/**
 * Full UMMA FMHA — HD=16, SK=128, with PV GEMM via tcgen05.mma TS
 * 
 * Pipeline: Q×K^T (SS) → softmax → P×V (TS) → epilogue
 * TMEM: columns 0-127 = P, columns 128-159 = O
 * 
 * PV GEMM: 8 K-tiles, each A=(128,16) from TMEM, B=(16,16) from SMEM
 * Accumulate O in TMEM at columns 128-143
 */

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

#include "dsv4/kernels/attention/fmha_common.cuh"
#include "dsv4/kernels/attention/fmha_umma_desc.cuh"

using namespace dsv4::kernels::attention;

static bf16_t f32_to_bf16_host(float f) { uint32_t u; memcpy(&u,&f,4); return (uint16_t)(u>>16); }
static float bf16_to_f32_host(bf16_t h) { uint32_t u=(uint32_t)h<<16; float f; memcpy(&f,&u,4); return f; }

constexpr int HD = 16, SK = 128, BLOCK_MN = 128;
constexpr int VKT = SK / MMA_K_BF16;  // 8 PV K-tiles
constexpr int V_TILE_SZ = MMA_K_BF16 * HD;  // 256 BF16 per V K-tile
constexpr int TMEM_P = 128;     // P: 128 columns
constexpr int TMEM_O = 32;      // O: 16 cols, round to 32 (min, power of 2)
constexpr int TMEM_N = 256;     // total: 128 + 32 = 160 → round to 256

__global__ void __launch_bounds__(128)
test_fmha_ts(const bf16_t* q, const bf16_t* k, const bf16_t* v,
             bf16_t* o_out, float* o_scalar, float scale)
{
    const int tid = threadIdx.x, wid = tid / 32, lane = tid % 32;

    // SMEM: tmem_base + sQ(128,16)+pad + sK(128,16) + V tiles (8 × 16×16)
    extern __shared__ char sbuf[];
    uint32_t* sTmemBase = (uint32_t*)sbuf;
    bf16_t* sQ = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
    bf16_t* sK = sQ + 128 * 16 + 4096;
    // Fresh aligned V buffer (ensure 128-byte alignment for descriptor)
    bf16_t* sV = (bf16_t*)(((uintptr_t)(sK + 128 * 16) + 127) & ~(uintptr_t)127);
    float* sQ_row = (float*)(sV + 16 * 16);

    for (int d = tid; d < HD; d += 128) sQ_row[d] = bf16_to_f32(q[d]);

    // TMEM alloc — 256 columns
    if (wid == 1) tmem_alloc(__cvta_generic_to_shared(sTmemBase), TMEM_N);
    __syncthreads();
    uint32_t tb = *sTmemBase;
    uint32_t tb_o = tb + TMEM_P;  // O at column 128

    // Load Q, K
    write_q_to_smem<HD>(sQ, q);
    write_k_to_smem<SK, HD>(sK, k);
    bf16_t* sQ_pad = sQ + 128 * 16;
    for (int i = tid; i < 4096; i += 128) sQ_pad[i] = 0;

    // Load V = all 1.0 into (16, 16) canonical (same as isolated test)
    for (int i = tid; i < 16 * 16; i += 128) sV[i] = 0;
    __syncthreads();
    for (int i = tid; i < 16 * 16; i += 128) {
        int r = i / 16, c = i % 16;
        int ck = c / 8, lc = c % 8;
        int tmn = r / 8, lr = r % 8;
        sV[ck * 2 * 64 + tmn * 64 + lr * 8 + lc] = f32_to_bf16(1.0f);
    }
    __syncthreads();

    // DEBUG: Skip QK and softmax, write P directly to TMEM
    // Write P = all 1.0 (128, 128) — only row 0 has 1.0, rest 0
    if (wid == 0) {
        for (int n = 0; n < SK / 8; n++) {
            // Lane 0 writes 1.0 for row 0, other lanes write 0
            float p0=(lane==0)?0.0078125f:0, p1=(lane==0)?0.0078125f:0;
            float p2=(lane==0)?0.0078125f:0, p3=(lane==0)?0.0078125f:0;
            float p4=(lane==0)?0.0078125f:0, p5=(lane==0)?0.0078125f:0;
            float p6=(lane==0)?0.0078125f:0, p7=(lane==0)?0.0078125f:0;
            asm volatile("tcgen05.st.sync.aligned.32x32b.x8.b32 [%0],{%1,%2,%3,%4,%5,%6,%7,%8};" :: "r"(tb+n*8),"f"(p0),"f"(p1),"f"(p2),"f"(p3),"f"(p4),"f"(p5),"f"(p6),"f"(p7));
        }
        tmem_fence_store();
    }
    __syncthreads();

