Bring Your Own Datatypes #5812

gussmith23 · 2020-06-15T07:52:43Z

Expands the functionality of the Bring Your Own Datatypes framework, originally introduced in #2900.

Enables the conversion of large workloads (e.g. whole models) to use a custom datatype
Expands API for registering custom datatypes and the lowering functions which implement them
Adds tests
Adds tutorial

Old task list:

tests/python/unittest/test_target_custom_datatypes.py

src/tir/transforms/lower_custom_datatypes.cc

tests/python/unittest/test_custom_datatypes_change_dtype.py

Ravenwater · 2020-07-30T10:39:39Z

@gussmith23 the posit standard is moving towards posit<8,2>, posit<16,2>, posit<32,2>, posit<64,2> specifically to make conversions better, faster, cheaper.

gussmith23 · 2020-07-30T14:26:31Z

Thanks, Theo! We'll be making heavy use of posits to test the framework, so we'll be sure to use the right configurations. Gus

…

On Thu, Jul 30, 2020, 3:39 AM Theodore Omtzigt ***@***.***> wrote: @gussmith23 <https://github.com/gussmith23> the posit standard is moving towards posit<8,2>, posit<16,2>, posit<32,2>, posit<64,2> specifically to make conversions better, faster, cheaper. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5812 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJFZAPAGX422Y663R3S2P3R6FEXVANCNFSM4N55HAPQ> .

gussmith23 · 2020-07-30T15:51:23Z

@tqchen looking for your input on testing (or feel free to point us to someone else who could advise us!)

@hypercubestart is working on making the BYODT tests more thorough. Yesterday we had a discussion about whether it's better to test entire networks (i.e. run resnet with posits, run resnet with floats, compare the results) or test operator-by-operator, or both. Do you have an opinion?

hypercubestart · 2020-08-01T19:38:29Z

tests/python/unittest/test_custom_datatypes_change_dtype.py

+    # Simplifying inference is essential right now, as batch norms (which get
+    # removed) are broken with custom datatypes.
+    #expr = relay.ir_pass.simplify_inference(expr)


@gussmith23 why is this here?

For a while (not sure if still true -- we should set up a test for it) batch norms were broken with custom datatypes. It was a very odd bug that I didn't end up figuring out. It seemed to go pretty deep, and be related to the code being generated. However, the batch norm operator can generally be optimized out when you're running a network for inference (and not training), which is what simplify_inference did. So we'd run simplify_inference to remove the batch norms and things would be hunky-dory.

@gussmith23 Still need to document this somewhere.

tests/python/unittest/test_custom_datatypes_change_dtype.py

3rdparty/nop-type/nop-type.cc

CMakeLists.txt

gussmith23 · 2020-08-12T23:18:02Z

Makefile


+# Test scripts
+pyunittest:
+	./tests/scripts/task_python_unittest.sh


Is this just a convenience thing?

I might have added this. We can probably remove it.

@gussmith23 remove

python/tvm/relay/backend/_backend.py

tests/python/unittest/test_custom_datatypes_change_dtype.py

gussmith23 · 2020-08-12T23:54:08Z

tests/python/unittest/test_custom_datatypes_change_dtype.py

+    # #expr = relay.ir_pass.simplify_inference(expr)
+
+def run_conv2d(src_dtype, dst_dtype):
+    def run_test_conv2d(src_dtype,


Should we be using your compare() here?

tests/python/unittest/test_custom_datatypes_change_dtype.py

hypercubestart · 2020-08-13T04:16:50Z

tests/python/unittest/test_custom_datatypes.py

+    #           'float32',
+    #           'custom[posites2]32',
+    #           num_classes=10)
+    # run_model(get_resnet, (3, 32, 32),


resnet (x is posits, y is floats):

Max absolute difference: 0.01120126 Max relative difference: 0.10072963 x: array([[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]], dtype=float32) y: array([[0.097013, 0.097102, 0.105917, 0.099155, 0.095107, 0.096329, 0.097079, 0.111201, 0.102866, 0.09823 ]], dtype=float32)

@gussmith23 should we simply increase atol/rtol or look more into why posits are less accurate?

