conv3d>(sess, op, kparams);

case HTP_OP_SUM_ROWS:
case HTP_OP_SUM:
case HTP_OP_ARGMAX:
Expand Down
Loading
Loading