weblas/index.js

GPU accelerated BLAS for node and the browser

var globals = require("./lib/globals"),
	pipeline = require("./lib/pipeline"),
	SGEMMCalculator = require("./lib/sgemmcalculator"),
	SAXPYCalculator = require("./lib/saxpycalculator"),
	SSCALCalculator = require("./lib/sscalcalculator"),
	SDWNSCalculator = require("./lib/sdwnscalculator"),
	SCLMPCalculator = require("./lib/sclmpcalculator");

// do we have a WebGL context?
if(globals.gl){
	// yes, load the library
	module.exports = createModule(globals.gl);
} else {
	// no, abort and export null
	module.exports = null;
}

function createModule(gl){

	var sgemmcalculator = new SGEMMCalculator(gl),
		saxpycalculator = new SAXPYCalculator(gl),
		sscalcalculator = new SSCALCalculator(gl),
		sdwnscalculator = new SDWNSCalculator(gl),
		sclmpcalculator = new SCLMPCalculator(gl);

	return {
		// level one
		"saxpy" : saxpy,
		"sscal" : sscal,   // single precision matrix scale
		// level two
		// level three
		"sgemm" : sgemm,   // single precision generalized matrix multiply
		// extra
		"sstd" : sstd,     // single precision Standard Score normalization
		"sdwns": sdwns,
		"sclmp": sclmp,
		// pipeline
		"pipeline" : pipeline,
		// internals
		"gpu" : {
					"gl" : gl,
		 			"sgemm": pipeline.sgemmcalculator.calculate.bind(pipeline.sgemmcalculator),
					"sscal" : pipeline.sscalcalculator.calculate.bind(pipeline.sscalcalculator),
					"sclmp" : pipeline.sclmpcalculator.calculate.bind(pipeline.sclmpcalculator),
					"sdwns" : pipeline.sdwnscalculator.calculate.bind(pipeline.sdwnscalculator),
					"encode" : gl.encode.bind(gl)
				},
		"util" : { "fromArray" : fromArray, "transpose" : transpose}
	};


	/* Wrap the GL calculation object in a (relatively) user friendly function that
		accepts TypedArrays

		* convert the data to (padded) textures in GPU memory
		* execute calculation
		* read result into an array, and return
	 */
	function sgemm(M, N, K, alpha, A, B, beta, C){

		if(C != null && C.length != N){
			throw new Error("Only vector C with length matching rows in A is currently supported.");
		}

		// pack each matrix into a single RGBA texel array, with the second transposed
		var texels0 = A,
			texels1,
			texels2 = C;


		texels1 = transpose(K, N, B);

		// create input textures from data
		var texture0 = gl.createDataTexture(M, K, texels0);
		var texture1 = gl.createDataTexture(N, K, texels1);
		var texture2 = null;
		if(texels2 != null){
			texture2 = gl.createDataTexture(1, N, texels2);
		}

		var texture3 = gl.createOutputTexture(M, N);

		sgemmcalculator.calculate(M, N, K, alpha, texture0, texture1, beta, texture2, texture3);

		// retrieve data
		rawBuffer = gl.readData(M, N);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture1);
		if(texture2 != null){
			gl.context.deleteTexture(texture2);
		}
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);

	}

	function saxpy(N, a, X, Y){

		var rawBuffer;


		var texels0 = X,
			texels1;

		// TODO: special shader for constant Y
		if(isFloat32Array(Y)){
			texels1 = Y;
		} else {
			texels1 = new Float32Array(N);
			texels1.fill(Y);
		}

		// create input textures from data
		var texture0 = gl.createDataTexture(1, N, texels0);
		var texture1 = gl.createDataTexture(1, N, texels1);

		var texture3 = gl.createOutputTexture(1, N);

		saxpycalculator.calculate(N, a, texture0, texture1, texture3);

		// retrieve data
		rawBuffer = gl.readData(1, N);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture1);
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);

	}

	function isFloat32Array(obj){
		return Object.prototype.toString.call(obj) === "[object Float32Array]";
	}
	/* a more general version of the BLAS Level 1 scale, that works on matrices
	   and includes an elementwise scalar addition

	   a * X + b

	   a - multiplicative scalar
	   b - additive scalar
	   X - matrix (M x N)

	   to get the standard BLAS scal set M = 1 and b = 0

	   this function is generally only cost effective to use in a pipeline
	*/
	function sscal(M, N, a, b, X){

		var rawBuffer;

		var texels0 = X;
		var texture0 = gl.createDataTexture(M, N, texels0);

		var texture3 = gl.createOutputTexture(M, N);

		sscalcalculator.calculate(M, N, a, b, texture0, texture3);