This isn't right! Something's going wrong. This sounds dumb, but I'm going to just try and apply SimplifyInference and see if it fixes things...

We should debug small Resnet. Something is going wrong. The easiest way to debug is to return early from the Resnet constructor:

modified python/tvm/relay/testing/resnet.py @@ -162,6 +162,7 @@ def resnet(units, data = relay.var("data", shape=data_shape, dtype=dtype) data = layers.batch_norm_infer(data=data, epsilon=2e-5, scale=False, name='bn_data') (_, _, height, _) = data_shape + return relay.Function(relay.analysis.free_vars(data), data) if height <= 32: # such as cifar10 body = layers.conv2d( data=data, channels=filter_list[0], kernel_size=(3, 3), @@ -194,7 +195,6 @@ def resnet(units, flat = relay.nn.batch_flatten(data=pool1) fc1 = layers.dense_add_bias(data=flat, units=num_classes, name='fc1') net = relay.nn.softmax(data=fc1) - return relay.Function(relay.analysis.free_vars(net), net) def get_net(batch_size,

This should work well with the current test file.
You can move the early return up and down to see where things diverge.
In my limited testing, it looks like posits are going to NaN after this first batch norm! Running the above version of resnet produces NaNs, but if we move the early return one layer up, then things are fine:

modified python/tvm/relay/testing/resnet.py @@ -160,6 +160,7 @@ def resnet(units, num_unit = len(units) assert num_unit == num_stages data = relay.var("data", shape=data_shape, dtype=dtype) + return relay.Function(relay.analysis.free_vars(data), data) data = layers.batch_norm_infer(data=data, epsilon=2e-5, scale=False, name='bn_data') (_, _, height, _) = data_shape if height <= 32: # such as cifar10 @@ -194,7 +195,6 @@ def resnet(units, flat = relay.nn.batch_flatten(data=pool1) fc1 = layers.dense_add_bias(data=flat, units=num_classes, name='fc1') net = relay.nn.softmax(data=fc1) - return relay.Function(relay.analysis.free_vars(net), net) def get_net(batch_size,

This version of Resnet literally doesn't do anything, and just returns the input. But we know it works, and we can see that things go wrong after that first batch norm.

My first thought was "oh, we need to run SimplifyInference!" So I did:

modified tests/python/unittest/test_custom_datatypes.py @@ -49,6 +49,8 @@ def change_dtype(src, dst, module, params): return module, params def compare(module, input, src_dtype, dst_dtype, rtol, atol, params = {}): + module = relay.transform.SimplifyInference()(module) + ex = relay.create_executor("graph", mod=module) correct = ex.evaluate()(*input, **params)

This fixes the batch norm...somewhat. We still get numerical errors between the outputs of the batch norms:

AssertionError: Not equal to tolerance rtol=0.0001, atol=0.0001 Mismatch: 99.2% Max absolute difference: 0.01599717 Max relative difference: 0.01600081 x: array([[[[0.692229, 0.071629, 0.014329, ..., 0.984048, 0.871523, 0.445013], [0.728206, 0.627117, 0.563761, ..., 0.378826, 0.326279,... y: array([[[[0.681328, 0.070501, 0.014103, ..., 0.968551, 0.857798, 0.438005], [0.716737, 0.61724 , 0.554882, ..., 0.37286 , 0.321141,...

There shouldn't be this amount of numerical error between posit32es2 and float32 -- posit32es2 should be far more accurate than float32 at values around [0,1].

Here's a debugging list:

I think we do still need to run SimplifyInference explicitly. This is confusing, because it's run in Optimize()...maybe it's not getting run?

Add SimplifyInference somewhere (maybe in compare() like I did? I think that might be the best place)

Figure out whether SimplifyInference is seemingly not getting run later on

I think we should add a test for batch norm itself, because it's been such a thorn in the side of this project. Have a test that runs just a batch norm through your compare(). Once we fix batch norm, hopefully everything else will be fixed.

To debug the numerical issues with batch norm, here are a few dead-simple ideas off the top of my head:

Make sure that the workload is using the posit32 C functions. Maybe your create_lower_func() call is lowering them to the wrong bitwidth?