    /*
    // STEP 1: QK GEMM — DISABLED for debug
    uint64_t desc_q = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sQ), BLOCK_MN);
    uint64_t desc_k = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sK), BLOCK_MN);
    uint32_t idesc_qk = make_idesc(BLOCK_MN, BLOCK_MN);
    if (lane == 0) umma_ss_f16(tb, desc_q, desc_k, idesc_qk, false);
    asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
    __syncthreads();

    // STEP 2: Softmax — DISABLED for debug
    if (wid == 0) {
        float s_vals[SK], row_max = -INFINITY;
        for (int n = 0; n < SK / 8; n++) {
            float tmp[8];
            asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];" : "=f"(tmp[0]),"=f"(tmp[1]),"=f"(tmp[2]),"=f"(tmp[3]),"=f"(tmp[4]),"=f"(tmp[5]),"=f"(tmp[6]),"=f"(tmp[7]) : "r"(tb + n*8));
            asm volatile("tcgen05.wait::ld.sync.aligned;");
            if (lane == 0) for (int c=0;c<8;c++) { s_vals[n*8+c] = tmp[c]*scale; row_max = fmaxf(row_max, tmp[c]*scale); }
        }
        row_max = wmax(row_max);
        float row_sum = 0.0f;
        if (lane == 0) for (int j=0;j<SK;j++) { s_vals[j] = expf(s_vals[j]-row_max); row_sum += s_vals[j]; }
        row_sum = wsum(row_sum);
        if (lane == 0) for (int j=0;j<SK;j++) s_vals[j] /= row_sum;

        for (int n = 0; n < SK / 8; n++) {
            float p0=(lane==0)?s_vals[n*8+0]:0, p1=(lane==0)?s_vals[n*8+1]:0;
            float p2=(lane==0)?s_vals[n*8+2]:0, p3=(lane==0)?s_vals[n*8+3]:0;
            float p4=(lane==0)?s_vals[n*8+4]:0, p5=(lane==0)?s_vals[n*8+5]:0;
            float p6=(lane==0)?s_vals[n*8+6]:0, p7=(lane==0)?s_vals[n*8+7]:0;
            asm volatile("tcgen05.st.sync.aligned.32x32b.x8.b32 [%0],{%1,%2,%3,%4,%5,%6,%7,%8};" :: "r"(tb+n*8),"f"(p0),"f"(p1),"f"(p2),"f"(p3),"f"(p4),"f"(p5),"f"(p6),"f"(p7));
        }
        tmem_fence_store();
    }
    __syncthreads();
    */

    // Zero O TMEM region at tb (overwriting P — same as isolated test)
    if (wid == 0) {
        for (int n = 0; n < 2; n++) {  // 16 cols / 8 = 2 iterations
            float z0=0,z1=0,z2=0,z3=0,z4=0,z5=0,z6=0,z7=0;
            asm volatile("tcgen05.st.sync.aligned.32x32b.x8.b32 [%0],{%1,%2,%3,%4,%5,%6,%7,%8};" :: "r"(tb+n*8),"f"(z0),"f"(z1),"f"(z2),"f"(z3),"f"(z4),"f"(z5),"f"(z6),"f"(z7));
        }
        tmem_fence_store();
    }
    __syncthreads();

    // STEP 3: PV GEMM via tcgen05.mma TS
    // A = P (TMEM, 128 rows × SK cols), B = V (SMEM K-tiles)
    // C = O (TMEM at tb_o, 128 rows × HD cols)
    // MMA: M=128, N=16. K=16 per MMA call, 8 K-tiles total.
    uint32_t idesc_pv = make_idesc(BLOCK_MN, HD);

    for (int kt = 0; kt < 1; kt++) {  // DEBUG: only first PV K-tile
        uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sV), 16);
        uint32_t tmem_a = tb;  // P's first 16 columns

        if (tid == 0) umma_ts_f16(tb, tmem_a, dv, idesc_pv, false);  // Write C to tb (overwriting P)
        asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
        __syncthreads();
    }
    asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
    __syncthreads();

    // STEP 4: Epilogue — read O from TMEM, apply MMA scale (0.5), write to GMEM
    if (wid == 0) {
        float o_vals[HD];
        for (int n = 0; n < HD / 8; n++) {
            float tmp[8];
            asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];" : "=f"(tmp[0]),"=f"(tmp[1]),"=f"(tmp[2]),"=f"(tmp[3]),"=f"(tmp[4]),"=f"(tmp[5]),"=f"(tmp[6]),"=f"(tmp[7]) : "r"(tb + n*8));
            asm volatile("tcgen05.wait::ld.sync.aligned;");
            if (lane == 0) for (int c=0;c<8;c++) o_vals[n*8+c] = tmp[c] * 2.0f;  // Undo MMA 0.5 scale
        }
        if (lane == 0) for (int d=0;d<HD;d++) o_out[d] = f32_to_bf16(o_vals[d]);
    }
    __syncthreads();

