#pragma once

#include <sys/time.h>
#include "dfatool_host_dpu.ah"

aspect DfatoolHostTiming : public DfatoolHostDPUTiming {

	unsigned int n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile;
	unsigned int element_size;

	virtual int getKernel() { return kernel; }

	DfatoolHostTiming() {
		element_size = sizeof(T);
	}

	advice call("% input_params(...)") : after() {
		Params* p = tjp->result();
		/*
		 * Input: (M_ * m) × (N_ * n) matrix
		 */
		n_rows_outer = p->M_;
		n_rows_tile = p->m;
		n_cols_outer = p->N_;
		n_cols_tile = p->n;
		printf("[>>] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
	}

	advice call("% trns_host(...)") : after() {
		printf("[--] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
	}

	advice execution("% main(...)") : after() {
		printf("[<<] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
	}
};