Make sure the params going into the batch norm are the same between float32 and posit32. (Batch norms take four parameters: mean, var, beta, gamma)

Let me know where you get with this! Good luck!

hypercubestart · 2020-08-13T04:19:27Z

tests/python/unittest/test_custom_datatypes.py

+    # run_model(get_inception, (3, 32, 32),
+    #           'float32',
+    #           'custom[posites2]32',
+    #           num_classes=10)


error: relay.concatenate requires all tensors have the same shape on non-concatenating axes

will need to take a closer look

turns out using default input size of (3, 299, 299) works, and doesn't run too slow. however (x is posit, y is float):

Mismatched elements: 4 / 10 (40%) Max absolute difference: 0.00023866 Max relative difference: 0.00239235 x: array([[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]], dtype=float32) y: array([[0.100027, 0.100217, 0.099981, 0.099959, 0.100235, 0.099952, 0.100033, 0.099761, 0.09995 , 0.099885]], dtype=float32)

Again, this isn't good. If you look at how softmax works (which is the last operator in these networks), part of what it does is normalizing a tensor so that it sums to 1. So the fact that the values are all coming out 0.1 means that all the values are equal going into the softmax, and are probably very different from the "true" values that go into the softmax in the f32 case.

One way to debug this (which I'm going to do now, but may be useful to you in the future) is to change the definition of the network, and have the network constructor return early. I.e. here, we can return the body of the network before the softmax is added, to check whether it's the softmax that is broken.

Strange, the default size is running REALLY slowly on my machine. On the order of minutes!

That's normal and expected -- that's how it always was. I'm just confused as to why it's running faster on your machine?

3rdparty/nop-type/nop-type.cc

tests/python/unittest/test_custom_datatypes_change_dtype.py

gussmith23 · 2020-08-13T00:14:05Z

Makefile


+# Test scripts
+pyunittest:
+	./tests/scripts/task_python_unittest.sh


I might have added this. We can probably remove it.

gussmith23 · 2020-08-13T00:17:16Z

python/tvm/relay/testing/mobilenet.py

    bias = relay.var('fc_bias')
    fc = relay.nn.dense(data=flatten, weight=weight, units=num_classes)
    fc = relay.nn.bias_add(fc, bias)
+    # TODO(gus) i think softmax is broken


Yes, remove this

gussmith23 · 2020-08-13T00:20:55Z

python/tvm/target/datatype.py

+                                        op.call_type == _Call.PureIntrinsic):
+            return _Call(dtype, extern_func_map[t.bits], convert(op.args),
+                              _Call.Extern)
+        return _Call(dtype, extern_func_map[t.bits], convert([op.a, op.b]),


Let's figure out -- what happens in the unary case?
How does this work for the unary case?
Are all of the unary ops intrinsics?
Does convert([op.a, op.b]) handle the case where op.b is none?

all unary ops are intrinsics

gussmith23 · 2020-08-13T00:21:48Z

python/tvm/target/datatype.py

+                intrinsic_name=None):
    """Register an external function which computes the given op.

    Currently, this will only work with Casts and binary expressions


Does it make sense to mention a and b? Is it just confusing to the reader?

tests/python/unittest/test_custom_datatypes_change_dtype.py

gussmith23 · 2020-08-13T00:34:36Z

tests/python/unittest/test_custom_datatypes_change_dtype.py

+    # run_model(get_resnet, (3, 32, 32),
+    #           'float32',
+    #           'custom[posit32]32',
+    #           num_classes=10)


Experiment with these -- are they slow?

If they're slow -- we need a contingency plan.
If we can't run whole models, then we should be testing layer by layer
--> partly achieved by testing unary/binary ops
--> but then we should also test the larger operators i.e. convolution (in which case we should beef up/update conv2d tests)

hypercubestart · 2020-08-19T02:38:23Z

tests/python/unittest/test_custom_datatypes.py

+            (8, 32): 'Posit8es2ToFloat',
+        }), 
+        "Cast", "llvm", "posites2", "float")
+    register_op(create_lower_func(


do we need this cast from int to posites2 since we don't actually use it at all

Sure, if it's not used we can remove it!

