Skip to content

Fix BinaryOp broadcasting for packed dim #2653

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

Closed
wants to merge 1 commit into from
Closed

Fix BinaryOp broadcasting for packed dim #2653

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
14 changes: 11 additions & 3 deletions backends/vulkan/runtime/graph/ops/glsl/binary_op.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,12 @@ layout(set = 0, binding = 5) uniform PRECISION restrict OtherSizes {
}
other_sizes;

layout(set = 0, binding = 6) uniform PRECISION restrict Alpha {
layout(set = 0, binding = 6) uniform PRECISION restrict BroadcastParams {
ivec2 data;
}
broadcast_params;

layout(set = 0, binding = 7) uniform PRECISION restrict Alpha {
float data;
}
alpha;
Expand All @@ -63,8 +68,11 @@ void main() {
COORD_TO_POS_${PACKING}(other_coord, other_sizes.data),
0));

// Detect broadcasting
if (PACKED_DIM_${PACKING}(other_sizes.data) < PACKED_DIM_${PACKING}(in_sizes.data)) {
// Check boolean broadcast flags; we use ivec2 instead of bvec2 for alignment.
if (broadcast_params.data.x > 0) {
in_texel = in_texel.xxxx;
}
if (broadcast_params.data.y > 0) {
other_texel = other_texel.xxxx;
}

Expand Down
7 changes: 6 additions & 1 deletion backends/vulkan/runtime/graph/ops/impl/BinaryOp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,6 @@ void check_binary_op_args(
const vTensor& other,
const vTensor& out) {
VK_CHECK_COND(check_same_memory_layout(self, other, out));
VK_CHECK_COND(check_broadcastable(self, other));
std::vector<int64_t> broadcasted_sizes =
calculate_broadcasted_output_size(self, other);
VK_CHECK_COND(out.sizes() == broadcasted_sizes);
Expand All @@ -36,6 +35,8 @@ void resize_binary_op_node(
const std::vector<ValueRef>& extra_args) {
(void)extra_args;
vTensor& out = graph->get_val(args[0].refs[0]).toTensor();

// TODO(T183442143): Verify tensors are broadcastable.
vTensor& self = graph->get_val(args[1].refs[0]).toTensor();
vTensor& other = graph->get_val(args[1].refs[1]).toTensor();

Expand Down Expand Up @@ -73,6 +74,9 @@ void add_binary_op_node(
alpha_val = extract_scalar<float>(graph.get_val(alpha));
}

const api::utils::ivec2 broadcast_params =
create_broadcast_params(t_in1, t_in2);

std::stringstream kernel_name;
kernel_name << "binary_" << op_name;
apply_memory_layout_suffix(kernel_name, t_out);
Expand All @@ -90,6 +94,7 @@ void add_binary_op_node(
{t_out.gpu_sizes_ubo(),
t_in1.gpu_sizes_ubo(),
t_in2.gpu_sizes_ubo(),
graph.create_params_buffer(broadcast_params),
graph.create_params_buffer(alpha_val)},
// Resizing
resize_binary_op_node));
Expand Down
39 changes: 23 additions & 16 deletions backends/vulkan/runtime/graph/ops/impl/utils/TensorUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,28 +72,35 @@ bool check_same_memory_layout(
return (t1.gpu_memory_layout() == t3.gpu_memory_layout());
}

bool check_broadcastable(const vTensor& t1, const vTensor& t2) {
size_t ndim = std::max(t1.sizes().size(), t2.sizes().size());
//
// Broadcast flag functions
//

// Match the sizes in reverse because sizes are in NCHW order
for (int i = -1; i >= -ndim; --i) {
int64_t t1_size = api::utils::val_at(i, t1.sizes());
int64_t t2_size = api::utils::val_at(i, t2.sizes());
// If the sizes are not equal, one of them must be 1
if (t1_size != t2_size) {
if (t1_size > 1 && t2_size != 1) {
return false;
} else if (t2_size > 1 && t1_size != 1) {
return false;
}
}
bool is_packed_dim_broadcasted(const vTensor& sndr, const vTensor& rcvr) {
// We assume that the tensors are broadcastable. If values aren't equal at
// some index, then the value of rcvr is 1 and hence should be broadcasted.
switch (sndr.gpu_memory_layout()) {
case api::GPUMemoryLayout::TENSOR_CHANNELS_PACKED:
return api::utils::val_at(-3, sndr.sizes()) >
api::utils::val_at(-3, rcvr.sizes());
case api::GPUMemoryLayout::TENSOR_HEIGHT_PACKED:
return api::utils::val_at(-2, sndr.sizes()) >
api::utils::val_at(-2, rcvr.sizes());
case api::GPUMemoryLayout::TENSOR_WIDTH_PACKED:
return api::utils::val_at(-1, sndr.sizes()) >
api::utils::val_at(-1, rcvr.sizes());
}
}

return true;
api::utils::ivec2 create_broadcast_params(
const vTensor& t1,
const vTensor& t2) {
return api::utils::make_ivec2(
{is_packed_dim_broadcasted(t2, t1), is_packed_dim_broadcasted(t1, t2)});
}

