Skip to content

added a jinja condition to bypass optimization ic caching #96

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

Open
wants to merge 4 commits into
base: lii/test-v2-opt
Choose a base branch
from
Open

added a jinja condition to bypass optimization ic caching #96

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
2e6d077
added a jinja condition to bypass optimization ic caching
kudomcho Mar 4, 2025
d70fc0e
moved is_cache inside the jinja condition
kudomcho Mar 5, 2025
40423e4
switched the condition of is_cache for applying optimization
kudomcho Mar 6, 2025
3b7bfe3
removed space
kudomcho Mar 6, 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
13 changes: 9 additions & 4 deletions fbgemm_gpu/codegen/training/forward/embedding_forward_split_kernel_v2_template.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -768,8 +768,10 @@ __global__ void split_embedding_codegen_forward_{{ wdesc }}_v2_kernel(
if (is_zero_total_L) {
const uint32_t D_start = D_offsets[t] / VEC_WIDTH;
const uint32_t load_D = (D_offsets[t + 1] / VEC_WIDTH) - D_start;
{%- if is_rocm %}
const uint32_t num_warps_per_row = DIV_ROUND_UP(load_D, (kWarpSize/2));
const auto placement = static_cast<PlacementType>(weights_placements[t]);
const auto is_cache = placement == PlacementType::MANAGED_CACHING;
{%- if is_rocm %}
const uint32_t num_warps_per_row = is_cache ? DIV_ROUND_UP(load_D, kWarpSize):DIV_ROUND_UP(load_D, (kWarpSize/2));
{%- else %}
const uint32_t num_warps_per_row = DIV_ROUND_UP(load_D, kWarpSize);
{% endif %}
Expand Down Expand Up @@ -813,15 +815,18 @@ __global__ void split_embedding_codegen_forward_{{ wdesc }}_v2_kernel(
load_D = (D_offsets[t + 1] / VEC_WIDTH) - D_start;
}
load_D = shfl_sync(load_D, 0);
{%- if is_rocm %}
const uint32_t num_warps_per_row = DIV_ROUND_UP(load_D, (kWarpSize/2));
const auto placement = static_cast<PlacementType>(weights_placements[t]);
const auto is_cache = placement == PlacementType::MANAGED_CACHING;
{%- if is_rocm %}
const uint32_t num_warps_per_row = is_cache ? DIV_ROUND_UP(load_D, kWarpSize):DIV_ROUND_UP(load_D, (kWarpSize/2));
{%- else %}
const uint32_t num_warps_per_row = DIV_ROUND_UP(load_D, kWarpSize);
{% endif %}
if (table_warp_id >= num_warps_per_row * (is_small_L ? num_warps_for_small_L : B)) {
return;
}


// Compute d (same for all Ls)
const uint32_t load_d = (table_warp_id % num_warps_per_row) * kWarpSize;
// Compute sample ID
Expand Down