Merge pull request #4818 from povik/macc_v2

Add `$macc_v2`

docs/source/cell/word_arith.rst

@@ -1,7 +1,7 @@
 Coarse arithmetics
 ------------------
 
-.. todo:: Add information about `$alu`, `$fa`, and `$lcu` cells.
+.. todo:: Add information about `$alu`, `$fa`, `$macc_v2`, and `$lcu` cells.
 
 The `$macc` cell type represents a generalized multiply and accumulate
 operation. The cell is purely combinational. It outputs the result of summing up

docs/source/code_examples/fifo/fifo.ys

@@ -67,7 +67,7 @@ show -color maroon3 @new_cells -notitle -format dot -prefix rdata_memrdv2 o:rdat
 # ========================================================
 
 alumacc
-select -set new_cells t:$alu t:$macc
+select -set new_cells t:$alu t:$macc_v2
 show -color maroon3 @new_cells -notitle -format dot -prefix rdata_alumacc o:rdata %ci*
 
 # ========================================================

docs/source/getting_started/example_synth.rst

@@ -523,7 +523,7 @@ That brings us to the fourth and final part for the iCE40 synthesis flow:
    :name: synth_coarse4
 
 Where before each type of arithmetic operation had its own cell, e.g. `$add`, we
-now want to extract these into `$alu` and `$macc` cells which can help identify
+now want to extract these into `$alu` and `$macc_v2` cells which can help identify
 opportunities for reusing logic.  We do this by running `alumacc`, which we can
 see produce the following changes in our example design:
 

kernel/celledges.cc

@@ -453,7 +453,7 @@ bool YOSYS_NAMESPACE_PREFIX AbstractCellEdgesDatabase::add_edges_from_cell(RTLIL
 	}
 
 	// FIXME: $mul $div $mod $divfloor $modfloor $slice $concat
-	// FIXME: $lut $sop $alu $lcu $macc $fa
+	// FIXME: $lut $sop $alu $lcu $macc $macc_v2 $fa
 	// FIXME: $mul $div $mod $divfloor $modfloor $pow $slice $concat $bweqx
 	// FIXME: $lut $sop $alu $lcu $macc $fa $logic_and $logic_or $bwmux
 

kernel/celltypes.h

@@ -144,6 +144,7 @@ struct CellTypes
 
 		setup_type(ID($lcu), {ID::P, ID::G, ID::CI}, {ID::CO}, true);
 		setup_type(ID($alu), {ID::A, ID::B, ID::CI, ID::BI}, {ID::X, ID::Y, ID::CO}, true);
+		setup_type(ID($macc_v2), {ID::A, ID::B, ID::C}, {ID::Y}, true);
 		setup_type(ID($fa), {ID::A, ID::B, ID::C}, {ID::X, ID::Y}, true);
 	}
 

kernel/consteval.h

@@ -310,7 +310,7 @@ struct ConstEval
 				}
 			}
 		}
-		else if (cell->type == ID($macc))
+		else if (cell->type.in(ID($macc), ID($macc_v2)))
 		{
 			Macc macc;
 			macc.from_cell(cell);

kernel/constids.inc

@@ -276,3 +276,11 @@ X(Y)
 X(Y_WIDTH)
 X(area)
 X(capacitance)
+X(NPRODUCTS)
+X(NADDENDS)
+X(PRODUCT_NEGATED)
+X(ADDEND_NEGATED)
+X(A_WIDTHS)
+X(B_WIDTHS)
+X(C_WIDTHS)
+X(C_SIGNED)

kernel/macc.h

@@ -82,7 +82,7 @@ struct Macc
 		new_ports.swap(ports);
 	}
 
-	void from_cell(RTLIL::Cell *cell)
+	void from_cell_v1(RTLIL::Cell *cell)
 	{
 		RTLIL::SigSpec port_a = cell->getPort(ID::A);
 
@@ -136,52 +136,128 @@ struct Macc
 		log_assert(port_a_cursor == GetSize(port_a));
 	}
 
