Implementation:Ggml org Ggml Hexagon htp main

Implementation Metadata

File Name	src/ggml-hexagon/htp/main.c
Repository	ggml-org/ggml
Lines	1043
Language	C
Domain Tags	DSP_Computing, Runtime_Dispatch, Hardware_Abstraction
Status	Active
Last Updated	2025-05-15 12:00 GMT
Knowledge Sources	ggml-org/ggml repository

Overview

main.c is the entry point for the DSP-side HTP skeleton, implementing the IDL interface (htp_iface_open/close) and the dspqueue message processing loop that dispatches tensor operations on the Hexagon processor. This is the DSP-side runtime core where host requests arrive and get executed on Hexagon hardware.

Description

The htp_iface_open function allocates an htp_context, configures DCVS power settings to maximum performance, powers on HVX and HMX coprocessors, and sets the client class to compute. The context is used as the remote handle.

The message processing loop reads htp_general_req messages from the dspqueue and dispatches them to operation functions based on the op code:

• op_matmul -- Matrix multiplication
• op_binary -- Element-wise add/mul/sub
• op_softmax -- Softmax with optional ALiBi
• op_rope -- Rotary position embedding
• op_act -- Activation functions (SwiGLU)
• op_flash_attn -- Flash attention

The runtime manages VTCM (Vector Tightly Coupled Memory) allocation, worker pool initialization, and DMA queue setup for high-performance data movement.

Usage

This code runs exclusively on the Hexagon DSP. It is loaded as a skeleton library by the FastRPC runtime when the host opens a session.

Code Reference

Source Location

Repository	File	Lines
ggml-org/ggml	src/ggml-hexagon/htp/main.c	1043

Key Signatures

// IDL interface entry points
AEEResult htp_iface_open(const char * uri, remote_handle64 * handle);
AEEResult htp_iface_close(remote_handle64 handle);

// Power configuration -- sets DCVS to max performance
HAP_power_request_t request;
request.dcvs_v3.dcvs_option = HAP_DCVS_V2_PERFORMANCE_MODE;
request.dcvs_v3.bus_params.min_corner = HAP_DCVS_VCORNER_MAX;

I/O Contract

Inputs

• dspqueue messages -- htp_general_req containing operation type, tensor descriptors, and parameters
• Shared memory buffers -- Tensor data in rpcmem-allocated regions

Outputs

• dspqueue responses -- htp_general_rsp with status, timing, and profiling data
• Computed tensor data -- Results written to shared memory for host access

Usage Examples

DSP-side lifecycle:

// 1. Host opens the remote handle
AEEResult err = htp_iface_open(uri, &handle);
// -> Allocates context, powers on HVX/HMX, sets max performance

// 2. Host sends operations via dspqueue
// -> Message loop dispatches to op_matmul, op_binary, op_softmax, etc.

// 3. Host closes the handle
htp_iface_close(handle);
// -> Frees context, releases resources

Related Pages

Implements Principle

• Principle:Ggml_org_Ggml_Hexagon_DSP_Computation

Related Implementations

• Implementation:Ggml_org_Ggml_Hexagon_backend -- Host-side counterpart
• Implementation:Ggml_org_Ggml_Hexagon_matmul_ops -- Matrix multiply operations
• Implementation:Ggml_org_Ggml_Hexagon_flash_attn -- Flash attention operations
• Implementation:Ggml_org_Ggml_Hexagon_softmax_ops -- Softmax operations
• Implementation:Ggml_org_Ggml_Hexagon_act_ops -- Activation operations
• Implementation:Ggml_org_Ggml_Hexagon_binary_ops -- Binary operations
• Implementation:Ggml_org_Ggml_Hexagon_rope_ops -- RoPE operations