fpga fixes
This commit is contained in:
@@ -37,9 +37,9 @@ static const scope_signal_t scope_signals[] = {
|
||||
{ 32, "dram_req_addr" },
|
||||
{ 1, "dram_req_rw" },
|
||||
{ 16, "dram_req_byteen" },
|
||||
{ 32, "dram_req_data" },
|
||||
{ 128, "dram_req_data" },
|
||||
{ 29, "dram_req_tag" },
|
||||
{ 32, "dram_rsp_data" },
|
||||
{ 128, "dram_rsp_data" },
|
||||
{ 29, "dram_rsp_tag" },
|
||||
|
||||
{ 32, "snp_req_addr" },
|
||||
@@ -55,12 +55,12 @@ static const scope_signal_t scope_signals[] = {
|
||||
|
||||
{ NW_BITS, "dcache_req_warp_num" },
|
||||
{ 32, "dcache_req_curr_PC" },
|
||||
{ 32, "dcache_req_addr" },
|
||||
{ 64, "dcache_req_addr" },
|
||||
{ 1, "dcache_req_rw" },
|
||||
{ 4, "dcache_req_byteen" },
|
||||
{ 32, "dcache_req_data" },
|
||||
{ 8, "dcache_req_byteen" },
|
||||
{ 64, "dcache_req_data" },
|
||||
{ NW_BITS, "dcache_req_tag" },
|
||||
{ 32, "dcache_rsp_data" },
|
||||
{ 64, "dcache_rsp_data" },
|
||||
{ NW_BITS, "dcache_rsp_tag" },
|
||||
|
||||
{ NW_BITS, "decode_warp_num" },
|
||||
@@ -70,14 +70,26 @@ static const scope_signal_t scope_signals[] = {
|
||||
{ 5, "decode_rs2" },
|
||||
|
||||
{ NW_BITS, "execute_warp_num" },
|
||||
{ 32, "execute_curr_PC" },
|
||||
{ 5, "execute_rd" },
|
||||
{ 32, "execute_a" },
|
||||
{ 32, "execute_b" },
|
||||
{ 64, "execute_a" },
|
||||
{ 64, "execute_b" },
|
||||
|
||||
{ NW_BITS, "writeback_warp_num" },
|
||||
{ 32, "writeback_curr_PC" },
|
||||
{ 2, "writeback_wb" },
|
||||
{ 5, "writeback_rd" },
|
||||
{ 32, "writeback_data" },
|
||||
{ 64, "writeback_data" },
|
||||
|
||||
{ 32, "bank_addr_st0" },
|
||||
{ 32, "bank_addr_st1" },
|
||||
{ 32, "bank_addr_st2" },
|
||||
{ 1, "scope_bank_is_mrvq_st1" },
|
||||
{ 1, "scope_bank_miss_st1" },
|
||||
{ 1, "scope_bank_dirty_st1" },
|
||||
{ 1, "scope_bank_tag_valid_st1" },
|
||||
{ 1, "scope_bank_tag_match_st1" },
|
||||
{ 1, "scope_bank_force_miss_st1" },
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -104,11 +116,17 @@ static const scope_signal_t scope_signals[] = {
|
||||
{ NUM_THREADS, "decode_valid" },
|
||||
{ NUM_THREADS, "execute_valid" },
|
||||
{ NUM_THREADS, "writeback_valid" },
|
||||
|
||||
{ 1, "schedule_delay" },
|
||||
{ 1, "memory_delay" },
|
||||
{ 1, "exec_delay" },
|
||||
{ 1, "gpr_stage_delay" },
|
||||
{ 1, "busy" },
|
||||
|
||||
{ 1, "bank_valid_st0" },
|
||||
{ 1, "bank_valid_st1" },
|
||||
{ 1, "bank_valid_st2" },
|
||||
{ 1, "bank_stall_pipe" },
|
||||
};
|
||||
|
||||
static const int num_signals = sizeof(scope_signals) / sizeof(scope_signal_t);
|
||||
|
||||
@@ -306,7 +306,7 @@ extern int vx_copy_to_dev(vx_buffer_h hbuffer, size_t dev_maddr, size_t size, si
|
||||
auto ls_shift = (int)std::log2(line_size);
|
||||
