Navigation

← Previous Changeset
Next Changeset →

Changeset 171

Timestamp:

01/08/08 07:11:22 (8 months ago)

Author:

Don Clugston

Message:

Added abs() intrinsic. Added codegen for unary minus. Bugfix for rank error messages. Bugfix for A=A.

Files:

trunk/blade/Blade.d (modified) (2 diffs)
trunk/blade/BladeDemo.d (modified) (2 diffs)
trunk/blade/BladeRank.d (modified) (3 diffs)
trunk/blade/BladeSimplify.d (modified) (6 diffs)
trunk/blade/CodegenX86.d (modified) (7 diffs)
trunk/blade/PostfixX86.d (modified) (4 diffs)
trunk/blade/SyntaxTree.d (modified) (2 diffs)

Legend:

: Unmodified
: Added
: Removed
: Modified
: Copied
: Moved

trunk/blade/Blade.d

r170	r171
13	13	*
14	14	* FEATURES:
15		* - Supports any mix of vector addition, subtraction, dot product, ~~slicing, and~~
16		* multiplication by a scalar.
	15	* - Supports any mix of vector addition, subtraction, dot product, unary minus,
	16	* multiplication by a scalar, sum(), abs(), and multidimensional slicing.
17	17	* - Generates either x87 asm code, SSE or SSE2 asm code or pure D, depending on
18	18	* the complexity of the expression, and the availability of inline asm.
…	…
143	143	}
144	144
	145	private:
	146	// Masks for setting or clearing the signbit in SSE registers.
	147	static ulong[2] SSE_SIGNMASKpd = [0x7FFF_FFFF_FFFF_FFFFL, 0x7FFF_FFFF_FFFF_FFFFL];
	148	static uint[4] SSE_SIGNMASKps = [0x7FFF_FFFF, 0x7FFF_FFFF, 0x7FFF_FFFF, 0x7FFF_FFFF];
	149	static ulong[2] SSE_SIGNBITpd = [0x8000_0000_0000_0000L, 0x8000_0000_0000_0000L];
	150	static uint[4] SSE_SIGNBITps = [0x8000_0000,0x8000_0000,0x8000_0000, 0x8000_0000];
145	151
146	152	private:

trunk/blade/BladeDemo.d

r169	r171
42	42	mixin(vectorize("a+=6*another[1,0..$]"));
43	43	mixin(vectorize("a+=6*(another[1,0..$]+another[1,0..$])"));
	44
44	45
45	46	mixin(vectorize("a+=6*another[1][0..$]"));
…	…
52	53	mixin(vectorize("another[0..$,1]=6*a[0..2]"));
53	54
54		// Simplifies to q= 2dot(q,q)dot(qq).
	55	// Simplifies to q= 2dot(q,q)*dot(q,q).
55	56	mixin(vectorize("q =dot(q,qdot(2*q,q))"));
56	57	double u;
57	58	mixin(vectorize("u = dot(q,q*dot(q,q))"));
58	59	mixin(vectorize("u = dot(a, q)"));
59		mixin(vectorize("u = sum(a)"));
	60	mixin(vectorize("u = sum(abs(-p))"));
	61	mixin(vectorize("a = -a"));
60	62
61	63	writefln("a=", a);

trunk/blade/BladeRank.d

r169	r171
93	93	RankMismatchDotProduct = -10,
94	94	ExtraCharsAfterArrayLiteral = -11,
95		ArrayLiteralRankMismatch = -12
	95	ArrayLiteralRankMismatch = -12,
	96	AbsDimensionality = -13
96	97	}
97	98
…	…
107	108	"Dimensionality mismatch (addition or subtraction)",
108	109	"Dimensionality mismatch in concatenation",
109		"Dimenionality error in dot product"
110		"Extra characters after array literal"
111		"Rank mismatch in array literal"
	110	"Dimenionality error in dot product",
	111	"Extra characters after array literal",
	112	"Rank mismatch in array literal",
	113	"Can only use abs with scalar or vector"
112	114	][-err-1];
113	115	}
…	…
135	137	if (lrank<0) return lrank; // propagate errors
136	138	return 0;
137		}
	139	} else if (func=="abs") {
	140	auto lrank = doVisit(this_,args[0]);
	141	if (lrank>1) return RankError.AbsDimensionality;
	142	return 0;
	143	}
138	144	assert(0, "BLADE ICE: Unsupported function:" ~ func);
139	145	return 0;