//
// Work Group Size Calculation Utilities
// Work group size calculation functions
//

api::utils::uvec3 adaptive_work_group_size(
Expand Down
8 changes: 6 additions & 2 deletions backends/vulkan/runtime/graph/ops/impl/utils/TensorUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,10 +47,14 @@ bool check_same_memory_layout(
const vTensor& t2,
const vTensor& t3);

bool check_broadcastable(const vTensor& t1, const vTensor& t2);
//
// Broadcast flag functions
//

api::utils::ivec2 create_broadcast_params(const vTensor& t1, const vTensor& t2);

//
// Work Group Size Calculation Utilities
// Work group size calculation functions
//

api::utils::uvec3 adaptive_work_group_size(
Expand Down
6 changes: 5 additions & 1 deletion backends/vulkan/test/test_vulkan_delegate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,16 +146,20 @@ class AddModule(torch.nn.Module):
def __init__(self):
super().__init__()

def forward(self, x, y):
def forward(self, x, y, w):
z = x + y
z = z + x
z = z + x
z = z + w
z = w + z
z = z + 3 # test scalar broadcasting
return z

add_module = AddModule()
sample_inputs = (
torch.rand(size=(2, 3), dtype=torch.float32),
torch.rand(size=(2, 3), dtype=torch.float32),
torch.rand(size=(2, 1), dtype=torch.float32), # test broadcasting
)

self.lower_module_and_test_output(add_module, sample_inputs)
Expand Down
30 changes: 13 additions & 17 deletions backends/vulkan/test/vulkan_compute_api_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -549,7 +549,7 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
std::vector<int64_t> size_big = {12, 64, 64};
std::vector<int64_t> size_small = {12, 64, 64};

// Build graph
// Build graph and regularly check allocation counts

IOValueRef a = graph.add_input_tensor(
size_big,
Expand All @@ -560,9 +560,8 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
api::kFloat,
/*shared_object_idx = */ 4);

// Allocation count will be 6:
// 4: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() for each staging shader
// 2: staging buffer for each input tensor
// +4: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() for each staging shader
// +2: staging buffer for each input tensor
EXPECT_TRUE(get_vma_allocation_count() == 6);

ValueRef c = graph.add_tensor(
Expand All @@ -578,11 +577,10 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
api::kFloat,
/*shared_object_idx = */ 2);

// Allocation count will be 11, 5 are new:
// 2: out.gpu_sizes_ubo(), alpha UBO for arithmetic shader
// 2: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() uniform buffer for staging shader
// 1: staging buffer for the input tensor
EXPECT_TRUE(get_vma_allocation_count() == 11);
// +3: out.gpu_sizes_ubo(), alpha UBO, broadcast UBO for arithmetic shader
// +2: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() uniform buffer for staging shader
// +1: staging buffer for the input tensor
EXPECT_TRUE(get_vma_allocation_count() == 12);

ValueRef e = graph.add_tensor(
size_big,
Expand All @@ -596,18 +594,16 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
out.value = e;
out.staging = graph.set_output_tensor(out.value);

// Allocation count will be 15, 4 are new:
// 1: alpha UBO for arithmetic shader
// 2: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() for staging shader
// 1 staging buffer for the input tensor
EXPECT_TRUE(get_vma_allocation_count() == 15);
// +2: alpha UBO, broadcast UBO for arithmetic shader
// +2: t.gpu_sizes_ubo(), t.cpu_sizes_ubo() for staging shader
// +1 staging buffer for the input tensor
EXPECT_TRUE(get_vma_allocation_count() == 17);

graph.prepare();
graph.encode_execute();

// Allocation count will be 18, 3 are new:
// 3: shared memory allocations for tensors
EXPECT_TRUE(get_vma_allocation_count() == 18);
// +3: shared memory allocations for tensors
EXPECT_TRUE(get_vma_allocation_count() == 20);

// Run graph

Expand Down