Implementation:Ggml org Ggml Hexagon binary ops

Implementation Metadata

File Name	src/ggml-hexagon/htp/binary-ops.c
Repository	ggml-org/ggml
Lines	343
Language	C
Domain Tags	ML_Infrastructure, DSP_Computing, Element_Wise_Operations
Status	Active
Last Updated	2025-05-15 12:00 GMT
Knowledge Sources	ggml-org/ggml repository

Overview

binary-ops.c is the DSP-side implementation of element-wise binary operations (multiply, add, subtract) on the Hexagon HVX vector processor. These operations are used throughout ML inference for residual connections, scaling, and masking.

Description

The file uses function pointer tables (func_table_HVX and func_table_HVX_opt) to dispatch to HVX-optimized variants. The optimized variants (hvx_mul_f32_aa, hvx_add_f32_aa, hvx_sub_f32_aa) are selected when both source and destination buffers are 128-byte aligned and row sizes are VLEN-aligned.

The binary_job_f32_per_thread function parallelizes work by distributing source rows across threads. It handles broadcasting by computing source row indices via modular arithmetic when source and destination dimensions differ.

The htp_binary_preamble macro extracts all tensor dimensions (ne00-ne13, ne0-ne3), strides (nb00-nb13, nb0-nb3), and per-thread row counts from the operation context.

Usage

These operations are dispatched from the DSP-side message loop when the host sends binary operation requests (GGML_OP_ADD, GGML_OP_MUL, GGML_OP_SUB).

Code Reference

Source Location

Repository	File	Lines
ggml-org/ggml	src/ggml-hexagon/htp/binary-ops.c	343

Key Signatures

typedef void (*hvx_elemwise_f32_func)(uint8_t * data_dst, const uint8_t * src0,
    const uint8_t * src1, const uint32_t num_elems);

static hvx_elemwise_f32_func func_table_HVX[]     = { hvx_mul_f32, hvx_add_f32, hvx_sub_f32 };
static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_aa, hvx_add_f32_aa, hvx_sub_f32_aa };

static void binary_job_f32_per_thread(struct htp_ops_context * octx,
    uint8_t * spad_data, uint32_t nth, uint32_t ith, enum htp_op op);

I/O Contract

Inputs

• octx -- Operation context containing source tensors, destination tensor, and per-thread row counts
• spad_data -- Scratchpad memory for intermediate storage
• nth, ith -- Total thread count and current thread index
• op -- Binary operation type (multiply, add, subtract)

Outputs

• dst tensor -- Result of element-wise binary operation written to destination buffer

Usage Examples

Internal dispatch:

// Called from main.c op_binary() dispatch
binary_job_f32_per_thread(octx, spad_data, nth, ith, HTP_OP_MUL);
binary_job_f32_per_thread(octx, spad_data, nth, ith, HTP_OP_ADD);
binary_job_f32_per_thread(octx, spad_data, nth, ith, HTP_OP_SUB);

Related Pages

Implements Principle

• Principle:Ggml_org_Ggml_Hexagon_DSP_Computation

Related Implementations

• Implementation:Ggml_org_Ggml_Hexagon_hvx_arith -- HVX arithmetic intrinsics used by this file
• Implementation:Ggml_org_Ggml_Hexagon_htp_main -- Message dispatcher
• Implementation:Ggml_org_Ggml_Hexagon_backend -- Host-side backend