trunk/blade/BladeSimplify.d

r169	r171
51	51	bool isBladeIntrinsic(char [] str)
52	52	{
53		return str=="dot" \|\| str=="sum";
	53	return str=="dot" \|\| str=="sum" \|\| str=="abs";
54	54	}
55	55
…	…
90	90	int numdups=0;
91	91	char [] mapping = ""; // The new letter which this symbol should become,
92		// or '!' if it is an intrinsic
	92	// or '!' if it is an intrinsic
93	93	for (int i=0; i<tree.symbolTable.length; ++i) {
94	94	char c = 'A'+i;
…	…
112	112	char c = tree.expression[i];
113	113	if (c>='A' && c<='Z') {
114		if (mapping[c-'A']=='!') e~=tree.symbolTable[c-'A'].value;
115		else e~= mapping[c-'A'];
	114	if (mapping[c-'A']=='!') e~=tree.symbolTable[c-'A'].value;
	115	else e~= mapping[c-'A'];
116	116	} else e~=c;
117	117	}
…	…
281	281	}
282	282	}
283		ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
284		if (func=="dot") { // dot product.
285		// Each element is reduced seperately
286		char [] left = wrapInParens(doVisit(this_,args[0]));
287		char [] right = wrapInParens(doVisit(this_, args[1]));
288		return func ~ "(" ~ left ~ "," ~ right ~ ")";
289		} else if (func=="sum") {
290		char [] left = wrapInParens(doVisit(this_,args[0]));
291		return func ~ "(" ~ left ~ ")";
292		}
293		assert(0, "BLADE ICE: Unsupported function");
294		return "";
	283	ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
	284	switch (func) {
	285	case "dot":
	286	// Each element is reduced seperately
	287	char [] left = wrapInParens(doVisit(this_,args[0]));
	288	char [] right = wrapInParens(doVisit(this_, args[1]));
	289	return func ~ "(" ~ left ~ "," ~ right ~ ")";
	290	case "sum":
	291	case "abs":
	292	char [] left = wrapInParens(doVisit(this_,args[0]));
	293	return func ~ "(" ~ left ~ ")";
	294	default:
	295	assert(0, "BLADE ICE: Unsupported function");
	296	return "";
	297	}
295	298	}
296	299	ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {
297	300	assert(this_.slicing.length==0, "BLADE ICE");
298		return op ~ ~~doVisit(this_, expr);~~
	301	return op ~ wrapInParens(doVisit(this_, expr));
299	302	}
300	303	ReturnType onVisitPostfix(This this_, char [] op, char [] expr) {
301	304	assert(this_.slicing.length==0, "BLADE ICE");
302		return ~~doVisit(this_, expr~~) ~ op;
	305	return wrapInParens(doVisit(this_, expr)) ~ op;
303	306	}
304	307	// Includes multi-dimensional slicing and indexing.
…	…
459	462	else return ScalarFold(sym, "");
460	463	}
461		ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
462		if (func=="dot") { // dot product.
	464	ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
	465	switch(func) {
	466	case "dot":
463	467	ScalarFold left = doVisit(this_,args[0]);
464	468	ScalarFold right = doVisit(this_, args[1]);
465	469	return ScalarFold("", combineMul(combineMul(left.multiplier, right.multiplier), "{" ~ func ~ "(" ~ wrapInParens(left.expr) ~ "," ~ wrapInParens(right.expr) ~ ")}"));
466		~~} else if (func=="sum") {~~
	470	case "sum":
467	471	ScalarFold left = doVisit(this_,args[0]);
468	472	return ScalarFold("", combineMul(left.multiplier, "{" ~ func ~ "(" ~ wrapInParens(left.expr) ~ ")}"));
469		} else {
	473	case "abs":
	474	ScalarFold left = doVisit(this_,args[0]);
	475	return ScalarFold((left.expr!="" ? func ~ "(" ~ wrapInParens(left.expr) ~ ")" : ""), (left.multiplier.length>0 && left.multiplier!="-")? func ~ "(" ~ wrapInParens(left.multiplier) ~ ")":"");
	476	default:
470	477	assert(0, "BLADE: Unsupported function");
471	478	return ScalarFold("","");
…	…
483	490	} else {
484	491	ScalarFold f = doVisit(this_, expr);
485		assert(f.expr=="");
	492	assert(f.expr=="", "BLADE ICE: Prefix error" ~ op ~ " " ~ expr);
486	493	return ScalarFold("", op ~ wrapInParens(f.multiplier));
487	494	}

