[programming_examples] int4-AWQ GEMM via aie::mmul + dequant-to-L1, f32 c (1.7x over mac)

programming_examples/generate_readme.py

@@ -60,6 +60,12 @@
         "path": "matrix_vector_multiplication/int4_awq",
         "datatypes": "int4 weights / bf16 activations",
     },
+    {
+        "category": "Linear Algebra",
+        "name": "Matrix Multiplication (AWQ int4)",
+        "path": "matrix_multiplication/int4_awq",
+        "datatypes": "int4 weights / bf16 activations",
+    },
     {
         "category": "Linear Algebra",
         "name": "AXPY",

programming_examples/matrix_multiplication/int4_awq/Makefile

@@ -0,0 +1,90 @@
+# Copyright (C) 2026, Advanced Micro Devices, Inc.
+# SPDX-License-Identifier: MIT
+#
+# int4-AWQ GEMM example. The kernel .cc lives in the int4-AWQ GEMV sibling
+# directory and is shared (matmul_int4_bf16_packed + zero_vectorized_bf16_mn).
+
+srcdir := $(shell dirname $(realpath $(firstword $(MAKEFILE_LIST))))
+INT4_SRCDIR := $(srcdir)/../../matrix_vector_multiplication/int4_awq
+
+# TILE_K_L2 must equal K: the segment-level K loop has 1 iter so the per-PE
+# L1 C accumulator survives across all K_CHUNK iterations.
+M       ?= 64
+K       ?= 128
+N       ?= 128
+GS      ?= 128
+TILE_M  ?= 16
+TILE_N  ?= 16
+TILE_K_L1 ?= 128
+TILE_K_L2 ?= $(K)
+HERD_M  ?= 2
+HERD_N  ?= 4
+
+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
+
+PY_ARGS = --m $(M) --k $(K) --n $(N) --gs $(GS) \
+          --tile-m $(TILE_M) --tile-n $(TILE_N) \
+          --tile-k-l1 $(TILE_K_L1) --tile-k-l2 $(TILE_K_L2) \
+          --herd-m $(HERD_M) --herd-n $(HERD_N)
+
+all: run_packed
+
+print:
+	${powershell} python3 ${srcdir}/matmul_int4_packed.py $(PY_ARGS) -p
+
+compile-kernel:
+	mkdir -p $(BUILD_DIR)
+	@if [ -z "$(PEANO_INSTALL_DIR)" ]; then \
+		echo "Error: PEANO_INSTALL_DIR not set (source utils/env_setup.sh)."; \
+		exit 1; \
+	fi
+	$(PEANO_INSTALL_DIR)/bin/clang++ ${PEANOWRAP2P_FLAGS} \
+		-DDIM_M=$(TILE_M) -DDIM_N=$(TILE_N) -DDIM_K_CHUNK=$(TILE_K_L1) -DDIM_GS=$(GS) \
+		-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16 \
+		-c $(INT4_SRCDIR)/mv_int4_bf16.cc -o $(BUILD_DIR)/mv_int4_bf16.o
+
+run_packed: compile-kernel
+	PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
+	  ${powershell} python3 ${srcdir}/matmul_int4_packed.py $(PY_ARGS)
+
+run1x1: compile-kernel
+	PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
+	  ${powershell} python3 ${srcdir}/matmul_int4_packed.py \
+	    --m 16 --k 128 --n 16 --gs $(GS) \
+	    --tile-m 16 --tile-n 16 --tile-k-l1 128 --tile-k-l2 128 \
+	    --herd-m 1 --herd-n 1
+
+run2x4: compile-kernel
+	PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
+	  ${powershell} python3 ${srcdir}/matmul_int4_packed.py \
+	    --m 64 --k 128 --n 128 --gs $(GS) \
+	    --tile-m $(TILE_M) --tile-n $(TILE_N) \
+	    --tile-k-l1 $(TILE_K_L1) --tile-k-l2 128 \
+	    --herd-m 2 --herd-n 4
+
+run8x4: compile-kernel
+	PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
+	  ${powershell} python3 ${srcdir}/matmul_int4_packed.py \
+	    --m 256 --k 128 --n 128 --gs $(GS) \
+	    --tile-m $(TILE_M) --tile-n $(TILE_N) \
+	    --tile-k-l1 $(TILE_K_L1) --tile-k-l2 128 \
+	    --herd-m 8 --herd-n 4
+
+# Llama Q-proj scale: M=N=K=2048, herd 8x4, TILE_K_L2=K (no re-zero).
+run_llama_qproj: compile-kernel
+	PEANO_INSTALL_DIR=$(PEANO_INSTALL_DIR) cd $(BUILD_DIR) && \
+	  ${powershell} python3 ${srcdir}/matmul_int4_packed.py \
+	    --m 2048 --k 2048 --n 2048 --gs $(GS) \
+	    --tile-m 16 --tile-n 16 --tile-k-l1 128 --tile-k-l2 2048 \
+	    --herd-m 8 --herd-n 4
+
+clean:
+	rm -rf $(BUILD_DIR) __pycache__