// SPDX-License-Identifier: BSD-3-Clause
/*
 * victory_atomics.cu - global atomicAdd/atomicMax histogram over 4 MB.
 *
 * A deterministic 4M-byte buffer (every value 0..255) is histogrammed on the
 * device into a 256-bin counter array with atomicAdd, while a global maximum
 * is tracked with atomicMax. The host recomputes the exact histogram and
 * maximum and verifies every bin, the max, and the total count.
 *
 * Output lines are stable and greppable:
 *
 *   DEVICE <name> cc=<major>.<minor> sm=<count> mem=<bytes>
 *   ATOMICS n=<count> bins=256 max=<value> bad=<mismatches> total=<sum>
 *   STATUS OK|FAIL
 *
 * Build with a static CUDA runtime (--cudart=static) so the guest only needs
 * the driver's libcuda.so, not a CUDA toolkit installation:
 *
 *   nvcc -O2 --cudart=static -o victory_atomics victory_atomics.cu
 *
 * Then, from the guest:  /tmp/victory_atomics
 */

#include <cstdio>
#include <cstdlib>
#include <cstdint>
#include <cstring>
#include <cuda_runtime.h>

#define CHECK(x)                                                         \
	do {                                                             \
		cudaError_t e_ = (x);                                    \
		if (e_ != cudaSuccess) {                                 \
			printf("CUDA_ERROR %s at line %d\n",             \
			       cudaGetErrorString(e_), __LINE__);        \
			printf("STATUS FAIL\n");                         \
			return 1;                                        \
		}                                                        \
	} while (0)

#define BINS 256

__global__ void histogram(const unsigned char *data, unsigned int *hist,
			  unsigned int *gmax, int n)
{
	int i = blockIdx.x * blockDim.x + threadIdx.x;

	if (i < n) {
		unsigned int v = data[i];

		atomicAdd(&hist[v], 1u);
		atomicMax(gmax, v);
	}
}

int main(void)
{
	cudaDeviceProp prop;
	const int n = 1 << 22; /* 4M bytes */
	unsigned char *h_data;
	unsigned int *h_hist;
	unsigned int h_max = 0, h_total = 0;
	unsigned char *d_data;
	unsigned int *d_hist, *d_max;
	unsigned int ref[BINS];
	unsigned int ref_max = 0;
	int bad = 0, i;

	setvbuf(stdout, NULL, _IONBF, 0);

	CHECK(cudaGetDeviceProperties(&prop, 0));
	printf("DEVICE %s cc=%d.%d sm=%d mem=%llu\n", prop.name,
	       prop.major, prop.minor, prop.multiProcessorCount,
	       (unsigned long long)prop.totalGlobalMem);

	h_data = (unsigned char *)malloc(n);
	h_hist = (unsigned int *)malloc(BINS * sizeof(unsigned int));
	if (!h_data || !h_hist)
		return 1;

	for (i = 0; i < n; i++)
		h_data[i] = (unsigned char)(((long long)i * 37 + i / 256) & 0xFF);

	memset(ref, 0, sizeof(ref));
	for (i = 0; i < n; i++) {
		ref[h_data[i]]++;
		if (h_data[i] > ref_max)
			ref_max = h_data[i];
	}

	CHECK(cudaMalloc(&d_data, n));
	CHECK(cudaMalloc(&d_hist, BINS * sizeof(unsigned int)));
	CHECK(cudaMalloc(&d_max, sizeof(unsigned int)));
	CHECK(cudaMemcpy(d_data, h_data, n, cudaMemcpyHostToDevice));
	CHECK(cudaMemset(d_hist, 0, BINS * sizeof(unsigned int)));
	CHECK(cudaMemset(d_max, 0, sizeof(unsigned int)));

	histogram<<<(n + 255) / 256, 256>>>(d_data, d_hist, d_max, n);
	CHECK(cudaGetLastError());
	CHECK(cudaDeviceSynchronize());
	CHECK(cudaMemcpy(h_hist, d_hist, BINS * sizeof(unsigned int),
			 cudaMemcpyDeviceToHost));
	CHECK(cudaMemcpy(&h_max, d_max, sizeof(unsigned int),
			 cudaMemcpyDeviceToHost));

	for (i = 0; i < BINS; i++) {
		if (h_hist[i] != ref[i])
			bad++;
		h_total += h_hist[i];
	}
	if (h_max != ref_max)
		bad++;

	printf("ATOMICS n=%d bins=256 max=%u bad=%d total=%u\n", n, h_max, bad,
	       h_total);

	if (bad != 0 || h_total != (unsigned int)n) {
		printf("STATUS FAIL\n");
		return 1;
	}

	printf("STATUS OK\n");
	CHECK(cudaFree(d_data));
	CHECK(cudaFree(d_hist));
	CHECK(cudaFree(d_max));
	free(h_data);
	free(h_hist);
	return 0;
}