    // Scalar reference
    if (tid == 0) {
        float s[SK];
        for (int j=0;j<SK;j++) {
            float dot = 0.0f;
            for (int d=0;d<HD;d++) dot += sQ_row[d] * bf16_to_f32(k[j*HD+d]);
            s[j] = dot * scale;
        }
        float mx = -INFINITY;
        for (int j=0;j<SK;j++) mx = fmaxf(mx, s[j]);
        float sm = 0.0f;
        for (int j=0;j<SK;j++) { s[j] = expf(s[j]-mx); sm += s[j]; }
        for (int j=0;j<SK;j++) s[j] /= sm;
        for (int d=0;d<HD;d++) {
            float ov = 0.0f;
            for (int j=0;j<SK;j++) ov += s[j] * bf16_to_f32(v[d*SK+j]);
            o_scalar[d] = ov;
        }
    }

    if (wid == 0) tmem_dealloc(tb, TMEM_N);
}

int main() {
    printf("=== Full UMMA FMHA HD=16 with PV GEMM (TS) ===\n");
    const float SCALE = 1.0f / sqrtf((float)HD);

    bf16_t* h_q = (bf16_t*)malloc(HD*sizeof(bf16_t));
    bf16_t* h_k = (bf16_t*)malloc(SK*HD*sizeof(bf16_t));
    bf16_t* h_v = (bf16_t*)malloc(HD*SK*sizeof(bf16_t));
    bf16_t* h_o = (bf16_t*)calloc(HD, sizeof(bf16_t));
    float* h_o_scalar = (float*)calloc(HD, sizeof(float));

    srand(42);
    for (int d=0;d<HD;d++) h_q[d] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
    for (int i=0;i<SK*HD;i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
    for (int i=0;i<HD*SK;i++) h_v[i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);

    bf16_t *d_q,*d_k,*d_v,*d_o; float *d_o_scalar;
    cudaMalloc(&d_q, HD*sizeof(bf16_t));
    cudaMalloc(&d_k, SK*HD*sizeof(bf16_t));
    cudaMalloc(&d_v, HD*SK*sizeof(bf16_t));
    cudaMalloc(&d_o, HD*sizeof(bf16_t));
    cudaMalloc(&d_o_scalar, HD*sizeof(float));
    cudaMemcpy(d_q, h_q, HD*sizeof(bf16_t), cudaMemcpyHostToDevice);
    cudaMemcpy(d_k, h_k, SK*HD*sizeof(bf16_t), cudaMemcpyHostToDevice);
    cudaMemcpy(d_v, h_v, HD*SK*sizeof(bf16_t), cudaMemcpyHostToDevice);

    int smem = (4+16 + 128*16*2+4096 + 128*16*2 + VKT*V_TILE_SZ*2 + 16*4 + 256 + 127) & ~127;
    test_fmha_ts<<<1, 128, smem>>>(d_q, d_k, d_v, d_o, d_o_scalar, SCALE);

    cudaError_t err = cudaDeviceSynchronize();
    if (err != cudaSuccess) { printf("CUDA ERROR: %s\n", cudaGetErrorString(err)); return 1; }

    cudaMemcpy(h_o, d_o, HD*sizeof(bf16_t), cudaMemcpyDeviceToHost);
    cudaMemcpy(h_o_scalar, d_o_scalar, HD*sizeof(float), cudaMemcpyDeviceToHost);

    printf("O[0..15] MMA:  "); for(int d=0;d<HD;d++) printf("%.6f ",bf16_to_f32_host(h_o[d])); printf("\n");
    printf("O[0..15] ref:  "); for(int d=0;d<HD;d++) printf("%.6f ",h_o_scalar[d]); printf("\n");

    float max_diff=0, max_val=0;
    for (int d=0;d<HD;d++) {
        max_diff = fmaxf(max_diff, fabsf(bf16_to_f32_host(h_o[d]) - h_o_scalar[d]));
        max_val = fmaxf(max_val, fabsf(h_o_scalar[d]));
    }
    float rel_err = max_val>0 ? max_diff/max_val : max_diff;
    float cos_sim=0,na=0,nb=0;
    for (int d=0;d<HD;d++) { float a=bf16_to_f32_host(h_o[d]),b=h_o_scalar[d]; cos_sim+=a*b; na+=a*a; nb+=b*b; }
    cos_sim /= (sqrtf(na)*sqrtf(nb)+1e-10f);
    printf("Max rel err: %.8f | cosine: %.8f\n", rel_err, cos_sim);
    printf("Test %s\n", cos_sim > 0.999f ? "PASSED" : "FAILED");

    cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_o_scalar);
    free(h_q); free(h_k); free(h_v); free(h_o); free(h_o_scalar);
    return cos_sim > 0.999f ? 0 : 1;
}