3rdparty/posit/posit-wrapper.cc

gussmith23 · 2020-08-19T23:28:48Z

3rdparty/posit/posit-wrapper.cc

+}
+
+extern "C" {
+TVM_DLL uint32_t RawPosit32es2(uint32_t in) { return in; }


What are these for?

convert uint_8 holding raw bits to posit number. this is useful for register_min_func, which has overflow errors otherwise

It's confusing me as is, because it's taking in a uint32_t and returning a uint32_t -- is there something I'm missing? It doesn't seem to be doing anything at the moment!

input is a IntImm, this is basically doing reinterpret_cast<posit>

gussmith23 · 2020-08-19T23:33:17Z

python/tvm/relay/frontend/change_datatype.py

+    def change_dtype(src, dst, expr, params):
+        cdtype = ChangeDatatype(src, dst)
+        expr = cdtype.visit(expr)
+        expr = relay.ir_pass.infer_type(expr)
+        params = dict((p, tvm.nd.array(params[p].asnumpy().astype(dst))) for p in params)
+        return expr, params


@gussmith23 update this code

gussmith23 · 2020-08-19T23:38:47Z

python/tvm/target/datatype.py


-    type_name : str
+    src_type_name : str
        The name of the custom datatype, e.g. posit (but not custom[posit]8).


Update to posites2

gussmith23 · 2020-08-19T23:40:47Z

python/tvm/target/datatype.py

+                          + src_type_name
    tvm._ffi.register_func(lower_func_name, lower_func)

+# TODO(gus) could probably make this a decorator if i want


@gussmith23 Document

gussmith23 · 2020-08-19T23:46:25Z

python/tvm/target/datatype.py

+    t = DataType(type_str)
+    return t.bits
+
+def lower_ite(ite_intrin):


@gussmith23 Document

python/tvm/target/datatype.py

gussmith23 · 2020-08-19T23:51:36Z

src/target/datatype/registry.h

   * same code. Generally, this should be straightforward, as the user will be manually registering
   * all of their custom types.
-   * \param type_name The name of the type, e.g. "bfloat"
+   * \param type_name The name of the type, e.g. "posit"


Change to posites2

tests/python/unittest/test_custom_datatypes.py

gussmith23 · 2020-08-20T16:30:58Z

@slyubomirsky I heard from Zach that you have been looking into how documentation is done in modern TVM. I need to document the datatypes framework, which is distributed across multiple files, and I'm trying to decide on a central place to keep the documentation. I'm wondering if anything you've learned recently would be relevant here?

python/tvm/relay/frontend/change_datatype.py

hypercubestart · 2020-08-24T23:46:52Z

tests/python/unittest/test_custom_datatypes.py

+    # run_model(get_mobilenet, (3, 224, 224), 'float32', 'custom[posit8]8')
+    # run_model(get_mobilenet, (3, 224, 224), 'float32', 'custom[posit32]32')
+    # run_model(get_inception, (3, 299, 299), 'float32', 'custom[posit32]32')
+    # run_model(get_resnet, (3, 224, 224), 'float32', 'custom[posit32]32')


can we remove these?

Let's talk about it -- I think we should at least leave a few around, to document which models are too slow to run. We can move them down, though.

hypercubestart

remove

hypercubestart · 2020-08-25T00:31:44Z

python/tvm/target/datatype.py

+        if isinstance(op, _Cast):
+            src_bits = bit_length(op.value.dtype)
+            return call_pure_extern(dtype, extern_func_map[(src_bits, t.bits)], op.value)
+        if isinstance(op, _FloatImm):
+            return call_pure_extern(dtype, extern_func_map[t.bits], op.value)
+        if isinstance(op, _Call):
+            return call_pure_extern(dtype, extern_func_map[t.bits], *op.args)
+        if isinstance(op, _BinaryOpExpr):
+            return call_pure_extern(dtype, extern_func_map[t.bits], op.a, op.b)


@gussmith23 should we improve debugging message here?