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_IO_ADDR, (buffer->io_addr + src_offset) >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_MEM_ADDR, (dev_maddr >> ls_shift) ));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_MEM_ADDR, dev_maddr >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_DATA_SIZE, asize >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CMD, CMD_TYPE_WRITE));
|
||||
|
||||
@@ -349,7 +349,7 @@ extern int vx_copy_from_dev(vx_buffer_h hbuffer, size_t dev_maddr, size_t size,
|
||||
auto ls_shift = (int)std::log2(line_size);
|
||||
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_IO_ADDR, (buffer->io_addr + dest_offset) >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_MEM_ADDR, (dev_maddr) >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_MEM_ADDR, dev_maddr >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_DATA_SIZE, asize >> ls_shift));
|
||||
CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CMD, CMD_TYPE_READ));
|
||||
|
||||
|
||||
@@ -140,7 +140,7 @@ reg[31:0] io_data;
|
||||
clk = 0;
|
||||
end
|
||||
|
||||
#5 clk <= ~clk;
|
||||
#5 clk <= !clk;
|
||||
end
|
||||
|
||||
endmodule
|
||||
|
||||
@@ -88,4 +88,7 @@ kill -9 <pid>
|
||||
lsof +D build_ase_1c
|
||||
|
||||
# quick off cache synthesis
|
||||
make -C pipeline > pipeline/build.log 2>&1 &
|
||||
make -C cache > cache/build.log 2>&1 &
|
||||
make -C vortex > vortex/build.log 2>&1 &
|
||||
make -C pipeline > pipeline/build.log 2>&1 &
|
||||
@@ -83,6 +83,10 @@ typedef logic [$bits(t_ccip_clData) + $bits(t_cci_rdq_tag)-1:0] t_cci_rdq_data;
|
||||
|
||||
state_t state;
|
||||
|
||||
`ifdef SCOPE
|
||||
`SCOPE_SIGNALS_DECL
|
||||
`endif
|
||||
|
||||
// Vortex ports ///////////////////////////////////////////////////////////////
|
||||
|
||||
logic vx_dram_req_valid;
|
||||
@@ -384,19 +388,19 @@ assign cci_dram_wr_req_enable = (state == STATE_WRITE)
|
||||
&& (cci_dram_wr_req_ctr < csr_data_size);
|
||||
|
||||
assign vx_dram_req_enable = vortex_enabled && (avs_pending_reads < AVS_RD_QUEUE_SIZE);
|
||||
assign vx_dram_rd_req_enable = vx_dram_req_enable && vx_dram_req_valid && ~vx_dram_req_rw;
|
||||
assign vx_dram_rd_req_enable = vx_dram_req_enable && vx_dram_req_valid && !vx_dram_req_rw;
|
||||
assign vx_dram_wr_req_enable = vx_dram_req_enable && vx_dram_req_valid && vx_dram_req_rw;
|
||||
|
||||
assign cci_dram_rd_req_fire = cci_dram_rd_req_enable && ~avs_waitrequest;
|
||||
assign cci_dram_wr_req_fire = cci_dram_wr_req_enable && ~avs_waitrequest;
|
||||
assign cci_dram_rd_req_fire = cci_dram_rd_req_enable && !avs_waitrequest;
|
||||
assign cci_dram_wr_req_fire = cci_dram_wr_req_enable && !avs_waitrequest;
|
||||
|
||||
assign vx_dram_rd_req_fire = vx_dram_rd_req_enable && ~avs_waitrequest;
|
||||
assign vx_dram_wr_req_fire = vx_dram_wr_req_enable && ~avs_waitrequest;
|
||||
assign vx_dram_rd_req_fire = vx_dram_rd_req_enable && !avs_waitrequest;
|
||||
assign vx_dram_wr_req_fire = vx_dram_wr_req_enable && !avs_waitrequest;
|
