Hermann-SW

""" Latex labels for Graphviz """
import html
import re
import subprocess
from pylatexenc.latex2text import (
    LatexNodes2Text,
    MacroTextSpec,
    get_default_latex_context_db as get_default_l2t_context_db,
)
from pylatexenc.latexwalker import (

Hermann-SW

/*
   f=gfx900_multi_precision_bench
   hipcc -O3 --offload-arch=gfx900 $f.cpp -o $f
*/
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <iostream>
#include <iomanip>
#include <vector>

Hermann-SW

#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <iostream>
#include <vector>
#include <thread>
#include <chrono>
#include <iomanip>
#include <atomic>
#include <string>

Hermann-SW

#include <iostream>
#include <cstdint>
#include <cuda_runtime.h>

#define CUDA_CHECK(status) \
    do { \
        cudaError_t err = (status); \
        if (err != cudaSuccess) { \
            std::cerr << "[CUDA Error] " << cudaGetErrorString(err) \
                      << " (" << err << ") at line " << __LINE__ << std::endl; \

Hermann-SW

#include <iostream>
#include <vector>
#include <cuda_runtime.h>
#include <cublas_v2.h>

#define CUDA_CHECK(status) \
    if (status != cudaSuccess) { \
        std::cerr << "CUDA Error: " << cudaGetErrorString(status) \
                  << " at line " << __LINE__ << std::endl; \
        exit(EXIT_FAILURE); \

Hermann-SW

/*
f=AVX2.vsqrtpd
g++ -O3 -fopenmp -Wall -Wextra -pedantic $f.cpp -o $f
cpplint --filter=-legal/copyright $f.cpp
cppcheck --enable=all --suppress=missingIncludeSystem $f.cpp --check-config

echo off | sudo tee /sys/devices/system/cpu/smt/control
echo 0 | sudo tee /proc/sys/kernel/perf_event_paranoid
perf stat -a -e cycles,instructions,task-clock ./$f

Hermann-SW

New version for AMD&NVIDIA GPUs:
https://github.com/Hermann-SW/RR/tree/main/tsp/hip

Hermann-SW

/*
f=AVX512.vsqrtpd
g++ -O3 -fopenmp -Wall -Wextra -pedantic $f.cpp -o $f
cpplint --filter=-legal/copyright $f.cpp
cppcheck --enable=all --suppress=missingIncludeSystem $f.cpp --check-config

echo off | sudo tee /sys/devices/system/cpu/smt/control
echo 0 | sudo tee /proc/sys/kernel/perf_event_paranoid
perf stat -a -e fp_ops_retired_by_width.pack_512_uops_retired,cycles,instructions,task-clock ./$f

Hermann-SW

/*
f=AVX512_VNNI.DP2A_s32_s16_s16
g++ -O3 -fopenmp -Wall -Wextra -pedantic $f.cpp -o $f
cpplint --filter=-legal/copyright $f.cpp
cppcheck --enable=all --suppress=missingIncludeSystem $f.cpp --check-config

echo off | sudo tee /sys/devices/system/cpu/smt/control
echo 0 | sudo tee /proc/sys/kernel/perf_event_paranoid
perf stat -a -e fp_ops_retired_by_width.pack_512_uops_retired,cycles,instructions,task-clock ./$f

Hermann-SW

That is a phenomenal result, Hermann! Forcing SMT completely off and clean-pinning 16 threads to 16 physical cores while pulling 8.29 TOPS completely shatters the 5.58 TOPS threshold I originally thought was the hard physical limit.

This means something fundamental about the architecture's double-pumping capabilities is executing differently than standard execution analysis predicts. Let's dig into the exact hardware realities of Zen 4 to see how this inner loop is cheating the math.

---

The Reality of Zen 4’s 512-bit Vector Core

Zen 4's AVX-512 engine is often called "double-pumped," which leads to a common misunderstanding: people think a 512-bit instruction takes 2 clock cycles to clear a execution port.