weblas/test/tensor.reshape.js

GPU accelerated BLAS for node and the browser

var tape = require('tape'),
	weblas = require('../index'),
	loader = require('floader'); // browserify aware file loader (xhr in browser)

weblas.test = require('../lib/test');

var RTOL = 1e-05,
	ATOL = 1e-07;

// reusing data from sscal
var dataDirectory = 'test/data/sscal/',
	testFile = 'small.json';

var gl = weblas.gpu.gl;

var matrixFiles = ['a.arr'];

tape("Tensor.reshape: 4 x 8", function(t){
	t.plan(1);

	var x = new Float32Array([ 1.0,  2.0,  3.0,  4.0,  5.0,  6.0,  7.0,  8.0,
							   9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0,
							  17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0,
							  25.0, 26.0, 27.0, 28.0, 29.0, 30.0, 31.0, 32.0]),
		expected = new Float32Array([ 1.0,  2.0,  3.0,  4.0,
			  						  5.0,  6.0,  7.0,  8.0,
									  9.0, 10.0, 11.0, 12.0,
									 13.0, 14.0, 15.0, 16.0,
									 17.0, 18.0, 19.0, 20.0,
									 21.0, 22.0, 23.0, 24.0,
									 25.0, 26.0, 27.0, 28.0,
									 29.0, 30.0, 31.0, 32.0]);

	var M = 4, N = 8,
		M2 = 8, N2 = 4,
		t0 = new weblas.pipeline.Tensor([M, N], x),
		t1;

	try{
		// when tranposing texture for t0 is deleted by default
		t1 = t0.reshape([M2, N2]);
		// when transfering texture for t1 is deleted by default
		var result = t1.transfer();
	}
	catch(ex){
		t.error(ex);
		return;
	}

	weblas.test.assert.allclose(t, result, expected, null, RTOL, ATOL);
});

tape("Tensor.reshape: 3 x 4", function(t){
	t.plan(2);

	var x = new Float32Array([ 1.0,  2.0,  3.0,  4.0,
							   5.0,  6.0,  7.0,  8.0,
							   9.0, 10.0, 11.0, 12.0]),
		expected = new Float32Array([ 1.0,  2.0,  3.0,
									  4.0,  5.0,  6.0,
									  7.0,  8.0,  9.0,
									 10.0, 11.0, 12.0]);

	var M = 3, N = 4,
		M2 = 4, N2 = 3,
		t0 = new weblas.pipeline.Tensor([M, N], x),
		t1;

	try{
		// when tranposing texture for t0 is deleted by default
		t1 = t0.reshape([M2, N2]);
		// when transfering texture for t1 is deleted by default
		var result = t1.transfer(true);
	}
	catch(ex){
		t.error(ex);
		return;
	}

	weblas.test.assert.allclose(t, result, expected, null, RTOL, ATOL);

	// use internals to check that texture is padded correctly
	var pad = 1,
		padded;

	try{
		padded = weblas.test.padData(M2, N2, pad, expected);
		out = weblas.gpu.gl.createOutputTexture(M2, N2 + pad);

		// float extraction
		weblas.gpu.encode(M2, N2 + pad, t1.texture, out);
		result = new Float32Array(weblas.gpu.gl.readData(M2, N2 + pad));

		weblas.gpu.gl.context.deleteTexture(out);
	}
	catch(ex){
		t.error(ex);
		return;
	}

	weblas.test.assert.allclose(t, result, padded, null, RTOL, ATOL);
	t1.delete();
});

function generateReshapeTestCase(prefix, M, N){
	return function(t){
		var pad = weblas.gpu.gl.getPad(M);
		if(pad == 0){
			t.plan(1);
		} else {
			t.plan(2);
		}

		var X, expected; // typed arrays

		// directory containing matrix data files for current test
		var testDirectory = dataDirectory + prefix + '/';

		// load matrices from files
		weblas.test.load(testDirectory, matrixFiles, function(err, matrices){

			if(err){
				t.error(err);
				if(pad != 0) t.notOk(false, "skipping padding test");

				return;
			}

			// matrices is an array which matches matrixFiles
			var X = matrices[0];

			if(!(X && X.length && X.length == M * N)){

				throw new Error("malformed data");
			}

			expected = X;

			var t0 = new weblas.pipeline.Tensor([M, N], X),
				t1;

			try{
				// when tranposing texture for t0 is deleted by default
				t1 = t0.reshape([N, M]);
			}
			catch(ex){
				t.error(ex);
				if(pad != 0) t.notOk(false, "skipping padding test");

				return;
			}

			// when transfering texture for t1 is deleted by default
			// testing with pad requires texture stay resident in GPU memory
			var result = t1.transfer(true);

			testWithPad(t, N, M, pad, result, expected, t1.texture, RTOL, ATOL);
			t1.delete();

		});
	};
}

loader.load(dataDirectory + testFile, function(err, config){

	var suite = JSON.parse(config);

	// suite configuration file uses directory name as key
	for(var i = 0; i < suite.length; i++){

		directory = String("0000" + (i + 1)).slice(-4);

		var test = suite[i];

		var input = test['in'],
			sizes = input['shape'],
			arg = test['arg'] || {};

		var m = input[0]['shape'][0],
			n = input[0]['shape'][1];

		//console.log("a: " + a + "; b: " + b);
		var testName = "Tensor.reshape: " + m + "x" + n;
		tape(testName, generateReshapeTestCase(directory, m, n));
	}

});


function testWithPad(t,  M, N, pad, result, expected, texture, RTOL, ATOL){

	weblas.test.assert.allclose(t, result, expected, null, RTOL, ATOL);

	if(pad > 0){

		// use internals to check that texture is padded correctly
		var padded;

		try{
			padded = weblas.test.padData(M, N, pad, expected);
			out = weblas.gpu.gl.createOutputTexture(M, N + pad);

			// float extraction
			weblas.gpu.encode(M, N + pad, texture, out);
			result = new Float32Array(weblas.gpu.gl.readData(M, N + pad));

			weblas.gpu.gl.context.deleteTexture(out);
		}
		catch(ex){
			t.assert(false, ex);
		}

		weblas.test.assert.allclose(t, result, padded, null, RTOL, ATOL);
	}
}