||||
|
||||
assign vx_dram_rd_rsp_fire = vx_dram_rsp_valid && vx_dram_rsp_ready;
|
||||
|
||||
assign avs_pending_reads_next = avs_pending_reads
|
||||
+ (((cci_dram_rd_req_fire || vx_dram_rd_req_fire) && ~avs_rdq_pop) ? 1 :
|
||||
+ (((cci_dram_rd_req_fire || vx_dram_rd_req_fire) && !avs_rdq_pop) ? 1 :
|
||||
(~(cci_dram_rd_req_fire || vx_dram_rd_req_fire) && avs_rdq_pop) ? -1 : 0);
|
||||
|
||||
if (`VX_DRAM_LINE_WIDTH != DRAM_LINE_WIDTH) begin
|
||||
@@ -575,10 +579,10 @@ assign cci_rdq_push = cci_rd_rsp_fire;
|
||||
assign cci_rdq_din = {cp2af_sRxPort.c0.data, t_cci_rdq_tag'(cp2af_sRxPort.c0.hdr.mdata)};
|
||||
|
||||
assign cci_pending_reads_next = cci_pending_reads
|
||||
+ ((cci_rd_req_fire && ~cci_rdq_pop) ? 1 :
|
||||
(~cci_rd_req_fire && cci_rdq_pop) ? -1 : 0);
|
||||
+ ((cci_rd_req_fire && !cci_rdq_pop) ? 1 :
|
||||
(!cci_rd_req_fire && cci_rdq_pop) ? -1 : 0);
|
||||
|
||||
assign af2cp_sTxPort.c0.valid = cci_rd_req_enable && ~cci_rd_req_wait;
|
||||
assign af2cp_sTxPort.c0.valid = cci_rd_req_enable && !cci_rd_req_wait;
|
||||
|
||||
// Send read requests to CCI
|
||||
always_ff @(posedge clk)
|
||||
@@ -672,12 +676,12 @@ assign cci_wr_req_fire = af2cp_sTxPort.c1.valid && !cp2af_sRxPort.c1TxAlmFull;
|
||||
assign cci_wr_rsp_fire = (STATE_READ == state) && cp2af_sRxPort.c1.rspValid;
|
||||
|
||||
assign cci_pending_writes_next = cci_pending_writes
|
||||
+ ((cci_wr_req_fire && ~cci_wr_rsp_fire) ? 1 :
|
||||
(~cci_wr_req_fire && cci_wr_rsp_fire) ? -1 : 0);
|
||||
+ ((cci_wr_req_fire && !cci_wr_rsp_fire) ? 1 :
|
||||
(!cci_wr_req_fire && cci_wr_rsp_fire) ? -1 : 0);
|
||||
|
||||
assign cmd_read_done = (0 == cci_wr_req_ctr) && (0 == cci_pending_writes);
|
||||
|
||||
assign af2cp_sTxPort.c1.valid = cci_wr_req_enable && ~avs_rdq_empty;
|
||||
assign af2cp_sTxPort.c1.valid = cci_wr_req_enable && !avs_rdq_empty;
|
||||
|
||||
// Send write requests to CCI
|
||||
always_ff @(posedge clk)
|
||||
@@ -798,87 +802,6 @@ begin
|
||||
end
|
||||
end
|
||||
|
||||
// SCOPE //////////////////////////////////////////////////////////////////////
|
||||
|
||||
`ifdef SCOPE
|
||||
|
||||
`SCOPE_SIGNALS_DECL
|
||||
localparam SCOPE_DATAW = $bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST});
|
||||
localparam SCOPE_SR_DEPTH = 2;
|
||||
|
||||
`SCOPE_ASSIGN(scope_dram_req_valid, vx_dram_req_valid);
|
||||
`SCOPE_ASSIGN(scope_dram_req_addr, {vx_dram_req_addr, 4'b0});
|
||||
`SCOPE_ASSIGN(scope_dram_req_rw, vx_dram_req_rw);
|
||||
`SCOPE_ASSIGN(scope_dram_req_byteen,vx_dram_req_byteen);
|
||||
`SCOPE_ASSIGN(scope_dram_req_data, vx_dram_req_data[31:0]);
|
||||
`SCOPE_ASSIGN(scope_dram_req_tag, vx_dram_req_tag);
|
||||
