/**
 * Map TMEM Layout D for PV MMA N=64 using 32x32b.x8 reads (all lanes).
 * All 32 lanes read, giving 128 positions per group of 8 columns.
 */

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

#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 = 64, SK = 128, BLOCK_MN = 128;
constexpr int LOCAL_MMA_K = 16;
constexpr int TILE_SZ = BLOCK_MN * LOCAL_MMA_K;
constexpr int V_TILE_SZ = (HD / 8) * 2 * 64;

// Global memory buffer for TMEM dump
// 64 columns × 128 positions = 8192 FP32
// But we read 8 columns at a time (32x32b.x8), 8 reads for 64 cols
// Each read: 8 columns × 128 positions (32 lanes × 4 FP32 per lane)
// But lane 0 gets 8 FP32 values (one per column). We only dump lane 0's data
// since T=1 decode only uses row 0.
// Actually, for row 0, lane 0 should have positions 0-3 of each column.
// But for Layout D, row 0 might be in a different lane's positions.
// So we dump ALL lanes' data via SMEM.

// We use a 2-step approach: read TMEM in kernel, store to GMEM,
// then analyze in host.

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

    extern __shared__ char sbuf[];
    uint32_t* sTmemBase = (uint32_t*)sbuf;
    bf16_t* sQ0 = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
    bf16_t* sK0 = sQ0 + 4 * TILE_SZ;
    bf16_t* sPk = (bf16_t*)(((uintptr_t)(sK0 + 4 * TILE_SZ) + 127) & ~(uintptr_t)127);
    bf16_t* sV = (bf16_t*)(((uintptr_t)(sPk + TILE_SZ) + 127) & ~(uintptr_t)127);
    float* s_p_vals = (float*)(sV + 8 * V_TILE_SZ);

    for (int kt = 0; kt < 4; kt++) {
        bf16_t* sq = sQ0 + kt * TILE_SZ;
        for (int i = tid; i < TILE_SZ; i += 128) sq[i] = 0;
        for (int d = tid; d < LOCAL_MMA_K; d += 128) {
            int ck = d / 8, lc = d % 8;
            sq[ck * 16 * 64 + lc] = q[kt * LOCAL_MMA_K + d];
        }
    }
    for (int kt = 0; kt < 4; kt++) {
        bf16_t* sk = sK0 + kt * TILE_SZ;
        for (int i = tid; i < TILE_SZ; i += 128) sk[i] = 0;
        for (int r = 0; r < SK; r++) {
            for (int d = tid; d < LOCAL_MMA_K; d += 128) {
                int ck = d / 8, lc = d % 8;
                int tmn = r / 8, lr = r % 8;
                sk[ck * 16 * 64 + tmn * 64 + lr * 8 + lc] = k[r * HD + kt * LOCAL_MMA_K + d];
            }
        }
    }
    for (int kt = 0; kt < 8; kt++) {
        bf16_t* sv = sV + kt * V_TILE_SZ;
        for (int i = tid; i < V_TILE_SZ; i += 128) sv[i] = 0;
        for (int d = tid; d < HD; d += 128) {
            for (int lr = 0; lr < LOCAL_MMA_K; lr++) {
                int r = kt * LOCAL_MMA_K + lr;
                int g_mn = d / 8, g_k = lr / 8;
                int llr = d % 8, lc = lr % 8;
                sv[g_k * 8 * 64 + g_mn * 64 + llr * 8 + lc] = v[d * SK + r];
            }
        }
    }
    __syncthreads();

    if (wid == 1) tmem_alloc(__cvta_generic_to_shared(sTmemBase), 128);
    __syncthreads();
    uint32_t tb = *sTmemBase;

    // QK GEMM
    {
        uint32_t idesc = make_idesc(BLOCK_MN, BLOCK_MN);
        for (int kt = 0; kt < 4; kt++) {
            bf16_t* sq = sQ0 + kt * TILE_SZ;
            bf16_t* sk = sK0 + kt * TILE_SZ;
            uint64_t dq = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sq), BLOCK_MN);
            uint64_t dk = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sk), BLOCK_MN);
            if (tid == 0) umma_ss_f16(tb, dq, dk, idesc, kt > 0);
            asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
            __syncthreads();
        }
    }

    // Softmax
    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;
        if (lane == 0) for (int j=0;j<SK;j++) s_p_vals[j] = s_vals[j];
    }
    __syncthreads();

    // PV MMA (N=64)
    {
        uint32_t idesc_pv = make_idesc(BLOCK_MN, HD);
        for (int kt = 0; kt < 8; kt++) {
            for (int i = tid; i < TILE_SZ; i += 128) sPk[i] = 0;
            if (tid < 16) {
                int c = tid;
                int ck = c / 8, lc = c % 8;
                sPk[ck * 16 * 64 + 0 * 64 + 0 * 8 + lc] = f32_to_bf16(s_p_vals[kt * LOCAL_MMA_K + c]);
            }
            __syncthreads();
            bf16_t* sv = sV + kt * V_TILE_SZ;
            uint64_t dp = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sPk), BLOCK_MN);
            uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sv), HD);
            if (tid == 0) umma_ss_f16(tb, dp, dv, idesc_pv, kt > 0);
            asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
            __syncthreads();
        }
    }

