sleepy/llama.cpp
1
0
Fork 0
Code Issues 11 Pull Requests Actions 114 Packages Projects Releases Wiki Activity

Add experimental ggml-hexagon backend for the Hexagon NPU (#16547)

Browse Source 
* model: add support for extra bufs for all devices

* hexagon: add experimental ggml-hexagon backend for the Hexagon NPU

This commit introduces a new experimental backend `ggml-hexagon` with support for the Hexagon NPU.

Highlights:
- Supports Hexagon versions: v73, v75, v79, and v81
- Targets Android devices based on Snapdragon SoCs: Gen3, 8-Elite, and 8-Elite Gen5
- Supports Q4_0, Q8_0, MXFP4, and FP32 data types
- Implements core LLM ops: MUL_MAT/MUL_MAT_ID, ADD/SUB/MUL/ADD_ID, RMS_NORM, ROPE, GLU/SWIGLU, SOFTMAX

**Note:** This backend is experimental and may exhibit instability or limited performance across supported devices.
It is intended for early testing and feedback from llama.cpp/ggml developer and user community.

Co-Authored-By: Rajdeep Ganguly <rganguly@qti.qualcomm.com>
Co-Authored-By: Todor Boinovski <todorb@qti.qualcomm.com>

* hexagon: fix format checker errors

* hexagon: update readme and cmake presets

* ci: add android-ndk-build jobs that build plain ARM64 and Snapdragon versions

* hexagon: add simple graph optimizer for stacking MUL_MAT ops with the same input

* hexagon: move ADB helper scripts into scripts/snapdragon/adb

* hexagon: replace all f/printfs with GGML_LOG_...

* readme: add hexagon to the list supported backends

* hexagon: stack malmuts with quantized inputs only

* hexagon: add TODO for fixing issues in hexagon_graph_optimize

* hexagon: update to hex-sdk 6.4.0 and add scripts for running on QDC

* scripts: fix lint errors

* scripts: update qdc pytest script to make linter happy

* hexagon: add reduce sum in fp32

* hexagon: reduce number of vector stores in matmul output

* hexagon: remove the need for vdelta in reduce-multiply-x8

* hexagon: consistent use of reduce_sum_fp32 for row_sums

* hexagon: some more matmul optimizations and comments

Optimize cases where tensor dims are not multiple of 1024 (e.g in Qwen models).
We've handled those cases already but at a higher overhead.

* hexagon: update cmake presets

* hexagon: add OPMASK support for run-bench.sh wrapper

* hexagon: update to use GGML_BACKEND_API

* hexagon: remove unused logic for setting tensor flags for the views

* hexagon: add asserts to set/get_tensor to make sure we handle complete tensors

Same asserts as the CPU backend.

* hexagon: use cpy_tensor slow path for non-host buffers

* hexagon: error checks in the buffer allocator

* cmake: move include(extProj) under ggml-hexagon

* hexagon: don't forget to delete the backend on free

* hexagon: set/get_tensor size assert apply only to quantized tensors

* hexagon: reintroduce HEX_VERBOSE wrapper for GGML_LOG_DEBUG for now

GGML_LOG_DEBUG is always enabled for test-backend-ops and the output gets in the way.
Ideally we need a bit more finer log levels.

* docs: typos in hexagon developer docs (libggm-...)

* hexagon: overhaul error handling in the session/device allocation

this should handle all failure paths in the session allocation.

* hexagon: update cmake presets to enable fp16 vectors

* hexagon: remove unused time_usec function

* hexagon: don't forget to release buffer contexts

* hexagon: fixed indents in hvx-utils (missed clang-format auto-format failure)