`SCOPE_ASSIGN(scope_dram_req_ready, vx_dram_req_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_valid, vx_dram_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_data, vx_dram_rsp_data[31:0]);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_tag, vx_dram_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_ready, vx_dram_rsp_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_snp_req_valid, vx_snp_req_valid);
|
||||
`SCOPE_ASSIGN(scope_snp_req_addr, {vx_snp_req_addr, 4'b0});
|
||||
`SCOPE_ASSIGN(scope_snp_req_invalidate, vx_snp_req_invalidate);
|
||||
`SCOPE_ASSIGN(scope_snp_req_tag, vx_snp_req_tag);
|
||||
`SCOPE_ASSIGN(scope_snp_req_ready, vx_snp_req_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_valid, vx_snp_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_tag, vx_snp_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_ready, vx_snp_rsp_ready);
|
||||
|
||||
wire scope_changed = (scope_icache_req_valid && scope_icache_req_ready)
|
||||
|| (scope_icache_rsp_valid && scope_icache_rsp_ready)
|
||||
|| ((| scope_dcache_req_valid) && scope_dcache_req_ready)
|
||||
|| ((| scope_dcache_rsp_valid) && scope_dcache_rsp_ready)
|
||||
|| (scope_dram_req_valid && scope_dram_req_ready)
|
||||
|| (scope_dram_rsp_valid && scope_dram_rsp_ready)
|
||||
|| (scope_snp_req_valid && scope_snp_req_ready)
|
||||
|| (scope_snp_rsp_valid && scope_snp_rsp_ready);
|
||||
|
||||
wire scope_start = vx_reset;
|
||||
|
||||
wire [SCOPE_DATAW+1:0] scope_data_in_st[SCOPE_SR_DEPTH-1:0];
|
||||
wire [SCOPE_DATAW+1:0] scope_data_in;
|
||||
assign scope_data_in_st[0] = {`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST, scope_changed, scope_start};
|
||||
assign scope_data_in = scope_data_in_st[SCOPE_SR_DEPTH-1];
|
||||
|
||||
genvar i;
|
||||
for (i = 1; i < SCOPE_SR_DEPTH; i++) begin
|
||||
VX_generic_register #(
|
||||
.N (SCOPE_DATAW+2)
|
||||
) scope_sr (
|
||||
.clk (clk),
|
||||
.reset (SoftReset),
|
||||
.stall (0),
|
||||
.flush (0),
|
||||
.in (scope_data_in_st[i-1]),
|
||||
.out (scope_data_in_st[i])
|
||||
);
|
||||
end
|
||||
|
||||
VX_scope #(
|
||||
.DATAW (SCOPE_DATAW),
|
||||
.BUSW (64),
|
||||
.SIZE (4096),
|
||||
.UPDW ($bits({`SCOPE_SIGNALS_UPD_LIST}))
|
||||
) scope (
|
||||
.clk (clk),
|
||||
.reset (SoftReset),
|
||||
.start (scope_data_in[0]),
|
||||
.stop (0),
|
||||
.changed (scope_data_in[1]),
|
||||
.data_in (scope_data_in[SCOPE_DATAW+1:2]),
|
||||
.bus_in (csr_scope_cmd),
|
||||
.bus_out (csr_scope_data),
|
||||
.bus_read (csr_scope_read),
|
||||
.bus_write(csr_scope_write)
|
||||
);
|
||||
|
||||
`endif
|
||||
|
||||
// Vortex /////////////////////////////////////////////////////////////////////
|
||||
|
||||
assign cmd_run_done = !vx_busy;
|
||||
@@ -887,7 +810,7 @@ Vortex #() vortex (
|
||||
`SCOPE_SIGNALS_ISTAGE_BIND
|
||||
`SCOPE_SIGNALS_LSU_BIND
|
||||
`SCOPE_SIGNALS_CORE_BIND
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
`SCOPE_SIGNALS_PIPELINE_BIND
|
||||
`SCOPE_SIGNALS_BE_BIND
|
||||
|
||||
@@ -941,4 +864,92 @@ Vortex #() vortex (
|
||||
`UNUSED_PIN (ebreak)
