Skip to content

CPU: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators #16083

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 7 commits into from
Oct 15, 2025
Merged

CPU: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators #16083

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
51040ed
CPU: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators
safranowith Sep 18, 2025
20aa499
docs: add documentation for ggml_trunc and ggml_trunc_inplace in ggml.h
safranowith Sep 21, 2025
95b3127
chore: remove trailing whitespace from ggml.h
safranowith Sep 25, 2025
4866260
Merged updates from llama/master with conflict resolution
safranowith Oct 15, 2025
f340abc
Remove unresolved merge markers
safranowith Oct 15, 2025
2bd2a5a
Apply review suggestions: cleanup formatting, enum order and leftover…
safranowith Oct 15, 2025
bcb1c8d
Regenerate ops.md using create_ops_docs.py
safranowith Oct 15, 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
4 changes: 4 additions & 0 deletions docs/ops.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@ Legend:
| ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
| ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| CEIL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
| CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
| CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
Expand All @@ -41,6 +42,7 @@ Legend:
| ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
| EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
| FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
| FLOOR | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
| GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
| GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
Expand Down Expand Up @@ -82,6 +84,7 @@ Legend:
| ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
| ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
| ROUND | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
Expand All @@ -107,4 +110,5 @@ Legend:
| SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
| TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| TRUNC | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
16 changes: 16 additions & 0 deletions docs/ops/CPU.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,6 +59,14 @@
"CPU","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
"CPU","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
"CPU","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
"CPU","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","FLOOR","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","CEIL","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","CEIL","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","ROUND","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","ROUND","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","TRUNC","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","TRUNC","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","ABS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","ABS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","SGN","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
Expand Down Expand Up @@ -119,6 +127,14 @@
"CPU","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
"CPU","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","CPU"
"CPU","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","CPU"
"CPU","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","FLOOR","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","CEIL","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","CEIL","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","ROUND","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","ROUND","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","TRUNC","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","CPU"
"CPU","TRUNC","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","CPU"
"CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","1","yes","CPU"
"CPU","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","1","yes","CPU"
"CPU","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","1","yes","CPU"
Expand Down
45 changes: 45 additions & 0 deletions ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -574,6 +574,11 @@ extern "C" {
GGML_UNARY_OP_HARDSIGMOID,
GGML_UNARY_OP_EXP,
GGML_UNARY_OP_GELU_ERF,
GGML_UNARY_OP_FLOOR,
GGML_UNARY_OP_CEIL,
GGML_UNARY_OP_ROUND,
GGML_UNARY_OP_TRUNC,


GGML_UNARY_OP_COUNT,
};
Expand Down Expand Up @@ -1148,6 +1153,46 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_floor(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_floor_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_ceil(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_ceil_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_round(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_round_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);

/**
* Truncates the fractional part of each element in the tensor (towards zero).
* For example: trunc(3.7) = 3.0, trunc(-2.9) = -2.0
* Similar to std::trunc in C/C++.
*/

GGML_API struct ggml_tensor * ggml_trunc(
struct ggml_context * ctx,
struct ggml_tensor * a);

GGML_API struct ggml_tensor * ggml_trunc_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);



// gated linear unit ops
// A: n columns, r rows,
// result is n / 2 columns, r rows,
Expand Down
4 changes: 4 additions & 0 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 @@ -2179,6 +2179,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_TRUNC:
{
n_tasks = 1;
} break;
Expand Down
16 changes: 16 additions & 0 deletions ggml/src/ggml-cpu/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9766,6 +9766,22 @@ void ggml_compute_forward_unary(
{
ggml_compute_forward_exp(params, dst);
} break;
case GGML_UNARY_OP_FLOOR:
{
ggml_compute_forward_floor(params, dst);
} break;
case GGML_UNARY_OP_CEIL:
{
ggml_compute_forward_ceil(params, dst);
} break;
case GGML_UNARY_OP_ROUND:
{
ggml_compute_forward_round(params, dst);
} break;
case GGML_UNARY_OP_TRUNC:
{
ggml_compute_forward_trunc(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
Expand Down
32 changes: 32 additions & 0 deletions ggml/src/ggml-cpu/unary-ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,6 +64,22 @@ static inline float op_log(float x) {
return logf(x);
}

static inline float op_floor(float x) {
return floorf(x);
}

static inline float op_ceil(float x) {
return ceilf(x);
}

static inline float op_round(float x) {
return roundf(x);
}

static inline float op_trunc(float x) {
return truncf(x);
}

template <float (*op)(float), typename src0_t, typename dst_t>
static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
Expand Down Expand Up @@ -184,3 +200,19 @@ void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor *
void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_log>(params, dst);
}

void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_floor>(params, dst);
}

void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_ceil>(params, dst);
}

void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_round>(params, dst);
}

void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_trunc>(params, dst);
}
7 changes: 7 additions & 0 deletions ggml/src/ggml-cpu/unary-ops.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,13 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);




#ifdef __cplusplus
}
Expand Down
62 changes: 61 additions & 1 deletion ggml/src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1143,9 +1143,13 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"HARDSIGMOID",
"EXP",
"GELU_ERF",
"FLOOR",
"CEIL",
"ROUND",
"TRUNC",
};

static_assert(GGML_UNARY_OP_COUNT == 15, "GGML_UNARY_OP_COUNT != 15");
static_assert(GGML_UNARY_OP_COUNT == 19, "GGML_UNARY_OP_COUNT != 19");


static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
Expand Down Expand Up @@ -2726,6 +2730,62 @@ static struct ggml_tensor * ggml_glu_impl(
return result;
}

// ggml_floor

struct ggml_tensor * ggml_floor(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_FLOOR);
}

struct ggml_tensor * ggml_floor_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_FLOOR);
}

// ggml_ceil

struct ggml_tensor * ggml_ceil(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_CEIL);
}

struct ggml_tensor * ggml_ceil_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_CEIL);
}

//ggml_round

struct ggml_tensor * ggml_round(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_ROUND);
}

struct ggml_tensor * ggml_round_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_ROUND);
}

//ggml_trunc

struct ggml_tensor * ggml_trunc(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

IMO the naming trunc is quite vast and can be misinterpreted as truncating only N fractional decimals. Naming like torch.frac or python's modf() makes more sense.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for your feedback!
I have added documentation above the ggml_trunc and ggml_trunc_inplace functions in ggml.h to clarify their behavior (they truncate the fractional part towards zero, similar to std::trunc).
Please let me know if further clarification is needed.

struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_TRUNC);
}

struct ggml_tensor * ggml_trunc_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_TRUNC);
}

struct ggml_tensor * ggml_glu(
struct ggml_context * ctx,
struct ggml_tensor * a,
Expand Down
Loading