Merge pull request #9 from histogrammar/run-in-gpu

Run in gpu

Author: jpivarski

README.md

@@ -25,32 +25,34 @@ Status
 
 The tests are thorough.
 
-   * `original.py` were written during development to test all features as they were added.
-   * `autogenerated.py` is from the language-independent testing suite ([histogrammar-multilang](https://github.com/histogrammar/histogrammar-multilang)), which provides greater coverage, value-explicitness in the test script, and cross-language agreement.
+   * `basic.py` were written during development to test all features as they were added.
    * `testnumpy.py` tests numerical agreement between the conventional implementation and the Numpy implementation, which are very different. Also tests much larger datasets and infinity/NaN handling.
      * contrary to its name, `testnumpy.py` also compares its implementation with the literal code given in [the specification](http://histogrammar.org/docs/specification/) as well.
-
-Primitive implementation is mature. Notes in the "Numpy" column are rough speedup factors for a range of tests.
-
-| Primitive         | Pure Python | Numpy            |
-|:------------------|:------------|:-----------------|
-| Count             | done        | 100X             |
-| Sum               | done        | 40-100X          |
-| Average           | done        | 40-100X          |
-| Deviate           | done        | 40-80X           |
-| Minimize          | done        | 50-150X          |
-| Maximize          | done        | 50-150X          |
-| Bag               | done        | 1.5-2X           |
-| Bin               | done        | 5-25X (100 bins) |
-| SparselyBin       | done        | 4-5X (~100 bins) |
-| CentrallyBin      | done        | 25-40X (10 bins) |
-| IrregularlyBin    | done        | 1-4X (10 plots)  |
-| Categorize        | done        | 1.5X             |
-| Fraction          | done        | 4-20X (100 bins) |
-| Stack             | done        | 2-12X (10 plots) |
-| Select            | done        | 4-20X (100 bins) |
-| Limit             | done        | pass-through     |
-| Label             | done        | pass-through     |
-| UntypedLabel      | done        | pass-through     |
-| Index             | done        | pass-through     |
-| Branch            | done        | pass-through     |
+   * `testrootcling.py` applies all of the Numpy tests to the ROOT/Cling implementation.
+   * `testgpu.py` applies all of the Numpy tests to the CUDA GPU implementation.
+
+Primitive implementation is mature. CUDA implementation has begun.
+
+| Primitive         | Pure Python | Numpy | ROOT JIT | CUDA GPU   |
+|:------------------|:------------|:------|:---------|:-----------|
+| Count             | done        | done  | done     | done       |
+| Sum               | done        | done  | done     | done       |
+| Average           | done        | done  | done     | done       |
+| Deviate           | done        | done  | done     | done       |
+| Minimize          | done        | done  | done     | done       |
+| Maximize          | done        | done  | done     | done       |
+| Bag               | done        | done  | done     | impossible |
+| Bin               | done        | done  | done     | done       |
+| SparselyBin       | done        | done  | done     | impossible |
+| CentrallyBin      | done        | done  | done     | done       |
+| IrregularlyBin    | done        | done  | done     | done       |
+| Categorize        | done        | done  | done     | impossible |
+| Fraction          | done        | done  | done     | done       |
+| Stack             | done        | done  | done     | done       |
+| Select            | done        | done  | done     | done       |
+| Label             | done        | done  | done     | done       |
+| UntypedLabel      | done        | done  | done     | done       |
+| Index             | done        | done  | done     | done       |
+| Branch            | done        | done  | done     | done       |
+
+* "impossible" for CUDA GPU means that the primitive requires a non-constant memory allocation (hashmaps in all three cases). There must be ways of doing this by preallocating more space than is needed, but I'm not going to get into that for this round.

histogrammar/__init__.py

@@ -26,7 +26,6 @@
 from histogrammar.primitives.deviate import *
 from histogrammar.primitives.fraction import *
 from histogrammar.primitives.irregularlybin import *
-from histogrammar.primitives.limit import *
 from histogrammar.primitives.minmax import *
 from histogrammar.primitives.select import *
 from histogrammar.primitives.sparselybin import *