|
||||
);
|
||||
|
||||
// SCOPE //////////////////////////////////////////////////////////////////////
|
||||
|
||||
`ifdef SCOPE
|
||||
|
||||
localparam SCOPE_DATAW = $bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST});
|
||||
localparam SCOPE_SR_DEPTH = 2;
|
||||
|
||||
`SCOPE_ASSIGN(scope_dram_req_valid, vx_dram_req_valid);
|
||||
`SCOPE_ASSIGN(scope_dram_req_addr, {vx_dram_req_addr, 4'b0});
|
||||
`SCOPE_ASSIGN(scope_dram_req_rw, vx_dram_req_rw);
|
||||
`SCOPE_ASSIGN(scope_dram_req_byteen,vx_dram_req_byteen);
|
||||
`SCOPE_ASSIGN(scope_dram_req_data, vx_dram_req_data);
|
||||
`SCOPE_ASSIGN(scope_dram_req_tag, vx_dram_req_tag);
|
||||
`SCOPE_ASSIGN(scope_dram_req_ready, vx_dram_req_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_valid, vx_dram_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_data, vx_dram_rsp_data);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_tag, vx_dram_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_dram_rsp_ready, vx_dram_rsp_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_snp_req_valid, vx_snp_req_valid);
|
||||
`SCOPE_ASSIGN(scope_snp_req_addr, {vx_snp_req_addr, 4'b0});
|
||||
`SCOPE_ASSIGN(scope_snp_req_invalidate, vx_snp_req_invalidate);
|
||||
`SCOPE_ASSIGN(scope_snp_req_tag, vx_snp_req_tag);
|
||||
`SCOPE_ASSIGN(scope_snp_req_ready, vx_snp_req_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_valid, vx_snp_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_tag, vx_snp_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_ready, vx_snp_rsp_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_valid, vx_snp_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_tag, vx_snp_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_snp_rsp_ready, vx_snp_rsp_ready);
|
||||
|
||||
wire scope_changed = (scope_icache_req_valid && scope_icache_req_ready)
|
||||
|| (scope_icache_rsp_valid && scope_icache_rsp_ready)
|
||||
|| ((| scope_dcache_req_valid) && scope_dcache_req_ready)
|
||||
|| ((| scope_dcache_rsp_valid) && scope_dcache_rsp_ready)
|
||||
|| (scope_dram_req_valid && scope_dram_req_ready)
|
||||
|| (scope_dram_rsp_valid && scope_dram_rsp_ready)
|
||||
|| (scope_snp_req_valid && scope_snp_req_ready)
|
||||
|| (scope_snp_rsp_valid && scope_snp_rsp_ready)
|
||||
|| scope_bank_valid_st0
|
||||
|| scope_bank_valid_st1
|
||||
|| scope_bank_valid_st2
|
||||
|| scope_bank_stall_pipe;
|
||||
|
||||
wire scope_start = vx_reset;
|
||||
|
||||
wire [SCOPE_DATAW+1:0] scope_data_in_st[SCOPE_SR_DEPTH-1:0];
|
||||
wire [SCOPE_DATAW+1:0] scope_data_in_ste;
|
||||
assign scope_data_in_st[0] = {`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST, scope_changed, scope_start};
|
||||
assign scope_data_in_ste = scope_data_in_st[SCOPE_SR_DEPTH-1];
|
||||
|
||||
genvar i;
|
||||
for (i = 1; i < SCOPE_SR_DEPTH; i++) begin
|
||||
VX_generic_register #(