trunk/blade/CodegenX86.d

r170	r171
38	38	* This string is converted to postfix. The postfix string is converted to
39	39	* a string containing x87 asm, which is then mixed into a function which accepts the tuple.
40		*
41	40	*/
42	41
…	…
44	43	* POTENTIAL FROM RECENT INSTRUCTION SETS:
45	44	* SSE5(AMD): fmaddpd can dramatically improve both performance and accuracy.
46		* SSE4(Intel): dppd has limited use.
	45	* SSE4(Intel): dppd has some limited use.
47	46	*/
48	47
…	…
399	398	++done;
400	399	numOnStack--;
	400	} else if (operations[done]=='a') {
	401	mainbody ~= " fabs;"\n;
	402	++done;
	403	} else if (operations[done]=='n') {
	404	mainbody ~= " fchs;"\n;
	405	++done;
401	406	} else if (!isInstruction(operations[done+1])){
402	407	// load a vector onto the FPU stack, to begin a new subexpression.
…	…
498	503
499	504	private:
	505
500	506	// split off from the template to make code coverage work
501	507	char [] generateCodeForSSEImpl(bool usingDoubles, char [] ranklist, char [] operations, char cumulatingOp=0)
…	…
564	570	++done;
565	571	numOnStack--;
	572	} else if (operations[done]=='a') { // abs
	573	mainbody ~= " andp" ~ suffix ~ XMM(numOnStack-1) ~ ", SSE_SIGNMASKp" ~ suffix ~"; // abs"\n;
	574	extra ~= " andp" ~ suffix ~ XMM(numOnStack-1) ~ ", SSE_SIGNMASKp" ~ suffix ~ "; // abs"\n;
	575	++done;
	576	} else if (operations[done]=='n') { // neg
	577	mainbody ~= " xorp" ~ suffix ~ XMM(numOnStack-1) ~ ", SSE_SIGNBITp" ~ suffix ~"; // negate"\n;
	578	extra ~= " xorp" ~ suffix ~ XMM(numOnStack-1) ~ ", SSE_SIGNBITp" ~ suffix ~ "; // negate"\n;
	579	++done;
566	580	} else if (!isInstruction(operations[done+1])){
567	581	// load a vector onto the FPU stack, to begin a new subexpression.
…	…
571	585	++done;
572	586	numOnStack++;
	587	} else if (operations[done]==',') {
	588	// operation on self, eg XX+ --> don't need to load it again.
	589	comment = "; // " ~ operations[done..done+2] ~ \n;
	590	mainbody ~= " " ~ opToSSE[operations[done+1]] ~ suffix ~ " " ~ XMM(numOnStack-1) ~ ", "
	591	~ XMM(numOnStack-1) ~ comment;
	592	extra ~= " " ~ opToSSESingle[operations[done+1]] ~ suffix ~ " " ~ XMM(numOnStack-1) ~ ", "
	593	~ XMM(numOnStack-1) ~ comment;
	594	done +=2;
573	595	} else if (ranklist[operations[done]-'A']=='1') {
574	596	// An operation will be performed between the stack top and a vector.
…	…
577	599	mainbody ~= " movap" ~ suffix ~ indexedSSEVector(ranklist, operations[$-2], vectorsize) ~ ", XMM" ~ itoa(numOnStack-1) ~ comment;
578	600	extra ~= " movs" ~ suffix ~ indexedSSENext(ranklist, operations[$-2], vectorsize) ~ ", XMM" ~ itoa(numOnStack-1) ~ comment;
579		~~} else~~
580		~~if (operations[done-1]==operations[done]) {~~
581		~~// operation on self, eg XX+ --> don't need to load it again.~~
582		~~int cumvector = (operations[done-1]=='0')? numScalarsOnStack : numOnStack-1;~~
583		~~mainbody ~= " " ~ opToSSE[operations[done+1]] ~ suffix ~ " " ~ XMM(numOnStack-1) ~ ", "~~
584		~~~ XMM(numOnStack-1) ~ comment;~~
585		~~extra ~= " " ~ opToSSESingle[operations[done+1]] ~ suffix ~ " " ~ XMM(numOnStack-1) ~ ", "~~
586		~~~ XMM(numOnStack-1) ~ comment;~~
587	601	} else {
588	602	mainbody ~= " " ~ opToSSE[operations[done+1]] ~ suffix ~ " " ~ XMM(numOnStack-1) ~ ", "

