Added Pico Co-Processor Interface (PCPI)

Makefile

@@ -1,6 +1,6 @@
 
 TEST_OBJS = $(addsuffix .o,$(basename $(wildcard tests/*.S)))
-FIRMWARE_OBJS = firmware/start.o firmware/irq.o firmware/print.o firmware/sieve.o firmware/stats.o
+FIRMWARE_OBJS = firmware/start.o firmware/irq.o firmware/print.o firmware/sieve.o firmware/multest.o firmware/stats.o
 
 test: testbench.exe firmware/firmware.hex
 	vvp -N testbench.exe

firmware/firmware.h

@@ -16,6 +16,13 @@ void print_hex(unsigned int val);
 // sieve.c
 void sieve();
 
+// multest.c
+uint32_t hard_mul(uint32_t a, uint32_t b);
+uint32_t hard_mulh(uint32_t a, uint32_t b);
+uint32_t hard_mulhsu(uint32_t a, uint32_t b);
+uint32_t hard_mulhu(uint32_t a, uint32_t b);
+void multest();
+
 // stats.c
 void stats();
 

firmware/multest.c

@@ -0,0 +1,79 @@
+#include "firmware.h"
+
+uint32_t xorshift32() {
+	static uint32_t x = 314159265;
+	x ^= x << 13;
+	x ^= x >> 17;
+	x ^= x << 5;
+	return x;
+}
+
+void multest()
+{
+	int i;
+	for (i = 0; i < 10; i++)
+	{
+		uint32_t a = xorshift32();
+		uint32_t b = xorshift32();
+
+		uint64_t au = a, bu = b;
+		int64_t as = (int32_t)a, bs = (int32_t)b;
+
+		print_str("input [");
+		print_hex(as >> 32);
+		print_str("] ");
+		print_hex(a);
+		print_str(" [");
+		print_hex(bs >> 32);
+		print_str("] ");
+		print_hex(b);
+		print_chr('\n');
+
+		uint32_t h_mul, h_mulh, h_mulhsu, h_mulhu;
+		print_str("hard   ");
+
+		h_mul = hard_mul(a, b);
+		print_hex(h_mul);
+		print_str("  ");
+
+		h_mulh = hard_mulh(a, b);
+		print_hex(h_mulh);
+		print_str("  ");
+
+		h_mulhsu = hard_mulhsu(a, b);
+		print_hex(h_mulhsu);
+		print_str("  ");
+
+		h_mulhu = hard_mulhu(a, b);
+		print_hex(h_mulhu);
+		print_chr('\n');
+
+		uint32_t s_mul, s_mulh, s_mulhsu, s_mulhu;
+		print_str("soft   ");
+
+		s_mul = a * b;
+		print_hex(s_mul);
+		print_str("  ");
+
+		s_mulh = (as * bs) >> 32;
+		print_hex(s_mulh);
+		print_str("  ");
+
+		s_mulhsu = (as * bu) >> 32;
+		print_hex(s_mulhsu);
+		print_str("  ");
+
+		s_mulhu = (au * bu) >> 32;
+		print_hex(s_mulhu);
+		print_str("  ");
+
+		if (s_mul != h_mul || s_mulh != h_mulh || s_mulhsu != h_mulhsu || s_mulhu != h_mulhu) {
+			print_str("ERROR!\n");
+			asm volatile ("sbreak");
+			return;
+		}
+
+		print_str(" OK\n");
+	}
+}
+

firmware/start.S

@@ -1,6 +1,11 @@
 	.section .text
 	.global irq
 	.global sieve
+	.global multest
+	.global hard_mul
+	.global hard_mulh
+	.global hard_mulhsu
+	.global hard_mulhu
 	.global stats
 
 #define TEST(n) \
@@ -220,6 +225,9 @@ start:
 	/* jump to sieve C code */
 	jal ra,sieve
 
+	/* jump to sieve C code */
+	jal ra,multest
+
 	/* jump to stats C code */
 	jal ra,stats
 
@@ -239,3 +247,19 @@ start:
 	/* break */
 	sbreak
 
+hard_mul:
+	mul a0, a0, a1
+	ret
+
+hard_mulh:
+	mulh a0, a0, a1
+	ret
+
+hard_mulhsu:
+	mulhsu a0, a0, a1
+	ret
+
+hard_mulhu:
+	mulhu a0, a0, a1
+	ret
+

picorv32.v

@@ -30,6 +30,7 @@ module picorv32 #(
 	parameter [ 0:0] ENABLE_REGS_16_31 = 1,
 	parameter [ 0:0] ENABLE_REGS_DUALPORT = 1,
 	parameter [ 0:0] LATCHED_MEM_RDATA = 0,
+	parameter [ 0:0] ENABLE_PCPI = 0,
 	parameter [ 0:0] ENABLE_IRQ = 0,
 	parameter [31:0] MASKED_IRQ = 32'h 0000_0000,
 	parameter [31:0] PROGADDR_RESET = 32'h 0000_0000,
@@ -47,14 +48,25 @@ module picorv32 #(
 	output reg [ 3:0] mem_wstrb,
 	input      [31:0] mem_rdata,
 