-	void to_cell(RTLIL::Cell *cell) const
+	void from_cell(RTLIL::Cell *cell)
 	{
-		RTLIL::SigSpec port_a;
-		std::vector<RTLIL::State> config_bits;
-		int max_size = 0, num_bits = 0;
+		if (cell->type == ID($macc)) {
+			from_cell_v1(cell);
+			return;
+		}
+		log_assert(cell->type == ID($macc_v2));
 
-		for (auto &port : ports) {
-			max_size = max(max_size, GetSize(port.in_a));
-			max_size = max(max_size, GetSize(port.in_b));
+		RTLIL::SigSpec port_a = cell->getPort(ID::A);
+		RTLIL::SigSpec port_b = cell->getPort(ID::B);
+		RTLIL::SigSpec port_c = cell->getPort(ID::C);
+
+		ports.clear();
+
+		int nproducts = cell->getParam(ID::NPRODUCTS).as_int();
+		const Const &product_neg = cell->getParam(ID::PRODUCT_NEGATED);
+		const Const &a_widths = cell->getParam(ID::A_WIDTHS);
+		const Const &b_widths = cell->getParam(ID::B_WIDTHS);
+		const Const &a_signed = cell->getParam(ID::A_SIGNED);
+		const Const &b_signed = cell->getParam(ID::B_SIGNED);
+		int ai = 0, bi = 0;
+		for (int i = 0; i < nproducts; i++) {
+			port_t term;
+
+			log_assert(a_signed[i] == b_signed[i]);
+			term.is_signed = (a_signed[i] == State::S1);
+			int a_width = a_widths.extract(16 * i, 16).as_int(false);
+			int b_width = b_widths.extract(16 * i, 16).as_int(false);
+
+			term.in_a = port_a.extract(ai, a_width);
+			ai += a_width;
+			term.in_b = port_b.extract(bi, b_width);
+			bi += b_width;
+			term.do_subtract = (product_neg[i] == State::S1);
+
+			ports.push_back(term);
+		}
+		log_assert(port_a.size() == ai);
+		log_assert(port_b.size() == bi);
+
+		int naddends = cell->getParam(ID::NADDENDS).as_int();
+		const Const &addend_neg = cell->getParam(ID::ADDEND_NEGATED);
+		const Const &c_widths = cell->getParam(ID::C_WIDTHS);
+		const Const &c_signed = cell->getParam(ID::C_SIGNED);
+		int ci = 0;
+		for (int i = 0; i < naddends; i++) {
+			port_t term;
+
+			term.is_signed = (c_signed[i] == State::S1);
+			int c_width = c_widths.extract(16 * i, 16).as_int(false);
+
+			term.in_a = port_c.extract(ci, c_width);
+			ci += c_width;
+			term.do_subtract = (addend_neg[i] == State::S1);
+
+			ports.push_back(term);
+		}
+		log_assert(port_c.size() == ci);
+	}
+
+	void to_cell(RTLIL::Cell *cell)
+	{
+		cell->type = ID($macc_v2);
+
+		int nproducts = 0, naddends = 0;
+		Const a_signed, b_signed, a_widths, b_widths, product_negated;
+		Const c_signed, c_widths, addend_negated;
+		SigSpec a, b, c;
+
+		for (int i = 0; i < (int) ports.size(); i++) {
+			SigSpec term_a = ports[i].in_a, term_b = ports[i].in_b;
+
+			if (term_b.empty()) {
+				// addend
+				c_widths.append(Const(term_a.size(), 16));
+				c_signed.append(ports[i].is_signed ? RTLIL::S1 : RTLIL::S0);
+				addend_negated.append(ports[i].do_subtract ? RTLIL::S1 : RTLIL::S0);
+				c.append(term_a);
+				naddends++;
+			} else {
+				// product
+				a_widths.append(Const(term_a.size(), 16));
+				b_widths.append(Const(term_b.size(), 16));
+				a_signed.append(ports[i].is_signed ? RTLIL::S1 : RTLIL::S0);
+				b_signed.append(ports[i].is_signed ? RTLIL::S1 : RTLIL::S0);
+				product_negated.append(ports[i].do_subtract ? RTLIL::S1 : RTLIL::S0);
+				a.append(term_a);
+				b.append(term_b);
+				nproducts++;
+			}
 		}
 