|
||||
.N (SCOPE_DATAW+2)
|
||||
) scope_sr (
|
||||
.clk (clk),
|
||||
.reset (SoftReset),
|
||||
.stall (0),
|
||||
.flush (0),
|
||||
.in (scope_data_in_st[i-1]),
|
||||
.out (scope_data_in_st[i])
|
||||
);
|
||||
end
|
||||
|
||||
VX_scope #(
|
||||
.DATAW (SCOPE_DATAW),
|
||||
.BUSW (64),
|
||||
.SIZE (4096),
|
||||
.UPDW ($bits({`SCOPE_SIGNALS_UPD_LIST}))
|
||||
) scope (
|
||||
.clk (clk),
|
||||
.reset (SoftReset),
|
||||
.start (scope_data_in_ste[0]),
|
||||
.stop (0),
|
||||
.changed (scope_data_in_ste[1]),
|
||||
.data_in (scope_data_in_ste[SCOPE_DATAW+1:2]),
|
||||
.bus_in (csr_scope_cmd),
|
||||
.bus_out (csr_scope_data),
|
||||
.bus_read (csr_scope_read),
|
||||
.bus_write(csr_scope_write)
|
||||
);
|
||||
|
||||
`endif
|
||||
|
||||
endmodule
|
||||
@@ -30,11 +30,11 @@ module VX_alu_unit (
|
||||
VX_divide #(
|
||||
.WIDTHN(32),
|
||||
.WIDTHD(32),
|
||||
.SPEED("HIGHEST"),
|
||||
.REP("UNSIGNED"),
|
||||
.PIPELINE(DIV_PIPELINE_LEN)
|
||||
) unsigned_div (
|
||||
.clock(clk),
|
||||
.aclr(1'b0),
|
||||
.clk(clk),
|
||||
.reset(reset),
|
||||
.clken(1'b1), // TODO this could be disabled on inactive instructions
|
||||
.numer(ALU_in1),
|
||||
.denom(ALU_in2),
|
||||
@@ -45,13 +45,11 @@ module VX_alu_unit (
|
||||
VX_divide #(
|
||||
.WIDTHN(32),
|
||||
.WIDTHD(32),
|
||||
.NREP("SIGNED"),
|
||||
.DREP("SIGNED"),
|
||||
.SPEED("HIGHEST"),
|
||||
.REP("SIGNED"),
|
||||
.PIPELINE(DIV_PIPELINE_LEN)
|
||||
) signed_div (
|
||||
.clock(clk),
|
||||
.aclr(1'b0),
|
||||
.clk(clk),
|
||||
.reset(reset),
|
||||
.clken(1'b1), // TODO this could be disabled on inactive instructions
|
||||
.numer(ALU_in1),
|
||||
.denom(ALU_in2),
|
||||
@@ -63,12 +61,11 @@ module VX_alu_unit (
|
||||
.WIDTHA(64),
|
||||
.WIDTHB(64),
|
||||
.WIDTHP(64),
|
||||
.SPEED("HIGHEST"),
|
||||
.FORCE_LE("YES"),
|
||||
.REP("UNSIGNED"),
|
||||
.PIPELINE(MUL_PIPELINE_LEN)
|
||||
) multiplier (
|
||||
.clock(clk),
|
||||
.aclr(1'b0),
|
||||
.clk(clk),
|
||||
.reset(reset),
|
||||
.clken(1'b1), // TODO this could be disabled on inactive instructions
|
||||
.dataa(mul_data_a),
|
||||
.datab(mul_data_b),
|
||||
|
||||
@@ -133,14 +133,16 @@ module VX_back_end #(
|
||||
|
||||
`SCOPE_ASSIGN(scope_execute_valid, exec_unit_req_if.valid);
|
||||
`SCOPE_ASSIGN(scope_execute_warp_num, exec_unit_req_if.warp_num);
|
||||
`SCOPE_ASSIGN(scope_execute_curr_PC, exec_unit_req_if.curr_PC);
|
||||
`SCOPE_ASSIGN(scope_execute_rd, exec_unit_req_if.rd);
|
||||
`SCOPE_ASSIGN(scope_execute_a, exec_unit_req_if.a_reg_data[0]);
|
||||
`SCOPE_ASSIGN(scope_execute_b, exec_unit_req_if.b_reg_data[0]);
|
||||
`SCOPE_ASSIGN(scope_execute_a, exec_unit_req_if.a_reg_data);
|
||||
`SCOPE_ASSIGN(scope_execute_b, exec_unit_req_if.b_reg_data);
|
||||
|
||||
`SCOPE_ASSIGN(scope_writeback_valid, writeback_if.valid);
|
||||
`SCOPE_ASSIGN(scope_writeback_wb, writeback_if.wb);