-	// look-ahead interface
+	// Look-Ahead Interface
 	output            mem_la_read,
 	output            mem_la_write,
 	output     [31:0] mem_la_addr,
 	output reg [31:0] mem_la_wdata,
 	output reg [ 3:0] mem_la_wstrb,
 
-	// IRQ interface
+	// Pico Co-Processor Interface (PCPI)
+	output reg        pcpi_insn_valid,
+	output reg [31:0] pcpi_insn,
+	output reg        pcpi_rs1_valid,
+	output     [31:0] pcpi_rs1,
+	output reg        pcpi_rs2_valid,
+	output     [31:0] pcpi_rs2,
+	input             pcpi_rd_valid,
+	input      [31:0] pcpi_rd,
+	input             pcpi_ready,
+
+	// IRQ Interface
 	input      [31:0] irq,
 	output reg [31:0] eoi
 );
@@ -71,6 +83,9 @@ module picorv32 #(
 	reg [31:0] cpuregs [0:regfile_size-1];
 	reg [4:0] reg_sh;
 
+	assign pcpi_rs1 = reg_op1;
+	assign pcpi_rs2 = reg_op2;
+
 	wire [31:0] next_pc;
 
 	reg irq_active;
@@ -318,6 +333,10 @@ module picorv32 #(
 		end
 
 		if (decoder_trigger && !decoder_pseudo_trigger) begin
+			if (ENABLE_PCPI) begin
+				pcpi_insn <= mem_rdata_q;
+			end
+
 			instr_beq   <= is_beq_bne_blt_bge_bltu_bgeu && mem_rdata_q[14:12] == 3'b000;
 			instr_bne   <= is_beq_bne_blt_bge_bltu_bgeu && mem_rdata_q[14:12] == 3'b001;
 			instr_blt   <= is_beq_bne_blt_bge_bltu_bgeu && mem_rdata_q[14:12] == 3'b100;
@@ -437,6 +456,9 @@ module picorv32 #(
 	reg [31:0] current_pc;
 	assign next_pc = latched_store && latched_branch ? reg_out : reg_next_pc;
 
+	reg [7:0] pcpi_timeout_counter;
+	reg pcpi_timeout;
+
 	reg [31:0] next_irq_pending;
 	reg do_waitirq;
 
@@ -489,6 +511,15 @@ module picorv32 #(
 
 		reg_alu_out <= alu_out;
 
+		if (ENABLE_PCPI) begin
+			if (pcpi_insn_valid) begin
+				if (pcpi_timeout_counter)
+					pcpi_timeout_counter <= pcpi_timeout_counter - 1;
+			end else
+				pcpi_timeout_counter <= ~0;
+			pcpi_timeout <= !pcpi_timeout_counter;
+		end
+
 		if (ENABLE_COUNTERS)
 			count_cycle <= resetn ? count_cycle + 1 : 0;
 
@@ -520,6 +551,9 @@ module picorv32 #(
 			latched_is_lu <= 0;
 			latched_is_lh <= 0;
 			latched_is_lb <= 0;
+			pcpi_insn_valid <= 0;
+			pcpi_rs1_valid <= 0;
+			pcpi_rs2_valid <= 0;
 			irq_active <= 0;
 			irq_mask <= ~0;
 			next_irq_pending = 0;
@@ -567,6 +601,12 @@ module picorv32 #(
 				reg_pc <= current_pc;
 				reg_next_pc <= current_pc;
 
