Add OpenAI SwiGLU benchmarks

src/CMakeLists.txt

@@ -548,6 +548,7 @@ set(HECBENCH_POC_BENCHMARKS
     silu
     softmax-fused
     softmax-online
+    swiglu-oai
     sundials
     ssm
     storeKVCache

src/swiglu-oai-cuda/CMakeLists.txt

@@ -0,0 +1,8 @@
+# swiglu-oai-cuda/CMakeLists.txt
+
+add_hecbench_benchmark(
+    NAME swiglu-oai
+    MODEL cuda
+    SOURCES main.cu
+    CATEGORIES algorithms
+)

src/swiglu-oai-cuda/Makefile

@@ -0,0 +1,91 @@
+#===============================================================================
+# User Options
+#===============================================================================
+
+# Compiler can be set below, or via environment variable
+CC        = nvcc
+OPTIMIZE  = yes
+DEBUG     = no
+# __nv_fp8_e4m3 (cuda_fp8.h) needs CUDA >= 11.8; the native FP4 conversions
+# (cuda_fp4.h) need CUDA >= 12.8, and older toolkits fall back to the software
+# E2M1/E8M0 codecs in main.cu. FP4 conversion is native on
+# architecture-specific Blackwell targets such as sm_100a; other targets use
+# CUDA's emulation path. sm_90 keeps this benchmark runnable on Hopper while
+# exercising native FP8 support.
+ARCH      = sm_90
+LAUNCHER  ?=
+
+# Run options. The default run covers five representative production paths.
+# Override the shape or repeat count on the make command line, for example:
+#   make run ROWS=128 DIM=256 REPEAT=10
+ROWS   ?= 8192
+DIM    ?= 1024
+REPEAT ?= 100
+
+#===============================================================================
+# Program name & source code list
+#===============================================================================
+
+program = main
+
+source = main.cu
+
+obj = $(source:.cu=.o)
+
+#===============================================================================
+# Sets Flags
+#===============================================================================
+
+# Standard Flags
+CFLAGS := $(EXTRA_CFLAGS) -std=c++17 -Xcompiler -Wall -arch=$(ARCH)
+
+# Linker Flags
+LDFLAGS =
+
+# Debug Flags
+ifeq ($(DEBUG),yes)
+  CFLAGS += -g -DDEBUG
+  LDFLAGS  += -g
+endif
+
+# Optimization Flags
+ifeq ($(OPTIMIZE),yes)
+  CFLAGS += -O3
+endif
+#===============================================================================
+# Targets to Build
+#===============================================================================
+
+$(program): $(obj) Makefile
+	$(CC) $(CFLAGS) $(obj) -o $@ $(LDFLAGS)
+
+%.o: %.cu reference.h Makefile
+	$(CC) $(CFLAGS) -c $< -o $@
+
+clean:
+	rm -rf $(program) $(obj)
+
+.PHONY: clean run
+
+#
+# Standard LLM inference: BF16 activation, no bias, no output quantization.
+# Bias-bearing FP16 model.
+# vLLM/SGLang-style fused SwiGLU followed by FP8 activation quantization.
+# Blackwell-style fused SwiGLU followed by MXFP4 activation quantization.
+# Low-precision MoE path with MXFP8 input, FP8 output, and BF16 bias.
+#
+run: $(program)
+	@run_case() { \
+	  output="$$($(LAUNCHER) ./$(program) "$$@" 2>&1)"; status=$$?; \
+	  printf '%s\n' "$$output"; \
+	  if [ $$status -ne 0 ] || \
+	     ! printf '%s\n' "$$output" | awk '$$0 == "PASS" { pass=1 } END { exit !pass }'; then \
+	    return 1; \
+	  fi; \
+	}; \
+	set -e; \
+	run_case $(ROWS) $(DIM) $(REPEAT) bf16 none none; \
+	run_case $(ROWS) $(DIM) $(REPEAT) fp16 none fp16; \
+	run_case $(ROWS) $(DIM) $(REPEAT) bf16 fp8 none; \
+	run_case $(ROWS) $(DIM) $(REPEAT) bf16 mxfp4 none; \
+	run_case $(ROWS) $(DIM) $(REPEAT) mxfp8 fp8 bf16