		// retrieve data
		rawBuffer = gl.readData(M, N);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);
	}

	/* Calculate the Standard Score normalization (subtract mean
	   ,divide by standard deviation).
	 */
	function sstd(M, N, mu, sigma, X){

		var rawBuffer;

		var texels0 = X;
		var texture0 = gl.createDataTexture(M, N, texels0);

		var texture3 = gl.createOutputTexture(M, N);

		// adjust the parameters (for inverse) and call the standard score normalization
		sscalcalculator.calculate(M, N, 1.0/sigma, -1.0 * mu/sigma, texture0, texture3);

		// retrieve data
		rawBuffer = gl.readData(M, N);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);
	}

	/* downsample an image (taking the max) for Pooling

		M - rows in input
		N - columns in input
		c - channels in input
		factor - the downsample factor (width of patch to sample)
		stride - width between pooling regions
		X - input image
	 */
	function sdwns(M, N, channels, factor, stride, X){


		var texels0 = X;

		var texture0 = gl.createDataTexture(M, N * channels, X);

		var N_out = Math.floor((N - factor) / stride) + 1;
		var M_out = Math.floor((M - factor) / stride) + 1;

		var texture3 = gl.createOutputTexture(M_out, N_out * channels);

		sdwnscalculator.calculate(M, N, channels, factor, stride, texture0, texture3);

		// retrieve data
		rawBuffer = gl.readData(M_out, N_out * channels);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);
	}
	/*  Elementwise clamp function for matrices on the interval [a, b]. Can also be
		used for min or max, by passing Number.MIN_VALUE for the first parameter and
		Number.MAX_VALUE for the second parameter, respectively.

		Passing `null` for either of these parameters will default to it's
		respective min or max value.

		M - number of rows in X
		N - number of columns in X
		a - lower bound (inclusize)
		b - upper bound (inclusive)
		X - matrix

	   to get the standard BLAS scal set M = 1 and b = 0

	   this function is generally only cost effective to use in a pipeline
	*/
	function sclmp(M, N, a, b, X){

		a = (a != null) ? a : Number.MIN_VALUE;
		b = (b != null) ? b : Number.MAX_VALUE;

		var rawBuffer;

		var texels0 = X;
		var texture0 = gl.createDataTexture(M, N, texels0);

		var texture3 = gl.createOutputTexture(M, N);

		sclmpcalculator.calculate(M, N, a, b, texture0, texture3);

		// retrieve data
		rawBuffer = gl.readData(M, N);

		// clean up
		gl.context.deleteTexture(texture0);
		gl.context.deleteTexture(texture3);

		// return result
		return new Float32Array(rawBuffer);
	}
	/*
	function saxpy(n, a, x, y){
		var i = 0,
			result = new Float32Array(n);

		// assert n = x.length
		// assert a is scalar
		// assert x is Float32Array

		if(isNumeric(y)){
			// shortcut for scalar y
			for(; i < n; i++){
				result[i] = a * x[i] + y;
			}
		} else {

			for(; i < n; i++){
				result[i] = a * x[i] + y[i];
			}
		}

		return result;

	}*/

	// add a String.format method, if none exists
	if (!String.prototype.format) {
	  String.prototype.format = function() {
		var args = arguments;
		return this.replace(/{(\d+)}/g, function(match, number) {
		  return typeof args[number] != 'undefined'
			? args[number]
			: match
		  ;
		});
	  };
	}

	function isNumeric( obj ) { return (obj - parseFloat( obj ) + 1) >= 0; }

	/* create a typed array from a 2D javascript array */
	function fromArray(array, type, tranpose) {
		var shape = [],
				data,
				c;   // number of columns

		if(!tranpose){
			shape[0] = array.length;
			shape[1] = array[0].length;
		} else {
			shape[1] = array.length;
			shape[0] = array[0].length;
		}
		c = shape[1];

		type = type || Float32Array;

		data = new type(shape[0]*shape[1]);

		for (var ii = 0; ii < shape[0]; ++ii)
			for (var jj = 0; jj < shape[1]; ++jj)
			if(!tranpose)
				data[ii*c + jj] = array[ii][jj];
			else
				data[ii*c + jj] = array[jj][ii];

		return data;
	};

	// tranpose a typed array in row major order, with the given row and column
	// numers
	function transpose(r, c, typedArray){
		var result = new typedArray.constructor(r*c);

		for(var i = 0; i < r; i++){
			for(var j = 0; j < c; j++){
				result[j * r + i] = typedArray[i * c + j];
			}
		}

		return result;
	}

}