trunk/blade/PostfixX86.d

r170	r171
14	14	* _ ST(1)-ST(0) and pop stack
15	15	* = store stack top and pop stack
	16	* , duplicate stack top (so ,* means ST=STST, ,+ means ST=2)
	17	* a abs
	18	* n unary negation
16	19	*
17	20	* NOT YET IMPLEMENTED:
18	21	* 1 the literal one (used to initialize a product, for example)
19		* sc ~~ST(0) = sine(ST(0)) ST(0) = cos(ST(0))~~
20		* q ~~ST(0) = sqrt(ST(0))~~
	22	* sc sin, cos
	23	* q sqrt
21	24	*/
22	25
…	…
57	60	return "0" ~ doVisit(this_,args[0]) ~ doVisit(this_, args[1]) ~ "*+";
58	61	}
	62	if (func=="sum") return "0" ~ doVisit(this_, args[0]) ~ "+";
	63	if (func=="abs") return doVisit(this_,args[0]) ~ "a";
59	64	assert(0, "BLADE ICE: Unsupported");
60	65	}
61	66	ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {
	67	if (op=="-") return doVisit(this_, expr) ~ "n"; // unary minus
62	68	assert(0, "BLADE ICE: Unsupported");
63	69	}
…	…
134	140	}
135	141	if (func=="sum") return "0" ~ doVisit(this_, args[0]) ~ "+";
	142	if (func=="abs") return doVisit(this_,args[0]) ~ "a";
136	143	assert(0, "BLADE ICE: Unsupported");
137	144	}
138	145	ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {
	146	if (op=="-") return doVisit(this_, expr) ~ "n"; // unary minus
139	147	assert(0, "BLADE ICE: Unsupported");
140	148	}
…	…
157	165	return second ~ first ~ "=";
158	166	}
	167	if (second == first) return first ~ "," ~ op;
159	168
160	169	// SSE OPTIMISATION #1

trunk/blade/SyntaxTree.d

r157	r171
178	178	}
179	179	if (symbol.length>0) {
180		// Find it, and add it to the symbol table if not already present.
181		int k=0;
182		while(k < symbols.length && symbol!=symbols[k]) ++k;
183		code ~= cast(char)('A' + k);
184		if (k==symbols.length) {
185		symbols ~= symbol;
186		}
187		symbol="";
	180	// Add the new symbol to the symbol table
	181	code ~= cast(char)('A' + symbols.length);
	182	symbols ~= symbol;
	183	symbol = "";
188	184	}
189	185	if (c=='.') { // it was opSlice. Skip the next . as well.
…	…
217	213	return symbols;
218	214	}
219
220	215
221	216	// ==== SYNTAX PASS ====

mathextra

Navigation

Changeset 171

Legend:

trunk/blade/Blade.d

trunk/blade/BladeDemo.d

trunk/blade/BladeRank.d

trunk/blade/BladeSimplify.d

trunk/blade/CodegenX86.d

trunk/blade/PostfixX86.d

trunk/blade/SyntaxTree.d

Download in other formats: