Skip to content

Add fused SwiGLU programming example for NPU2 #1511

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
erwei-xilinx wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Add fused SwiGLU programming example for NPU2 #1511

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
282b50c
Add fused SwiGLU programming example for NPU2
erwei-xilinx Apr 8, 2026
7493226
Address review comments on fused SwiGLU example
erwei-xilinx Apr 8, 2026
c6e2e03
Use 8x4 herd (full NPU2 array) in LIT test
erwei-xilinx Apr 8, 2026
72aed36
Enable ping-pong buffering for fused SwiGLU
erwei-xilinx Apr 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
162 changes: 162 additions & 0 deletions programming_examples/ffn_swiglu/fused/Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,162 @@
# Copyright (C) 2026, Advanced Micro Devices, Inc.
# SPDX-License-Identifier: MIT
#
# Fused SwiGLU: output = SiLU(x @ W_gate) * (x @ W_up)
# Single launch with time-multiplexed herds on NPU2 (AIE2P).
#
srcdir := $(shell dirname $(realpath $(firstword $(MAKEFILE_LIST))))

# Tile sizes
TILE_M ?= 64
TILE_K_L2 ?= 128
TILE_K_L1 ?= 32
TILE_N ?= 64

# Problem sizes
M ?= 512
K ?= 512
N ?= 512

# Herd dimensions (4 rows x 4 cols = 16 tiles on NPU2)
HERD_M ?= 4
HERD_N ?= 4

# Derived constants for kernel compilation
TILE_M_DIV_8 := $(shell echo $$(( $(TILE_M) / 8 )))
TILE_N_DIV_8 := $(shell echo $$(( $(TILE_N) / 8 )))

# Output format
OUTPUT_FORMAT ?= elf

# Determine build dir based on whether PEANO_INSTALL_DIR is set
ifdef PEANO_INSTALL_DIR
BUILD_DIR := build_peano
else
BUILD_DIR := build_chess
endif

AIEOPT_DIR = $(shell realpath $(dir $(shell which aie-opt))/..)
WARNING_FLAGS = -Wno-parentheses -Wno-attributes -Wno-macro-redefined -Wno-empty-body
PEANOWRAP2P_FLAGS = -O2 -std=c++20 --target=aie2p-none-unknown-elf ${WARNING_FLAGS} -DNDEBUG -I ${AIEOPT_DIR}/include

COMPILE_MODE ?= compile-and-run

all: run

print:
${powershell} python3 ${srcdir}/swiglu_fused.py -p \
--m $(M) --k $(K) --n $(N) \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m $(HERD_M) --herd-n $(HERD_N)

compile-kernel:
mkdir -p $(BUILD_DIR)
@if [ -n "$(PEANO_INSTALL_DIR)" ]; then \
echo "Compiling swiglu_fused.cc with Peano for AIE2P"; \
$(PEANO_INSTALL_DIR)/bin/clang++ ${PEANOWRAP2P_FLAGS} \
-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16 \
-DDIM_M=$(TILE_M) -DDIM_K=$(TILE_K_L1) -DDIM_N=$(TILE_N) \
-DDIM_M_DIV_8=$(TILE_M_DIV_8) -DDIM_N_DIV_8=$(TILE_N_DIV_8) \
-c ${srcdir}/swiglu_fused.cc \
-o $(BUILD_DIR)/swiglu_fused.o; \
elif command -v xchesscc_wrapper >/dev/null 2>&1; then \
echo "Compiling swiglu_fused.cc with xchesscc for AIE2P"; \
cd $(BUILD_DIR) && ${powershell} xchesscc_wrapper aie2p \
-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16 \
-DDIM_M=$(TILE_M) -DDIM_K=$(TILE_K_L1) -DDIM_N=$(TILE_N) \
-DDIM_M_DIV_8=$(TILE_M_DIV_8) -DDIM_N_DIV_8=$(TILE_N_DIV_8) \
-c ${srcdir}/swiglu_fused.cc \
-o swiglu_fused.o; \
else \
echo "Error: Neither PEANO_INSTALL_DIR nor xchesscc_wrapper found."; \
exit 1; \
fi

run: compile-kernel
mkdir -p $(BUILD_DIR)/air_project
cp $(BUILD_DIR)/swiglu_fused.o $(BUILD_DIR)/air_project/swiglu_fused.o
PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
${powershell} python3 ${srcdir}/swiglu_fused.py \
--m $(M) --k $(K) --n $(N) \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m $(HERD_M) --herd-n $(HERD_N) \
--output-format $(OUTPUT_FORMAT) \
--compile-mode $(COMPILE_MODE)

