scale function(_get_mean_var) updated for dense array, speedup upto ~4.65x

[ashish615]

Hi,
We are submitting PR for speed up of the _get_mean_var function.

	Time(sec)
Original	18.49
Updated	3.97
Speedup	4.65743073

experiment setup : AWS r7i.24xlarge

import time
import numpy as np

import pandas as pd

import scanpy as sc
from sklearn.cluster import KMeans

import os
import wget

import warnings



warnings.filterwarnings('ignore', 'Expected ')
warnings.simplefilter('ignore')
input_file = "./1M_brain_cells_10X.sparse.h5ad"

if not os.path.exists(input_file):
    print('Downloading import file...')
    wget.download('https://rapids-single-cell-examples.s3.us-east-2.amazonaws.com/1M_brain_cells_10X.sparse.h5ad',input_file)


# marker genes
MITO_GENE_PREFIX = "mt-" # Prefix for mitochondrial genes to regress out
markers = ["Stmn2", "Hes1", "Olig1"] # Marker genes for visualization

# filtering cells
min_genes_per_cell = 200 # Filter out cells with fewer genes than this expressed
max_genes_per_cell = 6000 # Filter out cells with more genes than this expressed

# filtering genes
min_cells_per_gene = 1 # Filter out genes expressed in fewer cells than this
n_top_genes = 4000 # Number of highly variable genes to retain

# PCA
n_components = 50 # Number of principal components to compute

# t-SNE
tsne_n_pcs = 20 # Number of principal components to use for t-SNE

# k-means
k = 35 # Number of clusters for k-means

# Gene ranking

ranking_n_top_genes = 50 # Number of differential genes to compute for each cluster

# Number of parallel jobs
sc._settings.ScanpyConfig.n_jobs = os.cpu_count()

start=time.time()
tr=time.time()
adata = sc.read(input_file)
adata.var_names_make_unique()
adata.shape
print("Total read time : %s" % (time.time()-tr))



tr=time.time()
# To reduce the number of cells:
USE_FIRST_N_CELLS = 1300000
adata = adata[0:USE_FIRST_N_CELLS]
adata.shape

sc.pp.filter_cells(adata, min_genes=min_genes_per_cell)
sc.pp.filter_cells(adata, max_genes=max_genes_per_cell)
sc.pp.filter_genes(adata, min_cells=min_cells_per_gene)
sc.pp.normalize_total(adata, target_sum=1e4)
print("Total filter and normalize time : %s" % (time.time()-tr))


tr=time.time()
sc.pp.log1p(adata)
print("Total log time : %s" % (time.time()-tr))


# Select highly variable genes
sc.pp.highly_variable_genes(adata, n_top_genes=n_top_genes, flavor = "cell_ranger")

# Retain marker gene expression
for marker in markers:
        adata.obs[marker + "_raw"] = adata.X[:, adata.var.index == marker].toarray().ravel()

# Filter matrix to only variable genes
adata = adata[:, adata.var.highly_variable]

ts=time.time()
#Regress out confounding factors (number of counts, mitochondrial gene expression)
mito_genes = adata.var_names.str.startswith(MITO_GENE_PREFIX)
n_counts = np.array(adata.X.sum(axis=1))
adata.obs['percent_mito'] = np.array(np.sum(adata[:, mito_genes].X, axis=1)) / n_counts
adata.obs['n_counts'] = n_counts


sc.pp.regress_out(adata, ['n_counts', 'percent_mito'])
print("Total regress out time : %s" % (time.time()-ts))

#scale

ts=time.time()
sc.pp.scale(adata)
print("Total scale time : %s" % (time.time()-ts))

add timer around _get_mean_var call

scanpy/scanpy/preprocessing/_scale.py

Line 167 in 706d4ef

mean, var = _get_mean_var(X)

we can also create _get_mean_var_std function that return std as well so we don't require to compute it in scale function(L168-L169).

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.31%. Comparing base (896e249) to head (7a1a62e).
Report is 127 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3099      +/-   ##
==========================================
- Coverage   76.31%   76.31%   -0.01%     
==========================================
  Files         109      109              
  Lines       12513    12516       +3     
==========================================
+ Hits         9549     9551       +2     
- Misses       2964     2965       +1     

Files with missing lines	Coverage Δ
src/scanpy/preprocessing/_utils.py	95.12% <100.00%> (-2.25%)	⬇️

[scverse-benchmark]

Benchmark changes

Change	Before [ad657ed]	After [e7a4662]	Ratio	Benchmark (Parameter)
+	259M	310M	1.2	preprocessing_log.FastSuite.peakmem_mean_var('pbmc68k_reduced')
+	1.16±0.04ms	1.97±0.5ms	1.69	preprocessing_log.FastSuite.time_mean_var('pbmc68k_reduced')
+	255M	315M	1.23	preprocessing_log.peakmem_highly_variable_genes('pbmc68k_reduced')
-	373M	322M	0.86	preprocessing_log.peakmem_pca('pbmc68k_reduced')
-	1.03G	779M	0.76	preprocessing_log.peakmem_scale('pbmc3k')
-	729±5ms	517±5ms	0.71	preprocessing_log.time_scale('pbmc3k')