    // ===== Read TMEM: dump all lanes' data for columns 0..63 =====
    // Use 32x32b.x8 read, 8 reads of 8 columns each
    // Each lane gets 8 FP32 values per read (one per column)
    // 32 lanes × 4 FP32 per column = 128 FP32 per column
    // For lane i, the 8 values are positions i*4+0..3 of each of 8 columns
    // We write: for read n (cols n*8..n*8+7), lane i, column c (0..7):
    //   output[(n*8 + c) * 128 + i*4 + 0..3] = tmp[0..3]  (but tmp only has 1 value per column)
    // Wait — 32x32b.x8 gives each lane 8 FP32 values, one per column
    // Lane i gets: for each of 8 columns, one value at position (i*4 + sub) within the column
    // Actually, the 32x32b.x8 format: each lane reads 8 FP32 from 8 consecutive columns
    // Lane i's 8 values are at position (i*4 + 0), (i*4 + 1), (i*4 + 2), (i*4 + 3)
    // of each of the 8 columns. No wait — the x8 reads 8 columns, each column gives 1 FP32 per lane.
    // So lane i gets tmp[0..7], where tmp[j] is from column (n*8 + j), position i.
    // But what "position i"? For 32 lanes, each lane reads one of 32 positions per column.
    // The mapping: lane i reads positions i*4+0..i*4+3 in a 16x256b read, but 32x32b.x8 is different.
    //
    // From the verified TMEM mapping:
    // 32x32b.x8: reads 8 columns. Each lane gets 8 FP32 values.
    // Lane i's 8 values correspond to the same position within each of the 8 columns.
    // The position within the column depends on lane i.
    // For the QK read (N=128), lane 0's values corresponded to output row 0 positions.
    // 
    // We dump: tmem_dump[(n*8 + col_idx) * 128 + lane*4 + 0] = tmp[col_idx]
    // But this might not be the right mapping. Let's just dump lane i's 8 values for each read
    // and figure out the mapping on the host.

    if (wid == 0) {
        for (int n = 0; n < 16; n++) {  // 16 reads of 8 columns = 128 columns
            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;");
            // Each lane writes its 8 values to GMEM
            // For column (n*8 + c), lane i's value goes to position lane in that column
            // Layout: tmem_dump[col * 32 + lane] = lane's value for column col
            for (int c = 0; c < 8; c++) {
                int col = n * 8 + c;
                if (col < 128) tmem_dump[col * 32 + lane] = tmp[c];
            }
        }
    }
    __syncthreads();

    if (wid == 0) tmem_dealloc(tb, 128);
}

int main() {
    printf("=== TMEM Layout D dump (32x32b.x8, all lanes) ===\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));

    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;
    float* d_tmem_dump;
    cudaMalloc(&d_q, HD*sizeof(bf16_t));
    cudaMalloc(&d_k, SK*HD*sizeof(bf16_t));
    cudaMalloc(&d_v, HD*SK*sizeof(bf16_t));
    // 128 columns × 32 lanes = 4096 FP32
    cudaMalloc(&d_tmem_dump, 128 * 32 * sizeof(float));
    cudaMemset(d_tmem_dump, 0, 128 * 32 * 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 + 4*TILE_SZ*2 + 4*TILE_SZ*2 + TILE_SZ*2 + 8*V_TILE_SZ*2 + SK*4 + 256 + 127) & ~127;
    cudaFuncSetAttribute(test_tmem_all_lanes, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
    test_tmem_all_lanes<<<1, 128, smem>>>(d_q, d_k, d_v, d_tmem_dump, SCALE);

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

    float* h_dump = (float*)malloc(128 * 32 * sizeof(float));
    cudaMemcpy(h_dump, d_tmem_dump, 128 * 32 * sizeof(float), cudaMemcpyDeviceToHost);

    // Reference
    float s[SK];
    for (int j=0;j<SK;j++) {
        float dot = 0.0f;
        for (int d=0;d<HD;d++) dot += bf16_to_f32_host(h_q[d]) * bf16_to_f32_host(h_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;
    float o_ref[HD];
    for (int d=0;d<HD;d++) {
        float ov = 0.0f;
        for (int j=0;j<SK;j++) ov += s[j] * bf16_to_f32_host(h_v[d*SK+j]);
        o_ref[d] = ov;
    }

    // Find mapping: for each output d, find (col, lane) that has the closest value
    printf("=== Mapping: output position d -> (col, lane) ===\n");
    for (int d = 0; d < HD; d++) {
        float target = o_ref[d];
        int best_col = -1, best_lane = -1;
        float best_diff = 1e10f;
        for (int col = 0; col < 128; col++) {
            for (int ln = 0; ln < 32; ln++) {
                float val = h_dump[col * 32 + ln];
                float diff = fabsf(val - target);
                if (diff < best_diff) {
                    best_diff = diff;
                    best_col = col;
                    best_lane = ln;
                }
            }
        }
        printf("  d=%2d: ref=%10.6f at (col=%2d, lane=%2d) val=%10.6f diff=%.2e\n",
               d, target, best_col, best_lane, h_dump[best_col*32+best_lane], best_diff);
    }

    // Print column summary: for each column, which lanes have non-zero values?
    printf("\n=== Non-zero lanes per column ===\n");
        for (int col = 0; col < 128; col++) {
            int nz = 0;
            for (int ln = 0; ln < 32; ln++) if (fabsf(h_dump[col*32+ln]) > 1e-6f) nz++;
            if (nz > 0) printf("  col %3d: %d non-zero lanes, lane0=%10.6f\n", col, nz, h_dump[col*32+0]);
        }

    cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_tmem_dump);
    free(h_q); free(h_k); free(h_v); free(h_dump);
    return 0;
}