+				if (ENABLE_PCPI) begin
+					pcpi_insn_valid <= 0;
+					pcpi_rs1_valid <= 0;
+					pcpi_rs2_valid <= 0;
+				end
+
 				latched_store <= 0;
 				latched_stalu <= 0;
 				latched_branch <= 0;
@@ -618,6 +658,20 @@ module picorv32 #(
 				$display("DECODE: 0x%08x %-0s", reg_pc, instruction ? instruction : "UNKNOWN");
 `endif
 				if (instr_trap) begin
+					if (ENABLE_PCPI) begin
+						pcpi_rs1_valid <= 1;
+						pcpi_insn_valid <= 1;
+						reg_op1 <= decoded_rs1 ? cpuregs[decoded_rs1] : 0;
+						if (ENABLE_REGS_DUALPORT) begin
+							pcpi_rs2_valid <= 1;
+							reg_sh <= decoded_rs2 ? cpuregs[decoded_rs2] : 0;
+							reg_op2 <= decoded_rs2 ? cpuregs[decoded_rs2] : 0;
+							if (pcpi_ready) begin
+								reg_out <= pcpi_rd;
+								latched_store <= pcpi_rd_valid;
+								cpu_state <= cpu_state_fetch;
+							end else
+							if (pcpi_timeout) begin
 `ifdef DEBUG
 								$display("SBREAK OR UNSUPPORTED INSN AT 0x%08x", reg_pc);
 `endif
@@ -626,6 +680,20 @@ module picorv32 #(
 									cpu_state <= cpu_state_fetch;
 								end else
 									cpu_state <= cpu_state_trap;
+							end
+						end else begin
+							cpu_state <= cpu_state_ld_rs2;
+						end
+					end else begin
+`ifdef DEBUG
+						$display("SBREAK OR UNSUPPORTED INSN AT 0x%08x", reg_pc);
+`endif
+						if (ENABLE_IRQ && !irq_mask[irq_sbreak] && !irq_active) begin
+							next_irq_pending[irq_sbreak] = 1;
+							cpu_state <= cpu_state_fetch;
+						end else
+							cpu_state <= cpu_state_trap;
+					end
 				end else
 				if (is_rdcycle_rdcycleh_rdinstr_rdinstrh) begin
 					(* parallel_case, full_case *)
@@ -718,6 +786,24 @@ module picorv32 #(
 `endif
 				reg_sh <= decoded_rs2 ? cpuregs[decoded_rs2] : 0;
 				reg_op2 <= decoded_rs2 ? cpuregs[decoded_rs2] : 0;
+				if (ENABLE_PCPI && pcpi_insn_valid) begin
+					pcpi_rs2_valid <= 1;
+					if (pcpi_ready) begin
+						reg_out <= pcpi_rd;
+						latched_store <= pcpi_rd_valid;
+						cpu_state <= cpu_state_fetch;
+					end else
+					if (pcpi_timeout) begin
+`ifdef DEBUG
+						$display("SBREAK OR UNSUPPORTED INSN AT 0x%08x", reg_pc);
+`endif
+						if (ENABLE_IRQ && !irq_mask[irq_sbreak] && !irq_active) begin
+							next_irq_pending[irq_sbreak] = 1;
+							cpu_state <= cpu_state_fetch;
+						end else
+							cpu_state <= cpu_state_trap;
+					end
+				end else
 				if (is_sb_sh_sw) begin
 					cpu_state <= cpu_state_stmem;
 					mem_do_rinst <= 1;
@@ -874,6 +960,70 @@ module picorv32 #(
 endmodule
 
 
+/***************************************************************
+ * picorv32_pcpi_mul
+ ***************************************************************/
+
+module picorv32_pcpi_mul (
+	input clk, resetn,
+
+	input             pcpi_insn_valid,
+	input      [31:0] pcpi_insn,
+	input             pcpi_rs1_valid,
+	input      [31:0] pcpi_rs1,
+	input             pcpi_rs2_valid,
+	input      [31:0] pcpi_rs2,
+	output reg        pcpi_rd_valid,
+	output reg [31:0] pcpi_rd,
+	output reg        pcpi_ready
+);
+	reg instr_mul, instr_mulh, instr_mulhsu, instr_mulhu;
+	wire instr_any_mul = |{instr_mul, instr_mulh, instr_mulhsu, instr_mulhu};
+	wire instr_any_mulh = |{instr_mulh, instr_mulhsu, instr_mulhu};
+	wire instr_rs1_signed = |{instr_mulh, instr_mulhsu};
+	wire instr_rs2_signed = |{instr_mulh};
+
+	always @(posedge clk) begin
+		instr_mul <= 0;
+		instr_mulh <= 0;
+		instr_mulhsu <= 0;
+		instr_mulhu <= 0;
+
+		if (pcpi_insn_valid && pcpi_insn[6:0] == 7'b0110011 && pcpi_insn[31:25] == 7'b0000001) begin
+			case (pcpi_insn[14:12])
+				3'b000: instr_mul <= 1;
+				3'b001: instr_mulh <= 1;
+				3'b010: instr_mulhsu <= 1;
+				3'b011: instr_mulhu <= 1;
+			endcase
+		end
+	end
+
+	// FIXME: This is just a behavioral model
+
+	reg [63:0] rs1, rs2;
+	always @(posedge clk) begin
+		pcpi_rd_valid <= 0;
+		pcpi_ready <= 0;
+		if (pcpi_rs1_valid && pcpi_rs2_valid && instr_any_mul) begin
+			if (instr_rs1_signed)
+				rs1 = $signed(pcpi_rs1);
+			else
+				rs1 = $unsigned(pcpi_rs1);
+
+			if (instr_rs2_signed)
+				rs2 = $signed(pcpi_rs2);
+			else
+				rs2 = $unsigned(pcpi_rs2);
+
+			pcpi_rd <= instr_any_mulh ? (rs1 * rs2) >> 32 : rs1 * rs2;
+			pcpi_rd_valid <= 1;
+			pcpi_ready <= 1;
+		end
+	end
+endmodule
+
+
 /***************************************************************
  * picorv32_axi
  ***************************************************************/
@@ -882,6 +1032,7 @@ module picorv32_axi #(
 	parameter [ 0:0] ENABLE_COUNTERS = 1,
 	parameter [ 0:0] ENABLE_REGS_16_31 = 1,
 	parameter [ 0:0] ENABLE_REGS_DUALPORT = 1,
+	parameter [ 0:0] ENABLE_MUL = 0,
 	parameter [ 0:0] ENABLE_IRQ = 0,
 	parameter [31:0] MASKED_IRQ = 32'h 0000_0000,
 	parameter [31:0] PROGADDR_RESET = 32'h 0000_0000,
@@ -926,6 +1077,16 @@ module picorv32_axi #(
 	wire        mem_ready;
 	wire [31:0] mem_rdata;
 