Comparison: https://github.com/scverse/scanpy/compare/ad657edfb52e9957b9a93b3a16fc8a87852f3f09..e7a466265b08f6973a5cf3fecfc27879104c02f4
Last changed: Tue, 18 Jun 2024 18:39:49 +0000

More details: https://github.com/scverse/scanpy/pull/3099/checks?check_run_id=26384736173

[Intron7]

I have some small improvements that I would like to add next week for more precision for larger matrices

[Intron7]

@ashish615 after doing some benchmarking myself I found out that your solution for axis=1 is under performing compared to axis=0 for larger arrays. I think that is because of the memory access pattern you choose. I rewrote the function with that in mind. I'll again make a PR to you, because for some reason you disallow us from making changes to your PR.

[Intron7]

Please merge IntelLabs/open-omics-scanpy#2

[Intron7]

This fixes the issues

[Intron7]

Looks good to me

[flying-sheep]

I don’t see the claimed speedup in the benchmarks, what’s missing?

Also is numba.get_num_threads() safe? E.g. I think _get_mean_var is also called in each dask chunks. Will numba.get_num_threads() return a reasonable number in that case?

Otherwise nice! I’m not a huge fan of how unpythonic numba code looks, but I don’t think anything can be done about that.

[flying-sheep]

Before your change, this line ran unconditionally, now it only runs for the not isinstance(X, np.ndarray) case. Is that intentional? Then you should mention that in _compute_mean_var’s docstring.

[flying-sheep]

We already have _get_mean_var. Maybe rename this to _get_mean_var_ndarray or _get_mean_var_dense?

[Intron7]

I think we can rename the kernel

[flying-sheep]

Suggested change

	X: _SupportedArray, axis: Literal[0, 1] = 0, n_threads=1
	X: _SupportedArray, axis: Literal[0, 1] = 0, n_threads: int = 1

[flying-sheep]

Please don’t duplicate identical lines.

Suggested change

	if axis == 0:
	axis_i = 1
	sums = np.zeros((n_threads, X.shape[axis_i]), dtype=np.float64)
	sums_squared = np.zeros((n_threads, X.shape[axis_i]), dtype=np.float64)
	mean = np.zeros(X.shape[axis_i], dtype=np.float64)
	var = np.zeros(X.shape[axis_i], dtype=np.float64)
	n = X.shape[axis]
	for i in numba.prange(n_threads):
	for r in range(i, n, n_threads):
	for c in range(X.shape[axis_i]):
	value = X[r, c]
	sums[i, c] += value
	sums_squared[i, c] += value * value
	for c in numba.prange(X.shape[axis_i]):
	sum_ = sums[:, c].sum()
	mean[c] = sum_ / n
	var[c] = (sums_squared[:, c].sum() - sum_ * sum_ / n) / (n - 1)
	else:
	axis_i = 0
	mean = np.zeros(X.shape[axis_i], dtype=np.float64)
	var = np.zeros(X.shape[axis_i], dtype=np.float64)
	for r in numba.prange(X.shape[0]):
	for c in range(X.shape[1]):
	value = X[r, c]
	mean[r] += value
	var[r] += value * value
	for c in numba.prange(X.shape[0]):
	mean[c] = mean[c] / X.shape[1]
	var[c] = (var[c] - mean[c] ** 2) / (X.shape[1] - 1)
	axis_i = 1 - axis
	mean = np.zeros(X.shape[axis_i], dtype=np.float64)
	var = np.zeros(X.shape[axis_i], dtype=np.float64)
	if axis == 0:
	sums = np.zeros((n_threads, X.shape[axis_i]), dtype=np.float64)
	sums_squared = np.zeros((n_threads, X.shape[axis_i]), dtype=np.float64)
	n = X.shape[axis]
	for i in numba.prange(n_threads):
	for r in range(i, n, n_threads):
	for c in range(X.shape[axis_i]):
	value = X[r, c]
	sums[i, c] += value
	sums_squared[i, c] += value * value
	for c in numba.prange(X.shape[axis_i]):
	sum_ = sums[:, c].sum()
	mean[c] = sum_ / n
	var[c] = (sums_squared[:, c].sum() - sum_ * sum_ / n) / (n - 1)
	else:
	for r in numba.prange(X.shape[0]):
	for c in range(X.shape[1]):
	value = X[r, c]
	mean[r] += value
	var[r] += value * value
	for c in numba.prange(X.shape[0]):
	mean[c] = mean[c] / X.shape[1]
	var[c] = (var[c] - mean[c] ** 2) / (X.shape[1] - 1)

[Intron7]

I think we can slim this down a bit. The two different loops need to be separate though

[Intron7]

The function should also work for 1 thread. numba.get_num_threads() is fine it works well with the sparse arrays. But I have no experience with it inside of dask.

[ilan-gold]

I don't think _SupportedArray is the right type annotation here. This doesn't run directly on dask.Array, unless I am misunderstanding something.