code/hamer/kernels/cmm/fused_compute.cpp · changh95/hamer-p150 at main

hamer-p150 / code /hamer /kernels /cmm /fused_compute.cpp

Optimized build (2026-10-03): model call 3.49 ms, device 3.36 ms

58ba6df verified about 1 month ago

535 Bytes

// SPDX-FileCopyrightText: © 2026 Tenstorrent Inc.
// SPDX-License-Identifier: Apache-2.0
// Round 8: cmm compute (matmul at the explicit fidelity CMM_MM_FID) then the add+LN compute (this kernel's
// MATH_FIDELITY, HiFi4) on the cmm output, which never leaves its CB (ADDLN_C_Y = the cmm c_out).
#define kernel_main cmm_compute_main
#include "compute.cpp"
#undef kernel_main
#define ADDLN_KERNEL_MAIN void addln_compute_main()
#include "../addln/compute.cpp"

void kernel_main() {
    cmm_compute_main();
    addln_compute_main();
}

	// SPDX-FileCopyrightText: © 2026 Tenstorrent Inc.
	// SPDX-License-Identifier: Apache-2.0
	// Round 8: cmm compute (matmul at the explicit fidelity CMM_MM_FID) then the add+LN compute (this kernel's
	// MATH_FIDELITY, HiFi4) on the cmm output, which never leaves its CB (ADDLN_C_Y = the cmm c_out).
	#define kernel_main cmm_compute_main
	#include "compute.cpp"
	#undef kernel_main
	#define ADDLN_KERNEL_MAIN void addln_compute_main()
	#include "../addln/compute.cpp"

	void kernel_main() {
	cmm_compute_main();
	addln_compute_main();
	}