happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

WIP: debugging

This commit is contained in:
bssrdf 2025-11-07 17:44:00 -05:00
parent df88b2c917
commit 76885c7697
1 changed files with 13 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-cuda/conv2d-implicit.cu
View File

@ -677,6 +677,7 @@ static __global__ void conv2d_implicit_kernel(const half * __restrict__ input,
uint32_t (&reg_)[2] = reinterpret_cast<uint32_t(&)[2]>(acc_register_[mma_m][mma_n]);
uint idx = output_sts_addr +
mma_m * MMA_M * BN / 2 + (mma_n - i * mma_tiles_per_warp_n/2) * MMA_N;
idx = idx ^ ((idx & 0b110000000000) >> 9);
idx = idx ^ ((idx & 0b1110000000) >> 4);
uint32_t* dst_ptr = reinterpret_cast<uint32_t*>(&smemoutput[idx]);
dst_ptr[0] = reg_[0];
@ -695,19 +696,24 @@ static __global__ void conv2d_implicit_kernel(const half * __restrict__ input,
const int n = fastdiv(gemm_i, param.OHOW_fastdiv);
const int col = fastmodulo(gemm_i, param.OHOW_fastdiv);
uint idx = output_lds_addr + subk*2 + j*32*BN/2;
idx = idx ^ ((idx & 0b110000000000) >> 9);
idx = idx ^ ((idx & 0b1110000000) >> 4);
uint32_t* dst_ptr = reinterpret_cast<uint32_t*>(&smemoutput[idx]);
// uint32_t* dst_ptr = reinterpret_cast<uint32_t*>(&smemoutput[idx]);
uint32_t dst_ptr = *(reinterpret_cast<uint32_t*>(&smemoutput[idx]));
half (&res_)[2] = reinterpret_cast<half(&)[2]>(dst_ptr);
if (n < param.n && row < param.k && col < PQ) {
if constexpr (ksplit > 0) {
const uint outOffset = z * NKPQ +
n * KPQ +
row * PQ + col;
// output[outOffset] = smemoutput[idx];
output[outOffset] = reinterpret_cast<half *>(dst_ptr)[0];
// output[outOffset] = reinterpret_cast<half *>(dst_ptr)[0];
output[outOffset] = res_[0];
} else {
const uint outOffset = n * KPQ + row * PQ + col;
// output[outOffset] = smemoutput[idx];
output[outOffset] = reinterpret_cast<half *>(dst_ptr)[0];
// output[outOffset] = reinterpret_cast<half *>(dst_ptr)[0];
output[outOffset] = res_[0];
}
}
if (n < param.n && row+1 < param.k && col < PQ) {
@ -716,11 +722,13 @@ static __global__ void conv2d_implicit_kernel(const half * __restrict__ input,
n * KPQ +
(row+1) * PQ + col;
// output[outOffset] = smemoutput[idx];
output[outOffset] = reinterpret_cast<half *>(dst_ptr)[1];
// output[outOffset] = reinterpret_cast<half *>(dst_ptr)[1];
output[outOffset] = res_[1];
} else {
const uint outOffset = n * KPQ + (row+1) * PQ + col;
// output[outOffset] = smemoutput[idx];
output[outOffset] = reinterpret_cast<half *>(dst_ptr)[1];
// output[outOffset] = reinterpret_cast<half *>(dst_ptr)[1];
output[outOffset] = res_[1];
}
}
}