/**
 * PV SS MMA with P as (128, 128) and K-tile at offset.
 * Extending test_pv_ss.cu to use the full P matrix layout.
 */

#include <cuda_runtime.h>
#include <cstdio>
#include <cmath>
#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); }

constexpr int HD = 16, SK = 128, BLOCK_MN = 128;

__global__ void __launch_bounds__(128)
test_pv_ss_128()
{
    const int tid = threadIdx.x, wid = tid / 32, lane = tid % 32;

    extern __shared__ char sbuf[];
    uint32_t* sTmemBase = (uint32_t*)sbuf;
    // P as (128, 128) canonical
    bf16_t* sP = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
    // V as (16, 16) canonical
    bf16_t* sV = (bf16_t*)(((uintptr_t)(sP + 128 * 128) + 127) & ~(uintptr_t)127);

    // Fill P: (128, 128) canonical, row 0 = all 0.5
    // (128, 128): CORES_MN=16, CORES_K=16
    // core(g_mn, g_k) at g_k * 16 * 64 + g_mn * 64 + lr * 8 + lc
    for (int i = tid; i < 128 * 128; i += 128) sP[i] = 0;
    __syncthreads();
    for (int j = tid; j < 128; j += 128) {
        int core_k = j / 8, lc = j % 8;
        int dst_idx = core_k * 16 * 64 + 0 * 64 + 0 * 8 + lc;
        sP[dst_idx] = f32_to_bf16(0.5f);  // All 128 positions = 0.5
    }
    __syncthreads();

    // Fill V: (16, 16) canonical, all 1.0
    for (int i = tid; i < 256; i += 128) sV[i] = 0;
    __syncthreads();
    for (int i = tid; i < 256; 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();

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

    // Zero TMEM first (instead of relying on accumulate=false)
    if (wid == 0) {
        for (int n = 0; n < 2; n++) {  // 16 cols / 8 = 2 iterations to zero O region
            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();

    // PV SS MMA: 8 K-tiles with accumulation (all accumulate=true)
    // K-tile kt of (128,128): g_k=[2*kt, 2*kt+1], offset = kt * 2048 BF16
    {
        uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sV), 16);
        uint32_t idesc = make_idesc(BLOCK_MN, HD);
        for (int kt = 0; kt < 2; kt++) {
            bf16_t* sp = sP + kt * 2048;
            uint64_t dp = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sp), BLOCK_MN);
            if (tid == 0) umma_ss_f16(tb, dp, dv, idesc, true);  // ALL accumulate
            asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
            __syncthreads();
        }
    }

    // Read O from TMEM
    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];
        }
        if (lane == 0) {
            printf("O[0,0..15] (raw): ");
            for (int d=0;d<HD;d++) printf("%.2f ", o_vals[d]);
            // Expected: sum(P[0,0..15] * V[j,0..15]) = 16 * 0.5 * 1.0 = 8.0 (raw, no MMA scale)
            printf("\nExpected ~8.0 (1 K-tile, MMA scale 1.0)\n");
        }
    }

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

int main() {
    printf("=== PV SS MMA with (128,128) P ===\n");
    // SMEM: tmem(4+12) + sP(32768) + sV(512) + align
    int smem = (4+16 + 128*128*2 + 256*2 + 256 + 127) & ~127;
    printf("SMEM: %d bytes\n", smem);
    test_pv_ss_128<<<1, 128, smem>>>();
    cudaError_t err = cudaDeviceSynchronize();
    if (err != cudaSuccess) { printf("CUDA ERROR: %s\n", cudaGetErrorString(err)); return 1; }
    printf("PASS\n");
    return 0;
}