GitLab

Improves performance for 4/6 GEMMs used by huggingface BERT models with batch_size>1 by using a non-batched rocBLAS call for GEMMs where the B input has a broadcasted batch dimension.
The four verify tests added reflect the actual configurations used by bert-base-cased, with varied batch sizes.

Also adds a matcher to simplify_reshapes to move multibroadcasts after concats.

* Fix accuracy bug when vectorizing slices

Using not and or improves readability. The cppcheck rule will help ensure we are doing it consistently.

Rewrite_gelu pass replaces the gelu formula of x * (1/2) * (1 + erf(x/sqrt(2))) with the sigmoid approximation of x * Sigmoid(x * 1.702)

This will rewrite dot operators like X(Y + b) to XY + Xb when b is constant as we can fold the add away.
This improves handling pointwise with broadcasted operators, this helps improves const propagation.
Improve gemm fusion with a mul_add
Improve support for broadcast shapes in gemm

* Add is_supported to the target
* Add get_target_assignments
* Rename assignment to target_assignments
* Add ref target header to test
* Add fpga target
* Make context const in compute

*In the verification tests, check that saving and reloading the program is the same program. This also fixes serialization to always load instructions in the same order. There is also fixes for deconv and quant_conv which didn't save the solution id, and was broken for serialization.

Updated each source file in the repo with the existing license.

prioritizing int8 over int8x4 when it is applicable
Amend return to continue in apply loop
Adding error handling in case int8x4 compilation failed
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>

* Upgrade to cppcheck 2.8

Add ref and gpu implementations for ONNX op GatherND

Resolves #1032

There is significant improvement on larger tensors with half almost 50% faster:

lens: [1024, 384, 768]
gpu::code_object[code_object=13832,symbol_name=kernel,global=39321600,local=256,]: 1.16685ms
gpu::reduce_sum[axes={2}]: 1.73126ms
Also for non-trivial layouts this can sometimes be over 2x faster:

lens: [64, 1024, 768, 4]
gpu::code_object[code_object=13832,symbol_name=kernel,global=39321600,local=256,]: 1.1706ms
gpu::reduce_sum[axes={1}]: 2.63375ms
Of course if the stride becomes larger this speed improvement diminishes due to poor memory access patterns. A lane_reduce instead of a block_reduce is needed for such type of kernels. I plan to address that in a future PR.

Finally, this also includes a MIGRAPHX_GPU_DUMP_ASM env variable which will print out the assembly when the kernel compiles.

Issue 1127 Updates the math.hpp header file to perform overloads of various standard functions (ops) for the hip half2 type. The half2 type is two 16-bit floats packed into a 32-bit number and therefore the overloads act on vectors of sizes that are multiples of 2. They are invoked in runtime compilation any time one of the ops is called on a tensor declared with the data type shape::half_type.

Defined new template, made instances of the template for those math operations that the hip library contains, added verify tests for the sqrt operator for three cases:

tensor size not divisible by 2
tensor size divisible by 2 but not by 4
tensor size divisible by 4

Change the "scatter" struct and op to a base/child set of three: scatter_none, scatter_add, scatter_mul to mirror Onnx' ScatterElements op. and its three reduction options. (Onnx Scatter op is deprecated and is equivalent to scatter_none.)

Provides both a reference op. and update to Onnx parsing. Tests updated and new test case added.

This adds the infrastructure so we can compile everything in parallel, whereas before only pointwise kernels were compiled in parallel. This will also directly integrate with lowering and the gpu-driver. The kernels for pointwise and roialign are using this infrastructure. Scatternd is not since it does require standard shape.

This also makes it easier to add new runtime compiled kernels in the future.

Add exclusive and reverse modes to gpu implementation of prefix_scan_sum, which completes support for ONNX op CumSum

Changed the pooling values for two structures from strings to specialized enum classes. Many test and operator parsing changes to support this. Introduces one new source file, op_enums.cpp.

Add onnx parser and ref and gpu implementations of ONNX op ScatterND

Implements the IsNaN operator, ref, gpu, and onnx parser.

Make doc/CMakeLists.txt standalone
Switch to use rocm-cmake modules for document generation
Add CONFIGURE_DEPENDS to file(GLOB) so it will update without an explicit cmake run
Add STRINGS property for build type to make it easier to switch build types with ccmake
Various fixes and improvements

There is now a MIGRAPHX_DISABLE_POINTWISE_FUSION to disable it

Changed MessagePack file extensions to mxr.

allow nonstd shape for the arg ops, non-standard shapes include broadcast, slice and transpose

Make clip a pointwise op

Fix pointwise compile error with half sqrt 

Resolves a problem in parsing the ssd-10 model.

The problem is, after inserting contiguous in the auto_contiguous pass, standard output shape of some operators becomes non-standard. Then, if the next operator requires standard input shape, an exception is throw.

For example, if we pass the following model:
Input (standard shape) -> transpose (transposed) -> softmax (transposed) -> transpose (standard) -> gather.
It works fine, and no contiguous is required.

In the auto_contiguous pass, a contiguous is inserted after the first transpose. Then we need to replace the first transpose with the contiguous and recompute all shapes. When it comes to the gather operator, its input is a transposed shape, and an exception is thrown.

The solution is in the recompute_shape() function. If it is called by the auto_contiguous pass and shape of an instruction is changed, and the shape is non_standard, we do not recompute shape of its output. The reason is: since its output shape is non_standard, a contiguous op will be added after the instruction, which will recompute shape for later operators.

This PR is to turn on a few gemm unit test with int8 input datatype. Before rocm4.4, int8 input data type requires matrix size to be no less than 4 in rocblas implementation. Because of this limitation, we turned off a few gemm unit tests with int8 input data type.

This limitation is removed in rocm4.4, so after we upgrade to rocm4.5, we can turn on these unit tests. Also we change to unit test conv_bn_add to adding instructions to module instead of program.
Co-authored-by: kahmed10 <15948690+kahmed10@users.noreply.github.com>

This PR is the ref implementation of the nonmaxsuppression operator. It always returns the max possible output shape, which is the problem tracked in issue #948.

GPU implementation of the roialign operator, using the jit approach to reduce the lib size.

Implementation of the roialign operator. For now, we have only the ref implementation. When we run a model on the GPU, we fall back the execution to use the ref implementation.

This PR is for the nonzero operator with static output shape.  
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

Previously dot operator was defined as C = alpha * A . B + beta * C where * is scalar multiplication and . is dot product or matrix multiplication depending on dimension of the inputs.

Aim is to have the definition of dot operator as C = A . B without having alpha or beta.

In order to achieve the same effect as alpha and beta (1) it multiplies the one of the inputs to the dot operator with alpha value. (2) if beta is present then, multiplies the C with beta and then adds into the output from step 1.

Add multinomial op to onnx parser with ref and GPU implementations.

The onnx parser inserts a literal of shape {batch_size, sample_size} with random values in the range [0, 1) and inserts existing ops to compute the cumulative density function. The multinomial operator multiplies the random values by the sum of the CDF and returns the index of the first element of the CDF that is greater than the result, representing samples randomly drawn from [0, class_size) that follow the log-probability distribution.

Resolves #821
Co-authored-by: Shucai Xiao <shucai@gmail.com>

Checks wavefront size, then changes implementation and number of threads for DPP reduce

This reverts commit 9e43cb8b.