Skip to content

speedup tabulate cuda kernel by reducing shm using #830

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

Merged
amcadmus merged 3 commits into from
Jul 7, 2021
Merged

speedup tabulate cuda kernel by reducing shm using #830

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
dc735f0
reduced the shm used in tabulate_fusion_fifth_order_polynomial cuda k…
darelbeida Jul 6, 2021
ad3e30e
formatted `MTILE` and `KTILE` used in tabulate kernels
darelbeida Jul 6, 2021
7d2be2d
formatted `warp_idx` used in tabulate kernel
darelbeida Jul 6, 2021
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
42 changes: 19 additions & 23 deletions source/lib/src/cuda/tabulate.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,17 +73,13 @@ __global__ void tabulate_fusion_fifth_order_polynomial(
const int nnei,
const int last_layer_size)
{
extern __shared__ int _data[];
const int block_idx = blockIdx.x; // nloc
const int thread_idx = threadIdx.x; // last_layer_size
FPTYPE ago = __shfl_sync(0xffffffff, em_x[block_idx * nnei + nnei - 1], 0);
bool unloop = false;
int breakpoint = nnei - 1;
FPTYPE * iteratorC = (FPTYPE*) &_data[0];
for (int kk = 0; kk < MTILE; kk++)
iteratorC[kk * last_layer_size + thread_idx] = 0.f;
__syncthreads();

FPTYPE sum[MTILE] = {0.f};
for (int ii = 0; ii < nnei; ii++) {
FPTYPE var[6];
FPTYPE xx = em_x[block_idx * nnei + ii];
Expand All @@ -102,12 +98,12 @@ __global__ void tabulate_fusion_fifth_order_polynomial(
FPTYPE res = var[0] + (var[1] + (var[2] + (var[3] + (var[4] + var[5] * xx) * xx) * xx) * xx) * xx;

for (int kk = 0; kk < MTILE; kk++) {
iteratorC[kk * last_layer_size + thread_idx] += (nnei - breakpoint) * em[block_idx * nnei * MTILE + ii * MTILE + kk] * res;
sum[kk] += (nnei - breakpoint) * em[block_idx * nnei * MTILE + ii * MTILE + kk] * res;
}
if (unloop) break;
}
for (int ii = 0; ii < MTILE; ii++) {
out[block_idx * MTILE * last_layer_size + ii * last_layer_size + thread_idx] = iteratorC[ii * last_layer_size + thread_idx];
out[block_idx * MTILE * last_layer_size + ii * last_layer_size + thread_idx] = sum[ii];
}
}

Expand All @@ -133,8 +129,8 @@ __global__ void tabulate_fusion_grad_fifth_order_polynomial(
extern __shared__ int _data[];
const int block_idx = blockIdx.x; // nloc
const int thread_idx = threadIdx.x; // KTILE * WARP_SIZE, usally 128 here~
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / WARP_SIZE, 0);
int lane_idx = threadIdx.x % WARP_SIZE;
int breakpoint = nnei - 1;
bool unloop = false;
FPTYPE * iteratorA = (FPTYPE *)&_data[0]; // dy
Expand All @@ -145,16 +141,16 @@ __global__ void tabulate_fusion_grad_fifth_order_polynomial(
}
__syncthreads();
FPTYPE ago = __shfl_sync(0xffffffff, em_x[block_idx * nnei + nnei - 1], 0);
for (int ii = 0; ii < nnei; ii += KTILE) {
FPTYPE xx = em_x[block_idx * nnei + ii + warp_idx];
for (int ii = warp_idx; ii < nnei; ii += KTILE) {
FPTYPE xx = em_x[block_idx * nnei + ii];
if (ago == xx) {
unloop = true;
breakpoint = ii + warp_idx;
breakpoint = ii;
}

int table_idx = 0;
locate_xx(xx, table_idx, lower, upper, max, stride0, stride1);
FPTYPE sum[KTILE] = {0.f};
FPTYPE sum[MTILE] = {0.f};
FPTYPE Csub = 0.f;
for (int jj = lane_idx; jj < last_layer_size; jj += WARP_SIZE) {
FPTYPE var[6];
Expand All @@ -167,25 +163,25 @@ __global__ void tabulate_fusion_grad_fifth_order_polynomial(
var[5] = table[table_idx * last_layer_size * 6 + 6 * jj + 5];
FPTYPE res = var[0] + (var[1] + (var[2] + (var[3] + (var[4] + var[5] * xx) * xx) * xx) * xx) * xx;

for (int kk = 0; kk < KTILE; kk++) {
for (int kk = 0; kk < MTILE; kk++) {
sum[kk] += (nnei - breakpoint) * iteratorA[kk * last_layer_size + jj] * res;
}
res = em[block_idx * nnei * MTILE + (ii + warp_idx) * 4 + 0] * iteratorA[0 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + (ii + warp_idx) * 4 + 1] * iteratorA[1 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + (ii + warp_idx) * 4 + 2] * iteratorA[2 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + (ii + warp_idx) * 4 + 3] * iteratorA[3 * last_layer_size + jj];
res = em[block_idx * nnei * MTILE + ii * 4 + 0] * iteratorA[0 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + ii * 4 + 1] * iteratorA[1 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + ii * 4 + 2] * iteratorA[2 * last_layer_size + jj];
res += em[block_idx * nnei * MTILE + ii * 4 + 3] * iteratorA[3 * last_layer_size + jj];
Csub += (nnei - breakpoint) * (var[1] + (2 * var[2] + (3 * var[3] + (4 * var[4] + 5 * var[5] * xx) * xx) * xx) * xx) * res;
}
__syncwarp();
for (int kk = 0; kk < KTILE; kk++) {
for (int kk = 0; kk < MTILE; kk++) {
warp_reduce(sum[kk]);
}
warp_reduce(Csub);
if (lane_idx == 0) {
for (int kk = 0; kk < KTILE; kk++) {
dy_dem[block_idx * nnei * MTILE + (ii + warp_idx) * 4 + kk] = sum[kk];
for (int kk = 0; kk < MTILE; kk++) {
dy_dem[block_idx * nnei * MTILE + ii * 4 + kk] = sum[kk];
}
dy_dem_x[block_idx * nnei + ii + warp_idx] = Csub;
dy_dem_x[block_idx * nnei + ii] = Csub;
}
if (unloop) break;
}
Expand All @@ -204,7 +200,7 @@ void tabulate_fusion_gpu_cuda(
const int last_layer_size)
{
if (nloc <= 0) {return;}
tabulate_fusion_fifth_order_polynomial<FPTYPE, MM, KK> <<<nloc, last_layer_size, sizeof(FPTYPE) * MM * last_layer_size>>>(
tabulate_fusion_fifth_order_polynomial<FPTYPE, MM, KK> <<<nloc, last_layer_size>>>(
out,
table, em_x, em, table_info[0], table_info[1], table_info[2], table_info[3], table_info[4], nnei, last_layer_size);
}
Expand Down