-		while (max_size)
-			num_bits++, max_size /= 2;
+		if (a_signed.empty())
+			a_signed = {RTLIL::Sx};
+		if (b_signed.empty())
+			b_signed = {RTLIL::Sx};
+		if (c_signed.empty())
+			c_signed = {RTLIL::Sx};
+		if (a_widths.empty())
+			a_widths = {RTLIL::Sx};
+		if (b_widths.empty())
+			b_widths = {RTLIL::Sx};
+		if (c_widths.empty())
+			c_widths = {RTLIL::Sx};
+		if (product_negated.empty())
+			product_negated = {RTLIL::Sx};
+		if (addend_negated.empty())
+			addend_negated = {RTLIL::Sx};
 
-		log_assert(num_bits < 16);
-		config_bits.push_back(num_bits & 1 ? State::S1 : State::S0);
-		config_bits.push_back(num_bits & 2 ? State::S1 : State::S0);
-		config_bits.push_back(num_bits & 4 ? State::S1 : State::S0);
-		config_bits.push_back(num_bits & 8 ? State::S1 : State::S0);
-
-		for (auto &port : ports)
-		{
-			if (GetSize(port.in_a) == 0)
-				continue;
-
-			config_bits.push_back(port.is_signed ? State::S1 : State::S0);
-			config_bits.push_back(port.do_subtract ? State::S1 : State::S0);
-
-			int size_a = GetSize(port.in_a);
-			for (int i = 0; i < num_bits; i++)
-				config_bits.push_back(size_a & (1 << i) ? State::S1 : State::S0);
-
-			int size_b = GetSize(port.in_b);
-			for (int i = 0; i < num_bits; i++)
-				config_bits.push_back(size_b & (1 << i) ? State::S1 : State::S0);
-
-			port_a.append(port.in_a);
-			port_a.append(port.in_b);
-		}
-
-		cell->setPort(ID::A, port_a);
-		cell->setPort(ID::B, {});
-		cell->setParam(ID::CONFIG, config_bits);
-		cell->setParam(ID::CONFIG_WIDTH, GetSize(config_bits));
-		cell->setParam(ID::A_WIDTH, GetSize(port_a));
-		cell->setParam(ID::B_WIDTH, 0);
+		cell->setParam(ID::NPRODUCTS, nproducts);
+		cell->setParam(ID::PRODUCT_NEGATED, product_negated);
+		cell->setParam(ID::NADDENDS, naddends);
+		cell->setParam(ID::ADDEND_NEGATED, addend_negated);
+		cell->setParam(ID::A_SIGNED, a_signed);
+		cell->setParam(ID::B_SIGNED, b_signed);
+		cell->setParam(ID::C_SIGNED, c_signed);
+		cell->setParam(ID::A_WIDTHS, a_widths);
+		cell->setParam(ID::B_WIDTHS, b_widths);
+		cell->setParam(ID::C_WIDTHS, c_widths);
+		cell->setPort(ID::A, a);
+		cell->setPort(ID::B, b);
+		cell->setPort(ID::C, c);
 	}
 
 	bool eval(RTLIL::Const &result) const

kernel/rtlil.cc

@@ -540,6 +540,12 @@ void RTLIL::Const::bitvectorize() const {
 	}
 }
 
+void RTLIL::Const::append(const RTLIL::Const &other) {
+	bitvectorize();
+	bitvectype& bv = get_bits();
+	bv.insert(bv.end(), other.begin(), other.end());
+}
+
 RTLIL::State RTLIL::Const::const_iterator::operator*() const {
 	if (auto bv = parent.get_if_bits())
 		return (*bv)[idx];
@@ -1461,6 +1467,40 @@ namespace {
 				return;
 			}
 
