// SPDX-License-Identifier: BSD-3-Clause
/*
 * transformer.cu - single-head attention forward + 4-layer FFN mini-transformer.
 *
 * Two independent checks in one binary, both verified against an internal
 * double-precision CPU reference (tolerance 1e-3 relative, 1e-6 absolute floor):
 *
 *  1. Attention: X[64x256], Wq/Wk/Wv/Wo[256x256] (head dim 256)
 *       Q = X*Wq, K = X*Wk, V = X*Wv
 *       S = Q*K^T / 16            (scale = 1/sqrt(head_dim))
 *       A = softmax(S)            (row-wise)
 *       Y = (A*V)*Wo
 *  2. Mini-transformer FFN head: h <- ReLU(h*Wf[l]) for 4 layers, then
 *       logits = h*Wh, P = softmax(logits)  (row-wise over 256)
 *
 * Inputs and weights are multiples of 0.25/0.125, so every GEMM is exact in
 * fp32; only the softmax exponentials differ from the CPU reference, hence
 * the relative tolerance.
 *
 * Output lines are stable and greppable:
 *
 *   DEVICE <name> cc=<major>.<minor> sm=<count> mem=<bytes>
 *   TRANSFORMER dh=256 seq=64 bad=<mismatches> ysum=<sum>
 *   FFN4 layers=4 bad=<mismatches> psum=<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 -arch=sm_75 -o transformer transformer.cu
 *
 * Then, from the guest:  /tmp/transformer
 */

#include <cstdio>
#include <cstdlib>
#include <cstdint>
#include <cmath>
#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 SEQ 64
#define DH 256
#define TILE 16
#define TOL 1e-3f

__global__ void gemm_nn(const float *A, const float *B, float *C,
			int M, int N, int K)
{
	__shared__ float As[TILE][TILE];
	__shared__ float Bs[TILE][TILE];
	int row = blockIdx.y * TILE + threadIdx.y;
	int col = blockIdx.x * TILE + threadIdx.x;
	float acc = 0.0f;

	for (int t = 0; t < K / TILE; t++) {
		As[threadIdx.y][threadIdx.x] = A[row * K + t * TILE + threadIdx.x];
		Bs[threadIdx.y][threadIdx.x] =
			B[(t * TILE + threadIdx.y) * N + col];
		__syncthreads();
		for (int k = 0; k < TILE; k++)
			acc += As[threadIdx.y][k] * Bs[k][threadIdx.x];
		__syncthreads();
	}
	C[row * N + col] = acc;
}

__global__ void gemm_nt(const float *A, const float *B, float *C,
			int M, int N, int K)
{
	/* C[i][j] = sum_k A[i][k] * B[j][k]  (B is [N][K], row-major). */
	__shared__ float As[TILE][TILE];
	__shared__ float Bs[TILE][TILE];
	int row = blockIdx.y * TILE + threadIdx.y;
	int col = blockIdx.x * TILE + threadIdx.x;
	float acc = 0.0f;

	for (int t = 0; t < K / TILE; t++) {
		As[threadIdx.y][threadIdx.x] = A[row * K + t * TILE + threadIdx.x];
		Bs[threadIdx.y][threadIdx.x] =
			B[(blockIdx.x * TILE + threadIdx.y) * K +
			  t * TILE + threadIdx.x];
		__syncthreads();
		for (int k = 0; k < TILE; k++)
			acc += As[threadIdx.y][k] * Bs[threadIdx.x][k];
		__syncthreads();
	}
	C[row * N + col] = acc;
}

__global__ void softmax_rows(const float *S, float *P, int rows, int cols,
			     float scale)
{
	extern __shared__ float sm[];
	int row = blockIdx.x;
	int t = threadIdx.x;
	float v = S[row * cols + t] * scale;
	int stride;

	sm[t] = v;
	__syncthreads();
	for (stride = blockDim.x >> 1; stride > 0; stride >>= 1) {
		if (t < stride)
			sm[t] = fmaxf(sm[t], sm[t + stride]);
		__syncthreads();
	}
	v = expf(v - sm[0]);
	sm[t] = v;
	__syncthreads();
	for (stride = blockDim.x >> 1; stride > 0; stride >>= 1) {
		if (t < stride)
			sm[t] += sm[t + stride];
		__syncthreads();
	}
	P[row * cols + t] = v / sm[0];
}

