Merge branch 'master' into repo-refactor-fix-kernels-tests

flexflow · Jan 17, 2025 · f07a906 · f07a906
2 parents 1225f15 + 8f64a0f
commit f07a906
Show file tree

Hide file tree

Showing 103 changed files with 1,352 additions and 354 deletions.
diff --git a/lib/kernels/include/kernels/array_shape.h b/lib/kernels/include/kernels/array_shape.h
@@ -3,7 +3,7 @@
 
 #include "legion_dim.h"
 #include "op-attrs/tensor_shape.dtg.h"
-#include "utils/stack_vector.h"
+#include "utils/stack_vector/stack_vector.h"
 #include "utils/visitable.h"
 #include <cstddef>
 #include <optional>

diff --git a/lib/kernels/src/legion_dim.cc b/lib/kernels/src/legion_dim.cc
@@ -7,7 +7,7 @@ legion_dim_t add_to_legion_dim(legion_dim_t legion_dim, int value) {
 }
 
 legion_dim_t legion_dim_from_ff_dim(ff_dim_t ff_dim, int num_dimensions) {
-  return legion_dim_t(num_dimensions - ff_dim.value - 1);
+  return legion_dim_t(num_dimensions - ff_dim.value.get_value() - 1);
 }
 
 } // namespace FlexFlow
diff --git a/lib/kernels/test/src/test_concat_kernel.cc b/lib/kernels/test/src/test_concat_kernel.cc
@@ -7,7 +7,7 @@ TEST_SUITE(FF_TEST_SUITE) {
   TEST_CASE("Test concat kernel forward and backward") {
     size_t num_inputs = 3;
     size_t size_per_input = 100;
-    ff_dim_t concat_axis = ff_dim_t(0);
+    ff_dim_t concat_axis = ff_dim_t{nonnegative_int{0}};
 
     ManagedPerDeviceFFHandle managed_handle{};
     ManagedFFStream managed_stream{};

diff --git a/lib/kernels/test/src/test_transpose_kernel.cc b/lib/kernels/test/src/test_transpose_kernel.cc
@@ -7,7 +7,8 @@ TEST_SUITE(FF_TEST_SUITE) {
   TEST_CASE("Test Transpose Kernel Operations") {
     std::size_t num_dims = 2;
 
-    std::vector<ff_dim_t> perm = {ff_dim_t(0), ff_dim_t(1)};
+    std::vector<ff_dim_t> perm = {ff_dim_t{nonnegative_int{0}},
+                                  ff_dim_t{nonnegative_int{1}}};
 
     ManagedPerDeviceFFHandle managed_handle{};
     ManagedFFStream managed_stream{};

diff --git a/lib/local-execution/include/local-execution/legion_tensor_shape.h b/lib/local-execution/include/local-execution/legion_tensor_shape.h
@@ -3,9 +3,9 @@
 
 #include "kernels/legion_dim.h"
 #include "op-attrs/datatype.h"
-#include "op-attrs/ff_dim.h"
+#include "op-attrs/ff_dim_t.h"
 #include "op-attrs/tensor_shape.dtg.h"
-#include "utils/stack_vector.h"
+#include "utils/stack_vector/stack_vector.h"
 #include "utils/visitable.h"
 #include <cstddef>
 

diff --git a/lib/local-execution/src/legion_tensor_shape.cc b/lib/local-execution/src/legion_tensor_shape.cc
@@ -1,14 +1,15 @@
 #include "local-execution/legion_tensor_shape.h"
+#include "kernels/legion_dim.h"
 #include "op-attrs/tensor_shape.h"
 
 namespace FlexFlow {
 
 legion_dim_t legion_dim_from_ff_dim(ff_dim_t ff_dim, size_t num_dims) {
-  return legion_dim_t(num_dims - ff_dim.value - 1);
+  return legion_dim_t(num_dims - ff_dim.value.get_value() - 1);
 }
 
 legion_dim_t legion_dim_from_ff_dim(ff_dim_t ff_dim, TensorShape const &shape) {
-  return legion_dim_t(num_dims(shape) - ff_dim.value - 1);
+  return legion_dim_from_ff_dim(ff_dim, num_dims(shape));
 }
 
 } // namespace FlexFlow
diff --git a/lib/local-execution/src/ops/linear.cc b/lib/local-execution/src/ops/linear.cc
@@ -1,7 +1,7 @@
 #include "linear.h"
 #include "kernels/linear_kernels.h"
 #include "local-execution/task_argument_accessor.h"
-#include "op-attrs/ff_dim.h"
+#include "op-attrs/ff_dim_t.h"
 #include "op-attrs/get_output_shapes.h"
 #include "utils/exception.h"
 #include "utils/hash-utils.h"
@@ -66,8 +66,8 @@ static DeviceSpecificDeviceStates
   auto input = acc.get_tensor<Permissions::RO>(INPUT);
   auto weight = acc.get_tensor<Permissions::RO>(WEIGHT);
   auto output = acc.get_tensor<Permissions::WO>(OUTPUT);
-  int out_dim = output.shape.at(ff_dim_t{0});
-  int batch_size = output.shape.at(ff_dim_t{1});
+  int out_dim = output.shape.at(ff_dim_t{nonnegative_int{0}});
+  int batch_size = output.shape.at(ff_dim_t{nonnegative_int{1}});
 
   float *one_ptr;
 