if the map does not contain the bit_width, this throws a KeyError on extern_func_map[t.bits], which may be somewhat cryptic to the user

Yeah, please do! Better error messages are never a bad idea 😄

hypercubestart · 2020-08-25T01:41:53Z

src/arith/rewrite_simplify.cc

+    // TODO(@gussmith23) is this ok?
+    // CHECK(op->dtype.is_float());


why do we need to do this @gussmith23 ?

i think a better idea may be to do
CHECK(op->dtype.is_float() || datatype::Registry::Global()->GetTypeRegistered(op->dtype.code()));

also, it seems FloatImm(custom_datatype, value) represents a custom_datatype with that float value. see here: https://github.com/gussmith23/tvm/blob/a45f7bb1975933118db7647261a6ddefb214595a/include/tvm/tir/op.h#L829, we freely mutate FloatImm as a double

jroesch · 2020-08-26T16:37:01Z

@slyubomirsky I heard from Zach that you have been looking into how documentation is done in modern TVM. I need to document the datatypes framework, which is distributed across multiple files, and I'm trying to decide on a central place to keep the documentation. I'm wondering if anything you've learned recently would be relevant here?

@hogepodge maybe you can chime in

slyubomirsky · 2020-08-26T19:05:10Z

@slyubomirsky I heard from Zach that you have been looking into how documentation is done in modern TVM. I need to document the datatypes framework, which is distributed across multiple files, and I'm trying to decide on a central place to keep the documentation. I'm wondering if anything you've learned recently would be relevant here?

As Jared said, @hogepodge would be best positioned to comment. What you are describing sounds like an "explainer" (describing the overall design and rationale as opposed to a tutorial), which should go in the docs directory with similar other "explainers."

hogepodge · 2020-08-26T19:19:48Z

At a high level, the docs organization is described in this discussion thread.

https://discuss.tvm.ai/t/rfc-tvm-documentation-refactor/7456

It sounds like the document you're producing is either an explainer or a reference guide. If it's a design guide, or a higher level overview of the design, explainer is the best classification for it. If it's actually documenting APIs and is meant to be used as a reference for building on top of it, treat it like a reference file. @gussmith23 I think the best thing to do is send a PR with the documentation and add me to the review. We can sort out where it belongs or how to organize it once we see what it looks like.

gussmith23 · 2020-08-26T22:38:12Z

Thanks all, will do!

python/tvm/target/datatype.py

hypercubestart · 2020-09-25T22:19:41Z

Bumping on this for review:
cc: @tqchen @jroesch @tmoreau89 @junrushao1994 @zhiics @ZihengJiang @hogepodge @comaniac

tqchen

last few minor comment, other parts lgtm

tqchen · 2020-09-25T22:32:58Z

cmake/config.cmake

 # Contrib libraries
 #---------------------------------------------
+# Whether to build with posit custom datatype
+set(USE_POSIT OFF)


consider to change USE_BYOC_POSIT given it is a software emulated version.

tqchen · 2020-09-25T22:33:33Z

LICENSE

 Apache Software Foundation License 2.0
 --------------------------------------

-3rdparty/bfloat16/bfloat16.cc


please also remove https://github.com/apache/incubator-tvm/blob/master/licenses/LICENSE.bfloat16.txt

gussmith23 · 2020-09-28T15:05:19Z

thanks to everyone involved!! so cool to see this work finally integrated into TVM :)

…

On Sat, Sep 26, 2020, 3:49 PM ziheng ***@***.***> wrote: Merged #5812 <#5812> into master. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5812 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJFZAPIOGWIBKUJVTU5N23SHZVXZANCNFSM4N55HAPQ> .

hypercubestart · 2020-10-01T03:32:59Z

@tqchen are tutorial docs generated using the Docker gpu image?
Is so, can we install the Universal repo to both the Docker cpu/gpu images so that the tutorial can use the posit codepaths?