__global__ void relu_inplace(float *h, int n)
{
	int i = blockIdx.x * blockDim.x + threadIdx.x;

	if (i < n && h[i] < 0.0f)
		h[i] = 0.0f;
}

/* ------------------------------------------------------------------ host */

static void launch_gemm_nn(const float *A, const float *B, float *C,
			   int M, int N, int K)
{
	gemm_nn<<<dim3(N / TILE, M / TILE), dim3(TILE, TILE)>>>(A, B, C, M, N, K);
}

static void launch_gemm_nt(const float *A, const float *B, float *C,
			   int M, int N, int K)
{
	gemm_nt<<<dim3(N / TILE, M / TILE), dim3(TILE, TILE)>>>(A, B, C, M, N, K);
}

static void launch_softmax(const float *S, float *P, int rows, int cols,
			   float scale)
{
	softmax_rows<<<rows, cols, (size_t)cols * sizeof(float)>>>(S, P, rows,
								   cols,
								   scale);
}

/* Deterministic inputs: multiples of 0.25 (X) and 0.125 (weights). */
static float xval(int i)
{
	return (float)(((i % 7) - 3)) * 0.25f;
}

static float wval(int i, int seed)
{
	return (float)(((i * 13 + seed * 101 + 7) % 5) - 2) * 0.125f;
}

static void ref_gemm(const double *A, const double *B, double *C,
		     int M, int N, int K)
{
	for (int i = 0; i < M; i++)
		for (int j = 0; j < N; j++) {
			double acc = 0.0;

			for (int k = 0; k < K; k++)
				acc += A[i * K + k] * B[k * N + j];
			C[i * N + j] = acc;
		}
}

static void ref_gemm_nt(const double *A, const double *B, double *C,
			int M, int N, int K)
{
	for (int i = 0; i < M; i++)
		for (int j = 0; j < N; j++) {
			double acc = 0.0;

			for (int k = 0; k < K; k++)
				acc += A[i * K + k] * B[j * K + k];
			C[i * N + j] = acc;
		}
}

static void ref_softmax(const double *S, double *P, int rows, int cols,
			double scale)
{
	for (int i = 0; i < rows; i++) {
		const double *s = S + i * cols;
		double *p = P + i * cols;
		double m = s[0] * scale, sum = 0.0;

		for (int j = 1; j < cols; j++)
			if (s[j] * scale > m)
				m = s[j] * scale;
		for (int j = 0; j < cols; j++) {
			p[j] = exp(s[j] * scale - m);
			sum += p[j];
		}
		for (int j = 0; j < cols; j++)
			p[j] /= sum;
	}
}

static void ref_relu(double *h, int n)
{
	for (int i = 0; i < n; i++)
		if (h[i] < 0.0)
			h[i] = 0.0;
}

static int count_bad(const float *gpu, const double *ref, int n)
{
	int bad = 0;

	for (int i = 0; i < n; i++) {
		double g = (double)gpu[i];
		double r = ref[i];

		if (fabs(g - r) > (double)TOL * (fabs(r) + 1e-6))
			bad++;
	}
	return bad;
}

static double dsum(const float *h, int n)
{
	double s = 0.0;

	for (int i = 0; i < n; i++)
		s += (double)h[i];
	return s;
}

#define NF 4 /* FFN layers */

int main(void)
{
	cudaDeviceProp prop;
	const int seqdh = SEQ * DH, dh2 = DH * DH, seq2 = SEQ * SEQ;
	const size_t b_seqdh = (size_t)seqdh * sizeof(float);
	const size_t b_dh2 = (size_t)dh2 * sizeof(float);
	const size_t b_seq2 = (size_t)seq2 * sizeof(float);
	float *h_x, *h_y, *h_p;
	double *r_x, *r_q, *r_k, *r_v, *r_s, *r_a, *r_av, *r_y;
	double *r_f0, *r_f1, *r_logits, *r_p;
	float *d_x, *d_q, *d_k, *d_v, *d_s, *d_a, *d_av, *d_y;
	float *d_wq, *d_wk, *d_wv, *d_wo;
	float *d_wf[NF], *d_wh, *d_f0, *d_f1, *d_logits, *d_p;
	float *d_ping, *d_pong, *tmpf;
	double *r_ping, *r_pong, *tmpd;
	int bad_y, bad_p, i, l;

	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);

