happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

cpu : add FLOOR, CEIL, ROUND and TRUNC unary operators (#16083) 

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

- Added the operators to unary op enum
- Implemented API functions
- Implemented forward and unary-op logic in CPU backend
- Updated ggml_get_n_tasks
- Updated operators names array and static_assert
- Updated docs and enabled automatic tests

* docs: add documentation for ggml_trunc and ggml_trunc_inplace in ggml.h

* chore: remove trailing whitespace from ggml.h

* Remove unresolved merge markers

* Apply review suggestions: cleanup formatting, enum order and leftover artifacts

* Regenerate ops.md using create_ops_docs.py
This commit is contained in:
safranowith 2025-10-15 22:24:51 +03:00 committed by GitHub
parent 0cb7a0683b
commit 466c1911ab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 181 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
docs/ops.md
View File

@ -22,6 +22,7 @@ Legend:
| ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
| ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| CEIL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
| CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
| CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
@ -41,6 +42,7 @@ Legend:
| ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
| EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
| FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
| FLOOR | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
| GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
| GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@ -82,6 +84,7 @@ Legend:
| ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
| ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
| ROUND | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@ -108,5 +111,6 @@ Legend:
| TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
| TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
| TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
| TRUNC | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
| XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

16
docs/ops/CPU.csv
View File

@ -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"
@ -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"

Can't render this file because it is too large.

44
ggml/include/ggml.h
View File

@ -577,6 +577,10 @@ extern "C" {
GGML_UNARY_OP_EXP,
GGML_UNARY_OP_GELU_ERF,
GGML_UNARY_OP_XIELU,
GGML_UNARY_OP_FLOOR,
GGML_UNARY_OP_CEIL,
GGML_UNARY_OP_ROUND,
GGML_UNARY_OP_TRUNC,
GGML_UNARY_OP_COUNT,
};
@ -1151,6 +1155,46 @@ 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);
// xIELU activation function
// x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
// where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions

4
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -2184,6 +2184,10 @@ 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;

16
ggml/src/ggml-cpu/ops.cpp
View File

@ -8993,6 +8993,22 @@ 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;
case GGML_UNARY_OP_XIELU:
{
ggml_compute_forward_xielu(params, dst);

32
ggml/src/ggml-cpu/unary-ops.cpp
View File

@ -73,6 +73,22 @@ 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;
@ -274,6 +290,22 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
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);
}
void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
const float alpha_n = ggml_get_op_params_f32(dst, 1);
const float alpha_p = ggml_get_op_params_f32(dst, 2);

4
ggml/src/ggml-cpu/unary-ops.h
View File

@ -22,6 +22,10 @@ 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);
void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus

62
ggml/src/ggml.c
View File

@ -1144,9 +1144,13 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"EXP",
"GELU_ERF",
"XIELU",
"FLOOR",
"CEIL",
"ROUND",
"TRUNC",
};
static_assert(GGML_UNARY_OP_COUNT == 16, "GGML_UNARY_OP_COUNT != 16");
static_assert(GGML_UNARY_OP_COUNT == 20, "GGML_UNARY_OP_COUNT != 20");
static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
"REGLU",
@ -2749,6 +2753,62 @@ 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(
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,