+	wire pcpi_insn_valid;
+	wire [31:0] pcpi_insn;
+	wire pcpi_rs1_valid;
+	wire [31:0] pcpi_rs1;
+	wire pcpi_rs2_valid;
+	wire [31:0] pcpi_rs2;
+	wire pcpi_rd_valid;
+	wire [31:0] pcpi_rd;
+	wire pcpi_ready;
+
 	picorv32_axi_adapter axi_adapter (
 		.clk            (clk            ),
 		.resetn         (resetn         ),
@@ -955,10 +1116,30 @@ module picorv32_axi #(
 		.mem_rdata      (mem_rdata      )
 	);
 
+	generate if (ENABLE_MUL) begin
+		picorv32_pcpi_mul pcpi_mul (
+			.clk            (clk            ),
+			.resetn         (resetn         ),
+			.pcpi_insn_valid(pcpi_insn_valid),
+			.pcpi_insn      (pcpi_insn      ),
+			.pcpi_rs1_valid (pcpi_rs1_valid ),
+			.pcpi_rs1       (pcpi_rs1       ),
+			.pcpi_rs2_valid (pcpi_rs2_valid ),
+			.pcpi_rs2       (pcpi_rs2       ),
+			.pcpi_rd_valid  (pcpi_rd_valid  ),
+			.pcpi_rd        (pcpi_rd        ),
+			.pcpi_ready     (pcpi_ready     )
+		);
+	end else begin
+		assign pcpi_rd = 1'bx;
+		assign pcpi_ready = 0;
+	end endgenerate
+
 	picorv32 #(
 		.ENABLE_COUNTERS     (ENABLE_COUNTERS     ),
 		.ENABLE_REGS_16_31   (ENABLE_REGS_16_31   ),
 		.ENABLE_REGS_DUALPORT(ENABLE_REGS_DUALPORT),
+		.ENABLE_PCPI         (ENABLE_MUL          ),
 		.ENABLE_IRQ          (ENABLE_IRQ          ),
 		.MASKED_IRQ          (MASKED_IRQ          ),
 		.PROGADDR_RESET      (PROGADDR_RESET      ),
@@ -967,6 +1148,7 @@ module picorv32_axi #(
 		.clk      (clk   ),
 		.resetn   (resetn),
 		.trap     (trap  ),
+
 		.mem_valid(mem_valid),
 		.mem_addr (mem_addr ),
 		.mem_wdata(mem_wdata),
@@ -974,6 +1156,17 @@ module picorv32_axi #(
 		.mem_instr(mem_instr),
 		.mem_ready(mem_ready),
 		.mem_rdata(mem_rdata),
+
+		.pcpi_insn_valid(pcpi_insn_valid),
+		.pcpi_insn      (pcpi_insn      ),
+		.pcpi_rs1_valid (pcpi_rs1_valid ),
+		.pcpi_rs1       (pcpi_rs1       ),
+		.pcpi_rs2_valid (pcpi_rs2_valid ),
+		.pcpi_rs2       (pcpi_rs2       ),
+		.pcpi_rd_valid  (pcpi_rd_valid  ),
+		.pcpi_rd        (pcpi_rd        ),
+		.pcpi_ready     (pcpi_ready     ),
+
 		.irq(irq),
 		.eoi(eoi)
 	);

testbench.v

@@ -45,6 +45,7 @@ module testbench;
 	reg  [31:0] mem_axi_rdata;
 
 	picorv32_axi #(
+		.ENABLE_MUL(1),
 		.ENABLE_IRQ(1)
 	) uut (
 		.clk            (clk            ),