@@ -96,8 +96,8 @@ static std::optional<float> forward_task_impl(TaskArgumentAccessor const &acc) {
   ProfilingSettings profiling = acc.get_argument<ProfilingSettings>(PROFILING);
   auto attrs = acc.get_argument<LinearAttrs>(ATTRS);
 
-  int in_dim = input.shape.at(ff_dim_t{0}) + 1;
-  int out_dim = output.shape.at(ff_dim_t{0}) + 1;
+  int in_dim = input.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
+  int out_dim = output.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
   int batch_size = output.shape.get_volume() / out_dim;
 
   float const *bias_ptr = NULL;
@@ -140,8 +140,8 @@ static std::optional<float>
     bias_ptr = bias.get_float_ptr();
   }
 
-  int in_dim = input.shape.at(ff_dim_t{0}) + 1;
-  int out_dim = output.shape.at(ff_dim_t{0}) + 1;
+  int in_dim = input.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
+  int out_dim = output.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
   int batch_size = output.shape.get_volume() / out_dim;
 
   return profile(backward_kernel,

diff --git a/lib/local-execution/src/ops/pool_2d.cc b/lib/local-execution/src/ops/pool_2d.cc
@@ -30,14 +30,14 @@ static DeviceSpecificDeviceStates
   auto input = acc.get_tensor<Permissions::RO>(INPUT);
   auto output = acc.get_tensor<Permissions::WO>(OUTPUT);
 
-  int input_w = input.shape.at(ff_dim_t(0)) + 1;
-  int input_h = input.shape.at(ff_dim_t(1)) + 1;
-  int input_c = input.shape.at(ff_dim_t(2)) + 1;
-  int input_n = input.shape.at(ff_dim_t(3)) + 1;
-  int output_w = output.shape.at(ff_dim_t(0)) + 1;
-  int output_h = output.shape.at(ff_dim_t(1)) + 1;
-  int output_c = output.shape.at(ff_dim_t(2)) + 1;
-  int output_n = output.shape.at(ff_dim_t(3)) + 1;
+  int input_w = input.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
+  int input_h = input.shape.at(ff_dim_t{nonnegative_int{1}}) + 1;
+  int input_c = input.shape.at(ff_dim_t{nonnegative_int{2}}) + 1;
+  int input_n = input.shape.at(ff_dim_t{nonnegative_int{3}}) + 1;
+  int output_w = output.shape.at(ff_dim_t{nonnegative_int{0}}) + 1;
+  int output_h = output.shape.at(ff_dim_t{nonnegative_int{1}}) + 1;
+  int output_c = output.shape.at(ff_dim_t{nonnegative_int{2}}) + 1;
+  int output_n = output.shape.at(ff_dim_t{nonnegative_int{3}}) + 1;
 
   printf("init pool (input): n(%d) c(%d) h(%d) "
          "w(%d)\n",

diff --git a/lib/local-execution/src/ops/reverse.cc b/lib/local-execution/src/ops/reverse.cc
@@ -53,11 +53,11 @@ static std::optional<float> forward_task_impl(TaskArgumentAccessor const &acc) {
   coord_t in_blk_size = 1, reverse_dim_size = 1, num_out_blks = 1;
   for (int i = 0; i < output.shape.get_dim(); i++) {
     if (i < axis.value) {
-      in_blk_size *= output.shape.at(ff_dim_t(i));
+      in_blk_size *= output.shape.at(ff_dim_t{nonnegative_int{i}});
     } else if (i == axis.value) {
-      reverse_dim_size = output.shape.at(ff_dim_t(i));
+      reverse_dim_size = output.shape.at(ff_dim_t{nonnegative_int{i}});
     } else {
-      num_out_blks *= output.shape.at(ff_dim_t(i));
+      num_out_blks *= output.shape.at(ff_dim_t{nonnegative_int{i}});
     }
   }
 
@@ -79,15 +79,15 @@ static std::optional<float>
   auto output_grad = acc.get_tensor_grad<Permissions::RO>(OUTPUT);
   auto attrs = acc.get_argument<ReverseAttrs>(ATTRS);
 