* Add ChangeDatatype pass and unittest * [WIP] Jared's work on Fri This was work that Jared did on my computer, trying to get Inception v3 running. * Fix simplify inference to work over different data types. * Formatting * Copy setup code from other test file * Logging in Relay * Remove duplicate TVM_DLL * Add Sub, Mul, Div, Max to bfloat lib * Fix previous broken rebased commit * Remove line * Add LowerCustomDatatypes to build passes * Upcast ints to custom datatypes too, as well as to floats * Add and use convert_ndarray * Lower Call * Relay: create constant scalars of custom dtypes We use the same method we use in TVM: store the value in a double. * Custom datatype formatting in Relay * Update unittests * Add simpler example that's not working yet * Add Python unittests to Makefile * Fix bug * Fix function name in GetPackedFunc call * convert_ndarray makes its own executor * Add simple test case * Move setup() calls * Use convert_ndarray * Change import to make it more specific * Fix another Registry::Get call * Allow users to register minimum functions for custom datatypes This commit allows users to register global functions named `tvm.datatype.min.<type name>` which take the number of bits in the custom type and return the corresponding minimum value (as a double). A similar commit will need to be created for max, whenever that ends up being needed! * Remove check for float * Add test * Fix inception test * Add MobileNet * Lower custom datatypes before intrinsics * Add exp and sqrt bfloat functions * [buggy commit] Lower intrinsics like sqrt, exp This commit has bugs in it, I'm fairly certain. * Formatting * Fix bug * Add lowering for new ops in test * Add int to bfloat * Remove print * Add all tests * Correct image size * Add TODO * Add "notbfloat" type This type is for testing purposes. It just stores a float in a uint32. It was used to confirm the fact that my bfloat "implementation" is very numerically unstable and was causing issues when running the model. * Convert arguments Not sure how necessary this actually is. * Rewrite custom datatype constants in Relay * Add test_ops * Print constants in Relay * Use topi.testing * Test conv2d * Add test_model * Comment out model tests * Register notbfloat This could be unregistered at some point later * Add commented code Remove later * Add posit tests * test_ops_same_function * [temporary] move incomplete commit to macbook * Add more to tests * Formatting * Uncomment add * Remove bad tests * Change comments * Change function name and docstring * Change main function * Restructure tests * Fix visibility of posit functions * YAPF * Switching keywords around to resolve build errors on some systems * Improve test by running smaller mobilenet * Add test_cast * Change datatype name; add simple test * Rename to posit32 * Merge 3 posit types into one file * Add a nop type * Remove bfloat * Refactor test comments * Refactor conv2d test * Add optional tolerance arguments * Add posit8 and posit16 * Add comment about posit8 * Whoops -- actually add noptype to CMakeLists * Add rtol, atol to run_workload * Add noptype to tests * Run noptype over other models, too * Pass correct arguments to calls * Fix line length errors * Raise tolerances (again) to avoid flaky test * fix style * add test for tanh, log, sigmoid * Remove references to bfloat, notbfloat * Change comments * Remove old test file * fix min func * refactoring unit test file * use posits es2 * cleanup * comment * coment if_then_else * support different bit widths * use random seed to create stable tests * update documentation * removed nop-type and code consistency * add batchnorm test * rebase and update * fix tests and format * pylint * change order of include * include order * fix style * remove posit c linkage * update universal * fix style * fix test * fix overflow error with minfunc and posits * style * use change_dtype to convert params * update universal * fix fatal error * fix constant repr * minor update to posites2 * update universal * fix rst * fix invalid import and sqrt * update universal * comments * comments and expand testing * increase atol/rtol for custom[posites2]32 * Re-add newline * Remove comment * Remove opt level and comment * Change docstring * Add TODO * Add file header and newline * Update docstring * Update file docstring * Update docstrings * Delete todos * create_min_lower_func * add better debugging message * docs * add BYODT tutorial * add todo * Reformat some of tutorial to RST, plus code fixes * tutorial notebook runs now * fix hyperlink * rebase * add to tutorial * fix mobilenet model * add skip tag * black lint * add compiler flag and add dummy float * myfloat and posites2 test * remove universal * lint * lint * add setup * build with USE_POSIT for CI/CD * fix posit cmake * add cd / * undo docker changes * change tutorial to use myfloat * move files * lint * fix * remove filter * fix lint * fix suggestions Co-authored-by: Jared Roesch <roeschinc@gmail.com> Co-authored-by: Andrew Liu <andrewlliu@gmail.com>

gussmith23 commented Jul 23, 2020

View reviewed changes

tests/python/unittest/test_target_custom_datatypes.py Outdated Show resolved Hide resolved

gussmith23 commented Jul 29, 2020

View reviewed changes

src/tir/transforms/lower_custom_datatypes.cc Outdated Show resolved Hide resolved

gussmith23 commented Jul 30, 2020

View reviewed changes

hypercubestart reviewed Aug 1, 2020

View reviewed changes

gussmith23 commented Aug 10, 2020

View reviewed changes

tests/python/unittest/test_custom_datatypes_change_dtype.py Outdated Show resolved Hide resolved

gussmith23 commented Aug 12, 2020

View reviewed changes

hypercubestart reviewed Aug 13, 2020

View reviewed changes

gussmith23 commented Aug 14, 2020

View reviewed changes

hypercubestart force-pushed the custom-datatypes branch from 6b385e9 to 2a93563 Compare August 16, 2020 00:02

hypercubestart mentioned this pull request Aug 17, 2020

Build errors when integrated with downstream TVM stillwater-sc/universal#144

Closed

hypercubestart reviewed Aug 19, 2020

View reviewed changes

gussmith23 commented Aug 20, 2020

View reviewed changes

hypercubestart force-pushed the custom-datatypes branch 2 times, most recently from 85ac76c to b0fc673 Compare August 23, 2020 18:23

hypercubestart reviewed Aug 24, 2020

View reviewed changes

python/tvm/relay/frontend/change_datatype.py Show resolved Hide resolved

hypercubestart reviewed Aug 24, 2020

View reviewed changes

python/tvm/relay/frontend/change_datatype.py Show resolved Hide resolved

hypercubestart reviewed Aug 24, 2020

View reviewed changes

hypercubestart reviewed Aug 25, 2020

View reviewed changes

hypercubestart force-pushed the custom-datatypes branch from e647731 to c5c77c5 Compare August 27, 2020 18:24

hypercubestart reviewed Aug 27, 2020

View reviewed changes

python/tvm/target/datatype.py Show resolved Hide resolved

hypercubestart added 14 commits September 23, 2020 09:26

remove universal

d3e9168

lint

25850a3

lint

2c34863

add setup

6998ab7

build with USE_POSIT for CI/CD

3b9fff2

fix posit cmake

e2898fe

add cd /

f96c95e

undo docker changes

baa453b

change tutorial to use myfloat

30b8b6e

move files

142a055

lint

21b7da9

fix

eba7389

remove filter

47fde5e

fix lint

3967ed9

hypercubestart force-pushed the custom-datatypes branch from 969b9f5 to 3967ed9 Compare September 23, 2020 16:26

tqchen reviewed Sep 25, 2020

View reviewed changes

fix suggestions

ccdc65b

tqchen approved these changes Sep 25, 2020

View reviewed changes

ZihengJiang approved these changes Sep 25, 2020

View reviewed changes

ZihengJiang merged commit 5aafff9 into apache:master Sep 26, 2020

hypercubestart mentioned this pull request Oct 9, 2020

[Docker][CI][BYODT] add universal to Docker image #6654

Merged

		// TODO(@gussmith23) is this ok?
		// CHECK(op->dtype.is_float());

Bring Your Own Datatypes #5812

Bring Your Own Datatypes #5812

Uh oh!

Conversation

gussmith23 commented Jun 15, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Ravenwater commented Jul 30, 2020

Uh oh!

gussmith23 commented Jul 30, 2020 via email

Uh oh!

gussmith23 commented Jul 30, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gussmith23 Aug 14, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

hypercubestart Aug 13, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gussmith23 Aug 14, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

gussmith23 commented Jun 15, 2020 •

edited

Loading

gussmith23 Aug 14, 2020 •

edited

Loading

hypercubestart Aug 13, 2020 •

edited

Loading

gussmith23 Aug 14, 2020 •

edited

Loading

hypercubestart Aug 26, 2020 •

edited

Loading