+			if (cell->type == ID($macc_v2)) {
+				if (param(ID::NPRODUCTS) < 0)
+					error(__LINE__);
+				if (param(ID::NADDENDS) < 0)
+					error(__LINE__);
+				param_bits(ID::PRODUCT_NEGATED, max(param(ID::NPRODUCTS), 1));
+				param_bits(ID::ADDEND_NEGATED, max(param(ID::NADDENDS), 1));
+				param_bits(ID::A_SIGNED, max(param(ID::NPRODUCTS), 1));
+				param_bits(ID::B_SIGNED, max(param(ID::NPRODUCTS), 1));
+				param_bits(ID::C_SIGNED, max(param(ID::NADDENDS), 1));
+				if (cell->getParam(ID::A_SIGNED) != cell->getParam(ID::B_SIGNED))
+					error(__LINE__);
+				param_bits(ID::A_WIDTHS, max(param(ID::NPRODUCTS) * 16, 1));
+				param_bits(ID::B_WIDTHS, max(param(ID::NPRODUCTS) * 16, 1));
+				param_bits(ID::C_WIDTHS, max(param(ID::NADDENDS) * 16, 1));
+				const Const &a_width = cell->getParam(ID::A_WIDTHS);
+				const Const &b_width = cell->getParam(ID::B_WIDTHS);
+				const Const &c_width = cell->getParam(ID::C_WIDTHS);
+				int a_width_sum = 0, b_width_sum = 0, c_width_sum = 0;
+				for (int i = 0; i < param(ID::NPRODUCTS); i++) {
+					a_width_sum += a_width.extract(16 * i, 16).as_int(false);
+					b_width_sum += b_width.extract(16 * i, 16).as_int(false);
+				}
+				for (int i = 0; i < param(ID::NADDENDS); i++) {
+					c_width_sum += c_width.extract(16 * i, 16).as_int(false);
+				}
+				port(ID::A, a_width_sum);
+				port(ID::B, b_width_sum);
+				port(ID::C, c_width_sum);
+				port(ID::Y, param(ID::Y_WIDTH));
+				check_expected();
+				return;
+			}
+
 			if (cell->type == ID($logic_not)) {
 				param_bool(ID::A_SIGNED);
 				port(ID::A, param(ID::A_WIDTH));
@@ -4093,6 +4133,11 @@ void RTLIL::Cell::fixup_parameters(bool set_a_signed, bool set_b_signed)
 		return;
 	}
 
+	if (type == ID($macc_v2)) {
+		parameters[ID::Y_WIDTH] = GetSize(connections_[ID::Y]);
+		return;
+	}
+
 	bool signedness_ab = !type.in(ID($slice), ID($concat), ID($macc));
 
 	if (connections_.count(ID::A)) {

kernel/rtlil.h

@@ -738,6 +738,8 @@ public:
 	bool empty() const;
 	void bitvectorize() const;
 
+	void append(const RTLIL::Const &other);
+
 	class const_iterator {
 	private:
 		const Const& parent;

kernel/satgen.cc

@@ -740,7 +740,7 @@ bool SatGen::importCell(RTLIL::Cell *cell, int timestep)
 		return true;
 	}
 
-	if (cell->type == ID($macc))
+	if (cell->type.in(ID($macc), ID($macc_v2)))
 	{
 		std::vector<int> a = importDefSigSpec(cell->getPort(ID::A), timestep);
 		std::vector<int> y = importDefSigSpec(cell->getPort(ID::Y), timestep);

passes/techmap/booth.cc

@@ -218,7 +218,7 @@ struct BoothPassWorker {
 
 				log_assert(cell->getParam(ID::A_SIGNED).as_bool() == cell->getParam(ID::B_SIGNED).as_bool());
 				is_signed = cell->getParam(ID::A_SIGNED).as_bool();
-			} else if (cell->type == ID($macc)) {
+			} else if (cell->type.in(ID($macc), ID($macc_v2))) {
 				Macc macc;
 				macc.from_cell(cell);
 

passes/techmap/maccmap.cc

@@ -403,7 +403,7 @@ struct MaccmapPass : public Pass {
 
 		for (auto mod : design->selected_modules())
 		for (auto cell : mod->selected_cells())
-			if (cell->type == ID($macc)) {
+			if (cell->type.in(ID($macc), ID($macc_v2))) {
 				log("Mapping %s.%s (%s).\n", log_id(mod), log_id(cell), log_id(cell->type));
 				maccmap(mod, cell, unmap_mode);
 				mod->remove(cell);