-  int axis = input_grad.shape.get_dim() - attrs.axis.value - 1;
+  int axis = input_grad.shape.get_dim() - attrs.axis.value.get_value() - 1;
   coord_t in_blk_size = 1, reverse_dim_size = 1, num_out_blks = 1;
   for (int i = 0; i < input_grad.shape.get_dim(); i++) {
     if (i < axis) {
-      in_blk_size *= input_grad.shape.at(ff_dim_t(i));
+      in_blk_size *= input_grad.shape.at(ff_dim_t{nonnegative_int{i}});
     } else if (i == axis) {
-      reverse_dim_size = input_grad.shape.at(ff_dim_t(i));
+      reverse_dim_size = input_grad.shape.at(ff_dim_t{nonnegative_int{i}});
     } else {
-      num_out_blks *= input_grad.shape.at(ff_dim_t(i));
+      num_out_blks *= input_grad.shape.at(ff_dim_t{nonnegative_int{i}});
     }
   }
 

diff --git a/lib/local-execution/src/ops/softmax.cc b/lib/local-execution/src/ops/softmax.cc
@@ -64,8 +64,13 @@ static DeviceSpecificDeviceStates
   int output_c = output.shape.at(legion_dim_t(2));
   int output_n = output.shape.at(legion_dim_t(3));
 
-  SoftmaxPerDeviceState per_device_state = init_kernel(
-      handle, attrs.dim.value, output_n, output_c, output_h, output_w);
+  SoftmaxPerDeviceState per_device_state =
+      init_kernel(handle,
+                  attrs.dim.value.get_value(),
+                  output_n,
+                  output_c,
+                  output_h,
+                  output_w);
 
   return DeviceSpecificDeviceStates{
       DeviceSpecific<SoftmaxPerDeviceState>::create(per_device_state)};

diff --git a/lib/local-execution/src/ops/split.cc b/lib/local-execution/src/ops/split.cc
@@ -47,11 +47,11 @@ OpTaskInvocation backward(SplitAttrs const &attrs) {
 void calc_block_size(coord_t &num_blocks,
                      coord_t &block_size,
                      ArrayShape const &array_shape,
-                     int axis) {
+                     ff_dim_t axis) {
   num_blocks = 1;
   block_size = 1;
   for (int d = 0; d < array_shape.num_elements(); d++) {
-    if (d <= axis) {
+    if (d <= axis.value.get_value()) {
       block_size *= array_shape.at(legion_dim_t(d));
     } else {
       num_blocks *= array_shape.at(legion_dim_t(d));
@@ -66,12 +66,12 @@ static std::optional<float> forward_task_impl(TaskArgumentAccessor const &acc) {
   auto attrs = acc.get_argument<SplitAttrs>(ATTRS);
 
   coord_t num_blocks, in_block_size, out_block_size[MAX_NUM_OUTPUTS];
-  calc_block_size(num_blocks, in_block_size, input.shape, attrs.axis.value);
+  calc_block_size(num_blocks, in_block_size, input.shape, attrs.axis);
 
   for (int i = 0; i < attrs.splits.size(); i++) {
     coord_t out_num_blocks;
     calc_block_size(
-        out_num_blocks, out_block_size[i], output.shape, attrs.axis.value);
+        out_num_blocks, out_block_size[i], output.shape, attrs.axis);
   }
   float *output_float_ptr = output.get_float_ptr();
   return profile(forward_kernel,
@@ -94,12 +94,11 @@ static std::optional<float>
   auto attrs = acc.get_argument<SplitAttrs>(ATTRS);
 
   coord_t num_blocks, in_block_size, out_block_size[MAX_NUM_OUTPUTS];
-  calc_block_size(
-      num_blocks, in_block_size, input_grad.shape, attrs.axis.value);
+  calc_block_size(num_blocks, in_block_size, input_grad.shape, attrs.axis);
   for (int i = 0; i < attrs.splits.size(); i++) {
     coord_t out_num_blocks;
     calc_block_size(
-        out_num_blocks, out_block_size[i], output_grad.shape, attrs.axis.value);
+        out_num_blocks, out_block_size[i], output_grad.shape, attrs.axis);
   }
   float const *output_grad_ptr = output_grad.get_float_ptr();
   return profile(backward_kernel,

diff --git a/lib/local-execution/src/ops/transpose.cc b/lib/local-execution/src/ops/transpose.cc
@@ -17,6 +17,7 @@
 #include "kernels/transpose_kernels.h"
 #include "op-attrs/get_output_shapes.h"
 #include "op-attrs/ops/transpose.h"
+#include "utils/integer_conversions.h"
 
 using namespace FlexFlow::Kernels::Transpose;
 
@@ -39,8 +40,17 @@ OpTaskInvocation init(TransposeAttrs const &attrs) {
 static DeviceSpecificDeviceStates
     init_task_impl(TaskArgumentAccessor const &acc) {
   auto const &attrs = acc.get_argument<TransposeAttrs>(ATTRS);
-  std::vector<ff_dim_t> perm = inner_to_outer_idxs(attrs.perm);
-  TransposePerDeviceState per_device_state = init_kernel(perm.size(), perm);
+  int size = int_from_size_t(attrs.perm.size());
+
+  std::vector<ff_dim_t> perm = [&] {
+    std::vector<ff_dim_t> result;
+    for (int i : range(size)) {
+      result.push_back(ff_dim_t{nonnegative_int{size - i - 1}});
+    }
+    return result;
+  }();
+
+  TransposePerDeviceState per_device_state = init_kernel(size, perm);
 
   return DeviceSpecificDeviceStates{
       DeviceSpecific<TransposePerDeviceState>::create(per_device_state)};