Skip to content

fix L2 shared-memory indexing under the ir_mul layout #212

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
asglover wants to merge 1 commit into
base: main
Choose a base branch
from
+87 −6
Open

fix L2 shared-memory indexing under the ir_mul layout #212

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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: 7 additions & 6 deletions openequivariance/openequivariance/templates/loop_unroll_tp.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
{%- from 'macros.jinja' import layout_load, layout_store, reg_store with context %}
{%- from 'macros.jinja' import layout_load, layout_store, reg_store,
l2_smem_index with context %}
{%- from 'wmm.cuh' import generate_matmul %}

{%- macro generate_segment_kernel_forward(id, segment, warp_size) %}
Expand Down Expand Up @@ -51,7 +52,7 @@ __device__ __forceinline__ void forward_loop_unroll_{{id}}(IRREP_T* __restrict__

#pragma unroll
for(int j = 0; j < {{L2[v].ir.dim}}; j++)
l2_vec[j] = L2_smem[j + {{L2.slices()[v].start}} + k * {{L2[v].ir.dim}}] * weight;
l2_vec[j] = L2_smem[{{ l2_smem_index(problem.layout, L2[v].mul, L2[v].ir.dim, L2.slices()[v].start, 'k', 'j') }}] * weight;
{%- elif problem.instructions[k].connection_mode == "uvw" %}
{# Stream weights here #}
{%- set slice_size = L3[w].mul * L1[u].mul %}
Expand All @@ -61,7 +62,7 @@ __device__ __forceinline__ void forward_loop_unroll_{{id}}(IRREP_T* __restrict__
}
#pragma unroll
for(int j = 0; j < {{L2[v].ir.dim}}; j++)
l2_vec[j] = L2_smem[j + {{L2.slices()[v].start}} + k * {{L2[v].ir.dim}}];
l2_vec[j] = L2_smem[{{ l2_smem_index(problem.layout, L2[v].mul, L2[v].ir.dim, L2.slices()[v].start, 'k', 'j') }}];
{%- endif %}

// ----------------- CORE CALCULATION -----------------
Expand Down Expand Up @@ -184,11 +185,11 @@ __device__ __forceinline__ void forward_loop_unroll_{{id}}(IRREP_T* __restrict__
{%- if k == 0 or interactions[k][1] != interactions[k-1][1] or L2[v].mul > 1 or L1[u].mul != L1[interactions[k-1][0]].mul %}
#pragma unroll
for(int j = 0; j < {{L2[v].ir.dim}}; j++) {
l2_vec[j] = L2_smem[j + {{L2.slices()[v].start}} + k * {{L2[v].ir.dim}}];
l2_vec[j] = L2_smem[{{ l2_smem_index(problem.layout, L2[v].mul, L2[v].ir.dim, L2.slices()[v].start, 'k', 'j') }}];
l2_grad[j] = 0.0;

{%- if double_bwd %}
l2_original[j] = L2_original[j + {{L2.slices()[v].start}} + k * {{L2[v].ir.dim}}];
l2_original[j] = L2_original[{{ l2_smem_index(problem.layout, L2[v].mul, L2[v].ir.dim, L2.slices()[v].start, 'k', 'j') }}];
{%- endif %}
}
{%- endif %}
Expand Down Expand Up @@ -287,7 +288,7 @@ __device__ __forceinline__ void forward_loop_unroll_{{id}}(IRREP_T* __restrict__
if(lane_id == 0) {
#pragma unroll
for(int j = 0; j < {{L2[v].ir.dim}}; j++)
L2_grad_smem[j + {{L2.slices()[v].start}} + k * {{L2[v].ir.dim}}] += l2_grad[j];
L2_grad_smem[{{ l2_smem_index(problem.layout, L2[v].mul, L2[v].ir.dim, L2.slices()[v].start, 'k', 'j') }}] += l2_grad[j];
}
{%- endif %}

Expand Down
8 changes: 8 additions & 0 deletions openequivariance/openequivariance/templates/macros.jinja
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,14 @@ Keys map to lists of tuples with (name, dtype, num_elements) of each subarray.
{%- endif %}
{%- endmacro %}

{%- macro l2_smem_index(layout, mul, dim, start, mul_var, dim_var) -%}
{%- if layout == "ir_mul" -%}
{{mul_var}} + {{start}} + {{dim_var}} * {{mul}}
{%- else -%}
{{dim_var}} + {{start}} + {{mul_var}} * {{dim}}
{%- endif -%}
{%- endmacro %}

{%- macro declare_smem_variables(segment, smem_base) %}
{%- for name in segment.smem %}
{%- if name != "total" %}
Expand Down
36 changes: 36 additions & 0 deletions tests/batch_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,6 +330,42 @@ class TestIrMul(TPCorrectness):
internal_weights=False,
label="ir_mul_repr_13x1x13_l535",
),
oeq.TPProblem(
"32x1e",
"3x1e",
"32x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul3_l111",
),
oeq.TPProblem(
"32x1e",
"2x2e",
"32x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul2_l121",
),
oeq.TPProblem(
"16x2e",
"8x2e",
"16x2e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul8_l222",
),
oeq.TPProblem(
"16x1e",
"40x1e",
"16x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul40_l111",
),
]

@pytest.fixture(params=tpps, ids=lambda x: x.label, scope="class")
Expand Down
36 changes: 36 additions & 0 deletions tests/conv_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -345,6 +345,42 @@ class TestIrMulLayout(ConvCorrectness):
internal_weights=False,
label="ir_mul_repr_13x1x13_l535",
),
oeq.TPProblem(
"32x1e",
"3x1e",
"32x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul3_l111",
),
oeq.TPProblem(
"32x1e",
"2x2e",
"32x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul2_l121",
),
oeq.TPProblem(
"16x2e",
"8x2e",
"16x2e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul8_l222",
),
oeq.TPProblem(
"16x1e",
"40x1e",
"16x1e",
[(0, 0, 0, "uvu", True)],
shared_weights=False,
internal_weights=False,
label="ir_mul_L2mul40_l111",
),
]

@pytest.fixture(params=production_model_tpps, ids=lambda x: x.label, scope="class")
Expand Down
Loading