Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Resolve race conditions in Marlin kernel #11493

Merged
merged 5 commits into from
Jan 2, 2025
Merged

Resolve race conditions in Marlin kernel #11493

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5389c53
Resolve race conditions in Marlin kernel
wchen61 Dec 24, 2024
ab586fb
Add reduce size when verify configs
wchen61 Dec 26, 2024
3a6e41f
Optimize the caculation of reduce shared memory size
wchen61 Jan 2, 2025
aa2f07a
Fix clang format issue
wchen61 Jan 2, 2025
4762127
Optimize caclulation of shared memory size for reduction
wchen61 Jan 2, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 21 additions & 19 deletions csrc/quantization/gptq_marlin/gptq_marlin.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -834,6 +834,7 @@ __global__ void Marlin(
int4* sh_g_idx = sh_b + (stages * b_sh_stage);
int4* sh_zp = sh_g_idx + (stages * g_idx_stage);
int4* sh_s = sh_zp + (stages * zp_sh_stage);
int4* sh_red = sh_s + (stages * s_sh_stage);

// Register storage for double buffer of shared memory reads.
FragA frag_a[2][thread_m_blocks];
Expand Down Expand Up @@ -932,11 +933,11 @@ __global__ void Marlin(
int4* sh_s_stage = sh_s + s_sh_stage * pipe;

if constexpr (group_blocks >= thread_k_blocks) {
if (s_sh_wr_pred) {
cp_async4(&sh_s_stage[s_sh_wr], &scales_ptr[s_gl_rd]);
}
// Only fetch scales if this tile starts a new group
if (pipe % (group_blocks / thread_k_blocks) == 0) {
if (s_sh_wr_pred) {
cp_async4(&sh_s_stage[s_sh_wr], &scales_ptr[s_gl_rd]);
}
if ((pipe + 1) % (group_blocks / thread_k_blocks) == 0) {
s_gl_rd += s_gl_rd_delta;
}
} else {
Expand Down Expand Up @@ -1038,9 +1039,7 @@ __global__ void Marlin(
// No act-order case
if constexpr (group_blocks != -1) {
if constexpr (group_blocks >= thread_k_blocks) {
int4* sh_s_stage =
sh_s + s_sh_stage * ((group_blocks / thread_k_blocks) *
(pipe / (group_blocks / thread_k_blocks)));
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
} else {
int warp_id = threadIdx.x / 32;
Expand Down Expand Up @@ -1339,15 +1338,15 @@ __global__ void Marlin(
int red_sh_wr =
red_sh_delta * j + (red_sh_rd - red_sh_stride * i);
if (i < red_off) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * j + red_sh_rd]);
float* c_wr = reinterpret_cast<float*>(&sh[red_sh_wr]);
float* c_rd = reinterpret_cast<float*>(
&sh_red[red_sh_delta * j + red_sh_rd]);
float* c_wr = reinterpret_cast<float*>(&sh_red[red_sh_wr]);
#pragma unroll
for (int k = 0; k < 4; k++)
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + j][k] +=
c_rd[k] + c_wr[k];
}
sh[red_sh_wr] =
sh_red[red_sh_wr] =
reinterpret_cast<int4*>(&frag_c)[4 * 2 * m_block + j];
}
}
Expand All @@ -1357,7 +1356,7 @@ __global__ void Marlin(
#pragma unroll
for (int i = 0; i < 4 * 2; i++) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * i + red_sh_rd]);
reinterpret_cast<float*>(&sh_red[red_sh_delta * i + red_sh_rd]);
#pragma unroll
for (int j = 0; j < 4; j++)
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + i][j] +=
Expand Down Expand Up @@ -1397,7 +1396,7 @@ __global__ void Marlin(
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
cp_async4_pred(
&sh[c_sh_wr + c_sh_wr_delta * i],
&sh_red[c_sh_wr + c_sh_wr_delta * i],
&C[c_gl_wr + c_gl_wr_delta_o * (i / 2) +
c_gl_wr_delta_i * (i % 2)],
i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m);
Expand All @@ -1410,7 +1409,7 @@ __global__ void Marlin(
for (int i = 0; i < thread_m_blocks * 4; i++) {
if (i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m) {
if (!first) {
int4 c_red = sh[c_sh_wr + i * c_sh_wr_delta];
int4 c_red = sh_red[c_sh_wr + i * c_sh_wr_delta];
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<float*>(
Expand Down Expand Up @@ -1461,10 +1460,10 @@ __global__ void Marlin(
float* frag_c_ptr = reinterpret_cast<float*>(&frag_c);
#pragma unroll
for (int k = 0; k < th_size; k++) {
sh[threadIdx.x] =
sh_red[threadIdx.x] =
C_tmp[c_cur_offset + active_threads * k + threadIdx.x];

float* sh_c_ptr = reinterpret_cast<float*>(&sh[threadIdx.x]);
float* sh_c_ptr = reinterpret_cast<float*>(&sh_red[threadIdx.x]);
#pragma unroll
for (int f = 0; f < 4; f++) {
frag_c_ptr[k * 4 + f] += sh_c_ptr[f];
Expand Down Expand Up @@ -1515,7 +1514,7 @@ __global__ void Marlin(
res = __hmul2(res, s[0]);
}

((scalar_t2*)sh)[idx] = res;
((scalar_t2*)sh_red)[idx] = res;
};

if (threadIdx.x / 32 < thread_n_blocks / 4) {
Expand Down Expand Up @@ -1543,7 +1542,7 @@ __global__ void Marlin(
i < div_ceil(16 * thread_m_blocks, threads / (2 * thread_n_blocks));
i++) {
if (c_gl_wr < c_gl_wr_end) {
C[c_gl_wr] = sh[c_sh_rd];
C[c_gl_wr] = sh_red[c_sh_rd];
c_gl_wr += c_gl_wr_delta;
c_sh_rd += c_sh_rd_delta;
}
Expand Down Expand Up @@ -1865,9 +1864,12 @@ bool is_valid_cache_size(thread_config_t const& th_config, int max_m_blocks,

float pipe_size = (a_size + b_size) * pipe_stages;

float reduce_size = max(th_config.num_threads * 32 * 4,
(tb_n / 64) * 32 * (tb_max_m / 16) * 4 * 2 * 4 * 2);

TORCH_CHECK(max_shared_mem / 2 > scales_cache_size); // Sanity

return pipe_size < 0.95f * (max_shared_mem - scales_cache_size);
return pipe_size + reduce_size < 0.95f * (max_shared_mem - scales_cache_size);
}

bool is_valid_config(thread_config_t const& th_config, int max_m_blocks,
Expand Down
Loading