passes/techmap/techmap.cc

@@ -554,8 +554,8 @@ struct TechmapWorker
 
 							if (extmapper_name == "maccmap") {
 								log("Creating %s with maccmap.\n", log_id(extmapper_module));
-								if (extmapper_cell->type != ID($macc))
-									log_error("The maccmap mapper can only map $macc (not %s) cells!\n", log_id(extmapper_cell->type));
+								if (!extmapper_cell->type.in(ID($macc), ID($macc_v2)))
+									log_error("The maccmap mapper can only map $macc/$macc_v2 (not %s) cells!\n", log_id(extmapper_cell->type));
 								maccmap(extmapper_module, extmapper_cell);
 								extmapper_module->remove(extmapper_cell);
 							}
@@ -600,8 +600,8 @@ struct TechmapWorker
 						}
 
 						if (extmapper_name == "maccmap") {
-							if (cell->type != ID($macc))
-								log_error("The maccmap mapper can only map $macc (not %s) cells!\n", log_id(cell->type));
+							if (!cell->type.in(ID($macc), ID($macc_v2)))
+								log_error("The maccmap mapper can only map $macc/$macc_v2 (not %s) cells!\n", log_id(cell->type));
 							maccmap(module, cell);
 						}
 

techlibs/common/simlib.v

@@ -1207,6 +1207,120 @@ end
 
 endmodule
 
+// --------------------------------------------------------
+//  |---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|
+//-
+//-     $macc_v2 (A, B, C, Y)
+//* group arith
+//-
+//- Multiply and add.
+//- This cell represents a generic fused multiply-add operation, it supersedes the
+//- earlier $macc cell.
+//-
+module \$macc_v2 (A, B, C, Y);
+
+parameter NPRODUCTS = 0;
+parameter NADDENDS = 0;
+parameter A_WIDTHS = 16'h0000;
+parameter B_WIDTHS = 16'h0000;
+parameter C_WIDTHS = 16'h0000;
+parameter Y_WIDTH = 0;
+
+parameter PRODUCT_NEGATED = 1'bx;
+parameter ADDEND_NEGATED = 1'bx;
+parameter A_SIGNED = 1'bx;
+parameter B_SIGNED = 1'bx;
+parameter C_SIGNED = 1'bx;
+
+function integer sum_widths1;
+	input [(16*NPRODUCTS)-1:0] widths;
+	integer i;
+	begin
+		sum_widths1 = 0;
+		for (i = 0; i < NPRODUCTS; i++) begin
+			sum_widths1 = sum_widths1 + widths[16*i+:16];
+		end
+	end
+endfunction
+
+function integer sum_widths2;
+	input [(16*NADDENDS)-1:0] widths;
+	integer i;
+	begin
+		sum_widths2 = 0;
+		for (i = 0; i < NADDENDS; i++) begin
+			sum_widths2 = sum_widths2 + widths[16*i+:16];
+		end
+	end
+endfunction
+
+input [sum_widths1(A_WIDTHS)-1:0] A; // concatenation of LHS factors
+input [sum_widths1(B_WIDTHS)-1:0] B; // concatenation of RHS factors
+input [sum_widths2(C_WIDTHS)-1:0] C; // concatenation of summands
+output reg [Y_WIDTH-1:0] Y; // output sum
+
+integer i, j, ai, bi, ci, aw, bw, cw;
+reg [Y_WIDTH-1:0] product;
+reg [Y_WIDTH-1:0] addend, oper_a, oper_b;
+
+always @* begin
+	Y = 0;
+	ai = 0;
+	bi = 0;
+	for (i = 0; i < NPRODUCTS; i = i+1)
+	begin
+		aw = A_WIDTHS[16*i+:16];
+		bw = B_WIDTHS[16*i+:16];
+
+		oper_a = 0;
+		oper_b = 0;
+		for (j = 0; j < Y_WIDTH && j < aw; j = j + 1)
+			oper_a[j] = A[ai + j];
+		for (j = 0; j < Y_WIDTH && j < bw; j = j + 1)
+			oper_b[j] = B[bi + j];
+		// A_SIGNED[i] == B_SIGNED[i] as RTLIL invariant
+		if (A_SIGNED[i] && B_SIGNED[i]) begin
+			for (j = aw; j > 0 && j < Y_WIDTH; j = j + 1)
+				oper_a[j] = oper_a[j - 1];
+			for (j = bw; j > 0 && j < Y_WIDTH; j = j + 1)
+				oper_b[j] = oper_b[j - 1];
+		end
+
+		product = oper_a * oper_b;
+
+		if (PRODUCT_NEGATED[i])
+			Y = Y - product;
+		else
+			Y = Y + product;
+
+		ai = ai + aw;
+		bi = bi + bw;
+	end
+
+	ci = 0;
+	for (i = 0; i < NADDENDS; i = i+1)
+	begin
+		cw = C_WIDTHS[16*i+:16];
+
+		addend = 0;
+		for (j = 0; j < Y_WIDTH && j < cw; j = j + 1)
+			addend[j] = C[ci + j];
+		if (C_SIGNED[i]) begin
+			for (j = cw; j > 0 && j < Y_WIDTH; j = j + 1)
+				addend[j] = addend[j - 1];
+		end
+
+		if (ADDEND_NEGATED[i])
+			Y = Y - addend;
+		else
+			Y = Y + addend;
+
+		ci = ci + cw;
+	end
+end
+
+endmodule
+
 // --------------------------------------------------------
 //* ver 2
 //* title Divider

techlibs/common/techmap.v

@@ -290,7 +290,7 @@ module _90_alu (A, B, CI, BI, X, Y, CO);
 endmodule
 
 (* techmap_maccmap *)
-(* techmap_celltype = "$macc" *)
+(* techmap_celltype = "$macc $macc_v2" *)
 module _90_macc;
 endmodule
 

tests/alumacc/basic.ys

@@ -0,0 +1,61 @@
+read_verilog <<EOF
+module gate(input signed [2:0] a1, input signed [2:0] b1,
+			input [1:0] a2, input [3:0] b2, input c, input d, output signed [3:0] y);
+  wire signed [3:0] ab1;
+  assign ab1 = a1 * b1;
+  assign y = ab1 + a2*b2 + c + d + 1;
+endmodule
+EOF
+prep
+equiv_opt -assert alumacc
+design -load postopt
+stat
+design -save save
+equiv_opt -assert maccmap
+design -load save
+equiv_opt -assert maccmap -unmap
+
+design -reset
+read_verilog <<EOF
+module gate(input signed [2:0] a1, input signed [1:0] b1, output signed [3:0] y);
+  assign y = a1 * b1;
+endmodule
+EOF
+prep
+equiv_opt -assert alumacc
+design -load postopt
+stat
+design -save save
+equiv_opt -assert maccmap
+design -load save
+equiv_opt -assert maccmap -unmap
+
+design -reset
+read_verilog <<EOF
+module gate(input [2:0] a, input [1:0] b, output [3:0] y);
+  assign y = a * b;
+endmodule
+EOF
+prep
+equiv_opt -assert alumacc
+design -load postopt
+stat
+design -save save
+equiv_opt -assert maccmap
+design -load save
+equiv_opt -assert maccmap -unmap
+
+design -reset
+read_verilog <<EOF
+module gate(input [2:0] a, input [1:0] b, input [1:0] c, output [3:0] y);
+  assign y = a * b - c;
+endmodule
+EOF
+prep
+equiv_opt -assert alumacc
+design -load postopt
+stat
+design -save save
+equiv_opt -assert maccmap
+design -load save
+equiv_opt -assert maccmap -unmap

tests/alumacc/macc_infer_n_unmap.ys

@@ -0,0 +1,19 @@
+read_verilog <<EOF
+module gate(input signed [2:0] a1, input signed [2:0] b1,
+			input [1:0] a2, input [3:0] b2, input c, input d, output signed [3:0] y);
+  wire signed [3:0] ab1;
+  assign ab1 = a1 * b1;
+  assign y = ab1 + a2*b2 + c + d + 1;
+endmodule
+EOF
+prep
+
+design -save gold
+alumacc
+opt_clean
+select -assert-count 1 t:$macc_v2
+maccmap -unmap
+design -copy-from gold -as gold gate
+equiv_make gold gate equiv
+equiv_induct equiv
+equiv_status -assert equiv