	/* host inputs + CPU reference workspace */
	h_x = (float *)malloc(b_seqdh);
	h_y = (float *)malloc(b_seqdh);
	h_p = (float *)malloc(b_seqdh);
	r_x = (double *)malloc((size_t)seqdh * sizeof(double));
	r_q = (double *)malloc((size_t)seqdh * sizeof(double));
	r_k = (double *)malloc((size_t)seqdh * sizeof(double));
	r_v = (double *)malloc((size_t)seqdh * sizeof(double));
	r_s = (double *)malloc((size_t)seq2 * sizeof(double));
	r_a = (double *)malloc((size_t)seq2 * sizeof(double));
	r_av = (double *)malloc((size_t)seqdh * sizeof(double));
	r_y = (double *)malloc((size_t)seqdh * sizeof(double));
	r_f0 = (double *)malloc((size_t)seqdh * sizeof(double));
	r_f1 = (double *)malloc((size_t)seqdh * sizeof(double));
	r_logits = (double *)malloc((size_t)seqdh * sizeof(double));
	r_p = (double *)malloc((size_t)seqdh * sizeof(double));
	if (!h_x || !h_y || !h_p || !r_x || !r_q || !r_k || !r_v || !r_s ||
	    !r_a || !r_av || !r_y || !r_f0 || !r_f1 || !r_logits || !r_p)
		return 1;

	for (i = 0; i < seqdh; i++) {
		h_x[i] = xval(i);
		r_x[i] = (double)xval(i);
	}

	/* device buffers */
	CHECK(cudaMalloc(&d_x, b_seqdh));
	CHECK(cudaMalloc(&d_q, b_seqdh));
	CHECK(cudaMalloc(&d_k, b_seqdh));
	CHECK(cudaMalloc(&d_v, b_seqdh));
	CHECK(cudaMalloc(&d_s, b_seq2));
	CHECK(cudaMalloc(&d_a, b_seq2));
	CHECK(cudaMalloc(&d_av, b_seqdh));
	CHECK(cudaMalloc(&d_y, b_seqdh));
	CHECK(cudaMalloc(&d_wq, b_dh2));
	CHECK(cudaMalloc(&d_wk, b_dh2));
	CHECK(cudaMalloc(&d_wv, b_dh2));
	CHECK(cudaMalloc(&d_wo, b_dh2));
	CHECK(cudaMalloc(&d_wh, b_dh2));
	CHECK(cudaMalloc(&d_f0, b_seqdh));
	CHECK(cudaMalloc(&d_f1, b_seqdh));
	CHECK(cudaMalloc(&d_logits, b_seqdh));
	CHECK(cudaMalloc(&d_p, b_seqdh));
	for (l = 0; l < NF; l++)
		CHECK(cudaMalloc(&d_wf[l], b_dh2));

	/* weights: deterministic, uploaded once */
	{
		float *w = (float *)malloc(b_dh2);

		if (!w)
			return 1;
		for (i = 0; i < dh2; i++)
			w[i] = wval(i, 1);
		CHECK(cudaMemcpy(d_wq, w, b_dh2, cudaMemcpyHostToDevice));
		for (i = 0; i < dh2; i++)
			w[i] = wval(i, 2);
		CHECK(cudaMemcpy(d_wk, w, b_dh2, cudaMemcpyHostToDevice));
		for (i = 0; i < dh2; i++)
			w[i] = wval(i, 3);
		CHECK(cudaMemcpy(d_wv, w, b_dh2, cudaMemcpyHostToDevice));
		for (i = 0; i < dh2; i++)
			w[i] = wval(i, 4);
		CHECK(cudaMemcpy(d_wo, w, b_dh2, cudaMemcpyHostToDevice));
		for (i = 0; i < dh2; i++)
			w[i] = wval(i, 20);
		CHECK(cudaMemcpy(d_wh, w, b_dh2, cudaMemcpyHostToDevice));
		for (l = 0; l < NF; l++) {
			for (i = 0; i < dh2; i++)
				w[i] = wval(i, 10 + l);
			CHECK(cudaMemcpy(d_wf[l], w, b_dh2,
					 cudaMemcpyHostToDevice));
		}
		free(w);
	}
	CHECK(cudaMemcpy(d_x, h_x, b_seqdh, cudaMemcpyHostToDevice));

	/* ---- check 1: attention forward ---- */
	launch_gemm_nn(d_x, d_wq, d_q, SEQ, DH, DH);
	launch_gemm_nn(d_x, d_wk, d_k, SEQ, DH, DH);
	launch_gemm_nn(d_x, d_wv, d_v, SEQ, DH, DH);
	launch_gemm_nt(d_q, d_k, d_s, SEQ, SEQ, DH);
	launch_softmax(d_s, d_a, SEQ, SEQ, 1.0f / 16.0f);
	launch_gemm_nn(d_a, d_v, d_av, SEQ, DH, SEQ);
	launch_gemm_nn(d_av, d_wo, d_y, SEQ, DH, DH);
	CHECK(cudaGetLastError());
	CHECK(cudaDeviceSynchronize());
	CHECK(cudaMemcpy(h_y, d_y, b_seqdh, cudaMemcpyDeviceToHost));

	{
		double *r_w = (double *)malloc((size_t)dh2 * sizeof(double));

		if (!r_w)
			return 1;
		for (i = 0; i < dh2; i++)
			r_w[i] = (double)wval(i, 1);
		ref_gemm(r_x, r_w, r_q, SEQ, DH, DH);
		for (i = 0; i < dh2; i++)
			r_w[i] = (double)wval(i, 2);
		ref_gemm(r_x, r_w, r_k, SEQ, DH, DH);
		for (i = 0; i < dh2; i++)
			r_w[i] = (double)wval(i, 3);
		ref_gemm(r_x, r_w, r_v, SEQ, DH, DH);
		ref_gemm_nt(r_q, r_k, r_s, SEQ, SEQ, DH);
		ref_softmax(r_s, r_a, SEQ, SEQ, 1.0 / 16.0);
		ref_gemm(r_a, r_v, r_av, SEQ, DH, SEQ);
		for (i = 0; i < dh2; i++)
			r_w[i] = (double)wval(i, 4);
		ref_gemm(r_av, r_w, r_y, SEQ, DH, DH);
		free(r_w);
	}

	bad_y = count_bad(h_y, r_y, seqdh);
	printf("TRANSFORMER dh=256 seq=64 bad=%d ysum=%.6f\n", bad_y,
	       dsum(h_y, seqdh));

	/* ---- check 2: 4-layer FFN stack + softmax head ---- */
	CHECK(cudaMemcpy(d_f0, h_x, b_seqdh, cudaMemcpyHostToDevice));
	d_ping = d_f0;
	d_pong = d_f1;
	for (l = 0; l < NF; l++) {
		launch_gemm_nn(d_ping, d_wf[l], d_pong, SEQ, DH, DH);
		relu_inplace<<<(seqdh + 255) / 256, 256>>>(d_pong, seqdh);
		tmpf = d_ping;
		d_ping = d_pong;
		d_pong = tmpf;
	}
	launch_gemm_nn(d_ping, d_wh, d_logits, SEQ, DH, DH);
	launch_softmax(d_logits, d_p, SEQ, DH, 1.0f);
	CHECK(cudaGetLastError());
	CHECK(cudaDeviceSynchronize());
	CHECK(cudaMemcpy(h_p, d_p, b_seqdh, cudaMemcpyDeviceToHost));

	{
		double *r_w = (double *)malloc((size_t)dh2 * sizeof(double));

		if (!r_w)
			return 1;
		memcpy(r_f0, r_x, (size_t)seqdh * sizeof(double));
		r_ping = r_f0;
		r_pong = r_f1;
		for (l = 0; l < NF; l++) {
			for (i = 0; i < dh2; i++)
				r_w[i] = (double)wval(i, 10 + l);
			ref_gemm(r_ping, r_w, r_pong, SEQ, DH, DH);
			ref_relu(r_pong, seqdh);
			tmpd = r_ping;
			r_ping = r_pong;
			r_pong = tmpd;
		}
		for (i = 0; i < dh2; i++)
			r_w[i] = (double)wval(i, 20);
		ref_gemm(r_ping, r_w, r_logits, SEQ, DH, DH);
		ref_softmax(r_logits, r_p, SEQ, DH, 1.0);
		free(r_w);
	}

	bad_p = count_bad(h_p, r_p, seqdh);
	printf("FFN4 layers=4 bad=%d psum=%.6f\n", bad_p, dsum(h_p, seqdh));

	if (bad_y != 0 || bad_p != 0) {
		printf("STATUS FAIL\n");
		return 1;
	}

	printf("STATUS OK\n");
	CHECK(cudaFree(d_x));
	CHECK(cudaFree(d_q));
	CHECK(cudaFree(d_k));
	CHECK(cudaFree(d_v));
	CHECK(cudaFree(d_s));
	CHECK(cudaFree(d_a));
	CHECK(cudaFree(d_av));
	CHECK(cudaFree(d_y));
	CHECK(cudaFree(d_wq));
	CHECK(cudaFree(d_wk));
	CHECK(cudaFree(d_wv));
	CHECK(cudaFree(d_wo));
	CHECK(cudaFree(d_wh));
	for (l = 0; l < NF; l++)
		CHECK(cudaFree(d_wf[l]));
	CHECK(cudaFree(d_f0));
	CHECK(cudaFree(d_f1));
	CHECK(cudaFree(d_logits));
	CHECK(cudaFree(d_p));
	free(h_x);
	free(h_y);
	free(h_p);
	free(r_x);
	free(r_q);
	free(r_k);
	free(r_v);
	free(r_s);
	free(r_a);
	free(r_av);
	free(r_y);
	free(r_f0);
	free(r_f1);
	free(r_logits);
	free(r_p);
	return 0;
}