* hexagon: remove custom can_repeat function and use ggml_can_repeat

---------

Co-authored-by: Rajdeep Ganguly <rganguly@qti.qualcomm.com>
Co-authored-by: Todor Boinovski <todorb@qti.qualcomm.com>
This commit is contained in:
Max Krasnyansky
2025-10-22 13:47:09 -07:00
committed by GitHub
parent a2e0088d92
commit 63d2fc46e1
45 changed files with 13530 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/htp/htp-msg.h
+156
View File
@@ -0,0 +1,156 @@
#ifndef HTP_MSG_H
#define HTP_MSG_H
#include <assert.h>
// ggml-common.h must be included prio to this header
// Mask to enable various stages of the Ops.
// Used for debugging and profiling.
enum {
HTP_OPMASK_QUEUE = (1 << 0), // Enable Queueing (ie calls into the DSP)
HTP_OPMASK_QUANTIZE = (1 << 1), // Enable Quantize
HTP_OPMASK_COMPUTE = (1 << 2), // Enable Compute
};
// Op flags
enum {
HTP_OPFLAGS_SKIP_QUANTIZE = (1 << 0), // Skip dynamic quantization (reuse quantized tensors)
HTP_OPFLAGS_SKIP_COMPUTE = (1 << 1), // Skip actual computation (used for profiling)
HTP_OPFLAGS_EARLY_WAKEUP = (1 << 2) // Send early wakeup notification
};
enum htp_status {
HTP_STATUS_OK = 1,
HTP_STATUS_INTERNAL_ERR = 2,
HTP_STATUS_NO_SUPPORT = 3,
HTP_STATUS_INVAL_PARAMS = 4,
HTP_STATUS_VTCM_TOO_SMALL = 5,
};
// The values must match the ggml_type.
// Duplicated here because we can't include full ggml.h in the htp build.
// We have some static_asserts in the cpp code to ensure things are in sync.
enum htp_data_type {
HTP_TYPE_F32 = 0,
HTP_TYPE_F16 = 1,
HTP_TYPE_Q4_0 = 2,
HTP_TYPE_Q8_0 = 8,
HTP_TYPE_MXFP4 = 39,
HTP_TYPE_COUNT
};
// These values are manually translated over to HTP
// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
enum htp_op {
HTP_OP_MUL = 0,
HTP_OP_ADD = 1,
HTP_OP_SUB = 2,
HTP_OP_DIV = 3,
HTP_OP_MUL_MAT = 4,
HTP_OP_MUL_MAT_ID = 5,
HTP_OP_RMS_NORM = 6,
HTP_OP_UNARY_SILU = 7,
HTP_OP_GLU_SWIGLU = 8,
HTP_OP_GLU_SWIGLU_OAI = 9,
HTP_OP_SOFTMAX = 10,
HTP_OP_ADD_ID = 11,
HTP_OP_ROPE = 12,
INVALID
};
static inline size_t htp_type_block_size(uint32_t t) {
switch (t) {
case HTP_TYPE_F32:
return 1;
case HTP_TYPE_F16:
return 1;
case HTP_TYPE_Q4_0:
return QK4_0;
case HTP_TYPE_Q8_0:
return QK8_0;
case HTP_TYPE_MXFP4:
return QK_MXFP4;
default:
assert(0 && "unsupported HTP data type");
}
return 0;
}
static inline size_t htp_type_nbytes(uint32_t t) {
switch (t) {
case HTP_TYPE_F32:
return 4;
case HTP_TYPE_F16:
return 2;
case HTP_TYPE_Q4_0:
return sizeof(block_q4_0);
case HTP_TYPE_Q8_0:
return sizeof(block_q8_0);
case HTP_TYPE_MXFP4:
return sizeof(block_mxfp4);
default:
assert(0 && "unsupported HTP data type");
}
return 0;
}
static const char * htp_type_name(uint32_t t) {
switch (t) {
case HTP_TYPE_F32:
return "fp32";
case HTP_TYPE_F16:
return "fp16";
case HTP_TYPE_Q4_0:
return "q4_0";
case HTP_TYPE_Q8_0:
return "q8_0";
case HTP_TYPE_MXFP4:
return "mxfp4";
}
return 0;
}
// Internal types
#define QK_Q4_0x4x2 256 // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
#define QK_Q8_0x4x2 256 // 4x Q8_0 blocks concat with next 4x Q8_0 blocks
#define QK_MXFP4x4x2 256 // 4x MXFP4 blocks concat with next 4x MXFP4 blocks
#define HTP_MAX_DIMS 4
struct htp_tensor {
uint32_t data; // Buffer offset in the messages, and data pointer on the NSP
uint32_t type; // Data type
uint32_t ne[HTP_MAX_DIMS]; // Number of elements
uint32_t nb[HTP_MAX_DIMS]; // Stride in bytes (see ggml.h ggml_tensor)
};
#define HTP_MAX_OP_PARAMS 64
struct htp_general_req {
uint32_t op; // GGML/HTP Op
int32_t op_params[HTP_MAX_OP_PARAMS / sizeof(int32_t)];
// Params for the op, e.g. epsilon of RMS norm
uint32_t flags; // Request flags
struct htp_tensor src0; // Input0 tensor
struct htp_tensor src1; // Input1 tensor
struct htp_tensor src2; // Input2 tensor
struct htp_tensor dst; // Output tensor
// should be multiple of 64 bytes (cacheline)
};
struct htp_general_rsp {
uint32_t op; // GGML/HTP Op
uint32_t status; // HTP_STATUS_...
uint32_t prof_usecs; // Number of usec per request
uint32_t prof_cycles; // Number of cycles per request
uint32_t prof_pkts; // Number of instruction packets per request
uint8_t unused[44]; // Pad to 64 bytes
};
#define HTP_MAX_MESSAGE_SIZE sizeof(struct htp_general_req)
#define HTP_MAX_PACKET_BUFFERS 4
#endif /* HTP_MSG_H */