Cast Op

[KateUnger]

Description of changes:

• Updates Cast operation

TODO: Change cast_kernels.cpp and cast_kernels.cu

Related Issues:

Linked Issues:

• Issue Update Cast operator #1046

Issues closed by this PR:

• Closes Update Cast operator #1046, Cast Measure Operator Cost #1167

---

This change is 

[lockshaw]

Reviewed 3 of 3 files at r1, 1 of 4 files at r2.
Reviewable status: 1 of 4 files reviewed, 9 unresolved discussions (waiting on @KateUnger)

---

lib/kernels/include/kernels/cast_kernels.h line 10 at r1 (raw file):

struct CastPerDeviceState {
  PerDeviceFFHandle handle;

Is there a reason this is being placed in the per device state rather than getting passed in as a function argument?

---

lib/kernels/src/cuda/cast_kernels.cu line 64 at r2 (raw file):

};

CastPerDeviceState init_kernel(PerDeviceFFHandle const &handle) {

Where'd the data type arguments go?

---

lib/kernels/src/cuda/cast_kernels.cu line 69 at r2 (raw file):

void forward_kernel(ffStream_t stream,
                    CastPerDeviceState const *m,

Pass by const reference not pointer

---

lib/runtime/src/ops/cast.h line 45 at r1 (raw file):

#endif

// template <typename IDT>

Delete

---

lib/runtime/src/ops/cast.h line 126 at r1 (raw file):

/* }; */

// void Cast::backward(FFModel const &ff) {

Delete

---

lib/runtime/src/ops/cast.cc line 32 at r1 (raw file):

enum Slots { INPUT, OUTPUT, ATTRS, PROFILING, PER_DEVICE_STATE, HANDLE };

// declare Legion names

Delete

---

lib/runtime/src/ops/cast.cc line 110 at r1 (raw file):

                 profiling,
                 "[Cast] forward_time = %.2lfms\n",
                 &per_device_state,

Should be passed as a reference not a ptr

---

lib/runtime/src/ops/cast.cc line 134 at r1 (raw file):

                 profiling,
                 "[Cast] forward_time = %.2lfms\n",
                 &per_device_state,

Pass as reference

---

lib/runtime/src/ops/cast.cc line 154 at r1 (raw file):

  auto env = sim.new_environment();

  // Assume cast has no cost

This isn't necessarily true--can you create an issue to track fixing this?

[KateUnger]

Reviewable status: 1 of 4 files reviewed, 9 unresolved discussions (waiting on @lockshaw)

---

lib/kernels/include/kernels/cast_kernels.h line 10 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Is there a reason this is being placed in the per device state rather than getting passed in as a function argument?

Done.

---

lib/kernels/src/cuda/cast_kernels.cu line 64 at r2 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Where'd the data type arguments go?

They're passed in with forward and backward

---

lib/kernels/src/cuda/cast_kernels.cu line 69 at r2 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Pass by const reference not pointer

Deleted PerDeviceState

---

lib/runtime/src/ops/cast.h line 45 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Delete

Done.

---

lib/runtime/src/ops/cast.h line 126 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Delete

Done.

---

lib/runtime/src/ops/cast.cc line 32 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Delete

Done.

---

lib/runtime/src/ops/cast.cc line 110 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Should be passed as a reference not a ptr

Done.

---

lib/runtime/src/ops/cast.cc line 134 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

Pass as reference

Done.

---

lib/runtime/src/ops/cast.cc line 154 at r1 (raw file):

Previously, lockshaw (Colin Unger) wrote…

This isn't necessarily true--can you create an issue to track fixing this?

Done.

[lockshaw]

Reviewed 1 of 4 files at r2, 2 of 3 files at r3, 2 of 2 files at r4, all commit messages.
Reviewable status: all files reviewed, 1 unresolved discussion (waiting on @KateUnger)

---

lib/kernels/src/hip/cast_kernels.cpp line 82 at r4 (raw file):

                    DataType input_type,
                    DataType output_type,
                    PerDeviceFFHandle handle) {

Why pass as value vs as const &?

[reyna-abhyankar]

Reviewed all commit messages.
Reviewable status: all files reviewed, 1 unresolved discussion (waiting on @KateUnger)

[lockshaw]

Reviewed 4 of 4 files at r5, all commit messages.
Reviewable status: all files reviewed, 1 unresolved discussion (waiting on @KateUnger)

[lockshaw]

Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @KateUnger)