Merge branch 'main' into package

Author: adarshyoga

.pre-commit-config.yaml

@@ -50,7 +50,7 @@ repos:
     hooks:
     -   id: pydocstyle
         # TODO: add packages one by one to enforce pydocstyle eventually
-        files: (^dpbench/config/|^scripts/|^dpbench/console/)
+        files: (^dpbench/config/|^scripts/|^dpbench/console/|^dpbench/infrastructure/benchmark_runner.py|^dpbench/infrastructure/benchmark_validation.py)
         args: ["--convention=google"]
         # D417 does not work properly:
         # https://github.com/PyCQA/pydocstyle/issues/459

README.md

@@ -7,7 +7,7 @@ SPDX-License-Identifier: Apache-2.0
 [![Code style: black](https://img.shields.io/badge/code%20style-black-000000.svg)](https://github.com/psf/black)
 [![pre-commit](https://github.com/IntelPython/dpbench/actions/workflows/pre-commit.yml/badge.svg)](https://github.com/IntelPython/dpbench/actions/workflows/pre-commit.yml)
 
-# DPBench - Benchmarks to evaluate Data-parallel Extensions for Python
+# DPBench - Benchmarks to evaluate Data-Parallel Extensions for Python
 
 * __*_numba_*.py__ : This file contains Numba implementations of the benchmarks. There are three modes: nopython-mode, nopython-mode-parallel and nopython-mode-parallel-range.
 * __*_numba_dpex_*.py__ : This file contains Numba-Dpex implementations of the benchmarks. There are three modes: kernel-mode, numpy-mode and prange-mode.

dpbench/benchmarks/black_scholes/black_scholes_initialize.py

@@ -7,15 +7,15 @@ def initialize(nopt, seed, types_dict):
     import numpy as np
     import numpy.random as default_rng
 
-    dtype = types_dict["float"]
-    S0L = 10.0
-    S0H = 50.0
-    XL = 10.0
-    XH = 50.0
-    TL = 1.0
-    TH = 2.0
-    RISK_FREE = 0.1
-    VOLATILITY = 0.2
+    dtype: np.dtype = types_dict["float"]
+    S0L = dtype.type(10.0)
+    S0H = dtype.type(50.0)
+    XL = dtype.type(10.0)
+    XH = dtype.type(50.0)
+    TL = dtype.type(1.0)
+    TH = dtype.type(2.0)
+    RISK_FREE = dtype.type(0.1)
+    VOLATILITY = dtype.type(0.2)
 
     default_rng.seed(seed)
     price = default_rng.uniform(S0L, S0H, nopt).astype(dtype)

dpbench/benchmarks/black_scholes/black_scholes_numba_dpex_k.py

@@ -4,13 +4,15 @@
 
 from math import erf, exp, log, sqrt
 
+import dpnp as np
 import numba_dpex as dpex
 
 
 @dpex.kernel
 def _black_scholes_kernel(nopt, price, strike, t, rate, volatility, call, put):
+    dtype = price.dtype
     mr = -rate
-    sig_sig_two = volatility * volatility * 2
+    sig_sig_two = volatility * volatility * dtype.type(2)
 
     i = dpex.get_global_id(0)
 
@@ -22,14 +24,14 @@ def _black_scholes_kernel(nopt, price, strike, t, rate, volatility, call, put):
     b = T * mr
 
     z = T * sig_sig_two
-    c = 0.25 * z
-    y = 1.0 / sqrt(z)
+    c = dtype.type(0.25) * z
+    y = dtype.type(1.0) / sqrt(z)
 
     w1 = (a - b + c) * y
     w2 = (a - b - c) * y
 
-    d1 = 0.5 + 0.5 * erf(w1)
-    d2 = 0.5 + 0.5 * erf(w2)
+    d1 = dtype.type(0.5) + dtype.type(0.5) * erf(w1)
+    d2 = dtype.type(0.5) + dtype.type(0.5) * erf(w2)
 
     Se = exp(b) * S
 

dpbench/benchmarks/dbscan/dbscan_initialize.py

@@ -64,8 +64,10 @@ class Params(NamedTuple):
         data_size, Params(eps=DEFAULT_EPS, minpts=DEFAULT_MINPTS)
     )
 
+    dtype: np.dtype = types_dict["float"]
+
     return (
-        X.flatten().astype(types_dict["float"]),
-        params.eps,
+        X.flatten().astype(dtype),
+        dtype.type(params.eps),
         params.minpts,
     )

dpbench/benchmarks/dbscan/dbscan_numba_dpex_k.py

@@ -63,7 +63,7 @@ def get_neighborhood(n, dim, data, eps, ind_lst, sz_lst, block_size, nblocks):
         i2 = n if ii + 1 == nblocks1 else i1 + block_size1
         for j in range(start, stop):
             for k in range(i1, i2):
-                dist = 0.0
+                dist = data.dtype.type(0.0)
                 for m in range(dim):
                     diff = data[k * dim + m] - data[j * dim + m]
                     dist += diff * diff

dpbench/benchmarks/gpairs/gpairs_dpnp.py

@@ -12,7 +12,8 @@ def _gpairs_impl(x1, y1, z1, w1, x2, y2, z2, w2, rbins):
         + np.square(z2 - z1[:, None])
     )
     return np.array(
-        [np.outer(w1, w2)[dm <= rbins[k]].sum() for k in range(len(rbins))]
+        [np.outer(w1, w2)[dm <= rbins[k]].sum() for k in range(len(rbins))],
+        device=x1.device,
     )
 