# Smaller config for quick testing / CI
run4x4: compile-kernel
mkdir -p $(BUILD_DIR)/air_project
cp $(BUILD_DIR)/swiglu_fused.o $(BUILD_DIR)/air_project/swiglu_fused.o
PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
${powershell} python3 ${srcdir}/swiglu_fused.py \
--m 256 --k 256 --n 256 \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m 4 --herd-n 4 \
--output-format $(OUTPUT_FORMAT) \
--compile-mode $(COMPILE_MODE)

# Compile-only (no XRT needed)
compile-only: compile-kernel
mkdir -p $(BUILD_DIR)/air_project
cp $(BUILD_DIR)/swiglu_fused.o $(BUILD_DIR)/air_project/swiglu_fused.o
PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
${powershell} python3 ${srcdir}/swiglu_fused.py \
--m $(M) --k $(K) --n $(N) \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m $(HERD_M) --herd-n $(HERD_N) \
--compile-mode compile-only --output-format none

# Compile xclbin (for profile / C++ test usage)
compile-xclbin: compile-kernel
mkdir -p $(BUILD_DIR)/air_project
cp $(BUILD_DIR)/swiglu_fused.o $(BUILD_DIR)/air_project/swiglu_fused.o
PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
${powershell} python3 ${srcdir}/swiglu_fused.py \
--m $(M) --k $(K) --n $(N) \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m $(HERD_M) --herd-n $(HERD_N) \
--compile-mode compile-only --output-format xclbin

# Profile: compile + run with Python-based timing
# Usage: make profile [M=...] [K=...] [N=...]
profile: compile-kernel
mkdir -p $(BUILD_DIR)/air_project
cp $(BUILD_DIR)/swiglu_fused.o $(BUILD_DIR)/air_project/swiglu_fused.o
PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
${powershell} python3 ${srcdir}/swiglu_fused.py \
--m $(M) --k $(K) --n $(N) \
--tile-m $(TILE_M) --tile-k-l2 $(TILE_K_L2) --tile-k-l1 $(TILE_K_L1) --tile-n $(TILE_N) \
--herd-m $(HERD_M) --herd-n $(HERD_N) \
--compile-mode profile

build-test-exe:
@GPP=$$( \
for bin in /usr/bin/g++-*; do \
ver=$$(echo $$bin | grep -oE '[0-9]+$$'); \
if [ "$$ver" -ge 13 ] 2>/dev/null; then \
echo "$$ver $$bin"; \
fi; \
done | sort -nr | head -n1 | awk '{print $$2}' \
); \
if [ -z "$$GPP" ]; then \
echo "Error: No g++ version >= 13 found in /usr/bin."; \
exit 1; \
fi; \
if [ -z "$$XILINX_XRT" ]; then \
echo "Error: XILINX_XRT environment variable not set. Please source xrt/setup.sh."; \
exit 1; \
fi; \
if [ -z "$(AIEOPT_DIR)" ]; then \
echo "Error: aie-opt not found on PATH. Please source utils/env_setup.sh."; \
exit 1; \
fi; \
echo "Using compiler: $$GPP"; \
mkdir -p $(BUILD_DIR); \
cd $(BUILD_DIR) && $$GPP ${srcdir}/test.cpp -o test.exe -std=c++23 -Wall \
-I$$XILINX_XRT/include -L$$XILINX_XRT/lib \
-I$(AIEOPT_DIR)/runtime_lib/x86_64/test_lib/include \
-L$(AIEOPT_DIR)/runtime_lib/x86_64/test_lib/lib \
-luuid -lxrt_coreutil -lrt -lstdc++ -ltest_utils

clean:
rm -rf build_peano build_chess __pycache__
5 changes: 5 additions & 0 deletions programming_examples/ffn_swiglu/fused/run_makefile_peano.lit
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
// RUN: make -f %S/Makefile clean PEANO_INSTALL_DIR=%PEANO_INSTALL_DIR
// RUN: make -f %S/Makefile run PEANO_INSTALL_DIR=%PEANO_INSTALL_DIR HERD_M=8 HERD_N=4 COMPILE_MODE=compile-and-run OUTPUT_FORMAT=elf 2>&1 | FileCheck %s
// CHECK: PASS!

// REQUIRES: ryzen_ai_npu2, peano
Loading
Loading