 

dpbench/benchmarks/gpairs/gpairs_numba_dpex_k.py

@@ -2,10 +2,7 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-import math
-
 import numba_dpex as dpex
-import numpy as np
 
 # This implementation is numba dpex kernel version with atomics.
 
@@ -27,6 +24,7 @@ def count_weighted_pairs_3d_intel_no_slm_ker(
     rbins_squared,
     result,
 ):
+    dtype = x0.dtype
     lid0 = dpex.get_local_id(0)
     gr0 = dpex.get_group_id(0)
 
@@ -38,9 +36,9 @@ def count_weighted_pairs_3d_intel_no_slm_ker(
 
     n_wi = 20
 
-    dsq_mat = dpex.private.array(shape=(20 * 20), dtype=np.float32)
-    w0_vec = dpex.private.array(shape=(20), dtype=np.float32)
-    w1_vec = dpex.private.array(shape=(20), dtype=np.float32)
+    dsq_mat = dpex.private.array(shape=(20 * 20), dtype=dtype)
+    w0_vec = dpex.private.array(shape=(20), dtype=dtype)
+    w1_vec = dpex.private.array(shape=(20), dtype=dtype)
 
     offset0 = gr0 * n_wi * lws0 + lid0
     offset1 = gr1 * n_wi * lws1 + lid1
@@ -80,7 +78,7 @@ def count_weighted_pairs_3d_intel_no_slm_ker(
 
     # update slm_hist. Use work-item private buffer of 16 tfloat elements
     for k in range(0, slm_hist_size, private_hist_size):
-        private_hist = dpex.private.array(shape=(16), dtype=np.float32)
+        private_hist = dpex.private.array(shape=(16), dtype=dtype)
         for p in range(private_hist_size):
             private_hist[p] = 0.0
 
@@ -95,7 +93,9 @@ def count_weighted_pairs_3d_intel_no_slm_ker(
                 pk = k
                 for p in range(private_hist_size):
                     private_hist[p] += (
-                        pw if (pk < nbins and dsq <= rbins_squared[pk]) else 0.0
+                        pw
+                        if (pk < nbins and dsq <= rbins_squared[pk])
+                        else dtype.type(0.0)
                     )
                     pk += 1
 

dpbench/benchmarks/kmeans/kmeans_initialize.py

@@ -7,10 +7,10 @@ def initialize(npoints, niters, seed, ndims, ncentroids, types_dict):
     import numpy as np
     import numpy.random as default_rng
 
-    f_dtype = types_dict["float"]
-    i_dtype = types_dict["int"]
-    XL = 1.0
-    XH = 5.0
+    f_dtype: np.dtype = types_dict["float"]
+    i_dtype: np.dtype = types_dict["int"]
+    XL = f_dtype.type(1.0)
+    XH = f_dtype.type(5.0)
 
     default_rng.seed(seed)
 

dpbench/benchmarks/kmeans/kmeans_numba_dpex_k.py

@@ -11,7 +11,8 @@
 def groupByCluster(arrayP, arrayPcluster, arrayC, num_points, num_centroids):
     idx = dpex.get_global_id(0)
     # if idx < num_points: # why it was removed??
-    minor_distance = -1
+    dtype = arrayC.dtype
+    minor_distance = dtype.type(-1)
     for i in range(num_centroids):
         dx = arrayP[idx, 0] - arrayC[i, 0]
         dy = arrayP[idx, 1] - arrayC[i, 1]
@@ -41,8 +42,9 @@ def calCentroidsSum2(arrayP, arrayPcluster, arrayCsum, arrayCnumpoint):
 @dpex.kernel
 def updateCentroids(arrayC, arrayCsum, arrayCnumpoint, num_centroids):
     i = dpex.get_global_id(0)
-    arrayC[i, 0] = arrayCsum[i, 0] / arrayCnumpoint[i]
-    arrayC[i, 1] = arrayCsum[i, 1] / arrayCnumpoint[i]
+    dtype = arrayC.dtype
+    arrayC[i, 0] = arrayCsum[i, 0] / dtype.type(arrayCnumpoint[i])
+    arrayC[i, 1] = arrayCsum[i, 1] / dtype.type(arrayCnumpoint[i])
 
 
 @dpex.kernel