|
||||
`SCOPE_ASSIGN(scope_writeback_warp_num, writeback_if.warp_num);
|
||||
`SCOPE_ASSIGN(scope_writeback_curr_PC, writeback_if.curr_PC);
|
||||
`SCOPE_ASSIGN(scope_writeback_wb, writeback_if.wb);
|
||||
`SCOPE_ASSIGN(scope_writeback_rd, writeback_if.rd);
|
||||
`SCOPE_ASSIGN(scope_writeback_data, writeback_if.data[0]);
|
||||
`SCOPE_ASSIGN(scope_writeback_data, writeback_if.data);
|
||||
|
||||
endmodule
|
||||
@@ -6,7 +6,7 @@ module VX_cluster #(
|
||||
`SCOPE_SIGNALS_ISTAGE_IO
|
||||
`SCOPE_SIGNALS_LSU_IO
|
||||
`SCOPE_SIGNALS_CORE_IO
|
||||
`SCOPE_SIGNALS_ICACHE_IO
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
`SCOPE_SIGNALS_PIPELINE_IO
|
||||
`SCOPE_SIGNALS_BE_IO
|
||||
|
||||
@@ -120,7 +120,7 @@ module VX_cluster #(
|
||||
`SCOPE_SIGNALS_ISTAGE_BIND
|
||||
`SCOPE_SIGNALS_LSU_BIND
|
||||
`SCOPE_SIGNALS_CORE_BIND
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
`SCOPE_SIGNALS_PIPELINE_BIND
|
||||
`SCOPE_SIGNALS_BE_BIND
|
||||
|
||||
@@ -319,7 +319,9 @@ module VX_cluster #(
|
||||
.NUM_SNP_REQUESTS (`NUM_CORES),
|
||||
.SNP_REQ_TAG_WIDTH (`L2SNP_TAG_WIDTH),
|
||||
.SNP_FWD_TAG_WIDTH (`DSNP_TAG_WIDTH)
|
||||
) gpu_l2cache (
|
||||
) l2cache (
|
||||
`SCOPE_SIGNALS_CACHE_UNBIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
|
||||
@@ -36,7 +36,7 @@
|
||||
`endif
|
||||
|
||||
`ifndef SHARED_MEM_BASE_ADDR
|
||||
`define SHARED_MEM_BASE_ADDR 32'hFE000000
|
||||
`define SHARED_MEM_BASE_ADDR 32'h6FFFF000
|
||||
`endif
|
||||
|
||||
`ifndef STACK_BASE_ADDR
|
||||
|
||||
@@ -6,7 +6,7 @@ module VX_core #(
|
||||
`SCOPE_SIGNALS_ISTAGE_IO
|
||||
`SCOPE_SIGNALS_LSU_IO
|
||||
`SCOPE_SIGNALS_CORE_IO
|
||||
`SCOPE_SIGNALS_ICACHE_IO
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
`SCOPE_SIGNALS_PIPELINE_IO
|
||||
`SCOPE_SIGNALS_BE_IO
|
||||
|
||||
@@ -237,7 +237,7 @@ module VX_core #(
|
||||
VX_mem_unit #(
|
||||
.CORE_ID(CORE_ID)
|
||||
) mem_unit (
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -43,6 +43,6 @@ module VX_dcache_arb (
|
||||
assign out_core_rsp_if.core_rsp_data = rsp_select0 ? in0_core_rsp_if.core_rsp_data : in1_core_rsp_if.core_rsp_data;
|
||||
assign out_core_rsp_if.core_rsp_tag = rsp_select0 ? in0_core_rsp_if.core_rsp_tag : in1_core_rsp_if.core_rsp_tag;
|
||||
assign in0_core_rsp_if.core_rsp_ready = out_core_rsp_if.core_rsp_ready && rsp_select0;
|
||||
assign in1_core_rsp_if.core_rsp_ready = out_core_rsp_if.core_rsp_ready && ~rsp_select0;
|
||||
assign in1_core_rsp_if.core_rsp_ready = out_core_rsp_if.core_rsp_ready && !rsp_select0;
|
||||
|
||||
endmodule
|
||||
@@ -56,14 +56,14 @@ module VX_icache_stage #(
|
||||
end
|
||||
|
||||
// Icache Request
|
||||
assign icache_req_if.core_req_valid = valid_inst && ~mrq_full;
|
||||
assign icache_req_if.core_req_valid = valid_inst && !mrq_full;
|
||||
assign icache_req_if.core_req_rw = 0;
|
||||
assign icache_req_if.core_req_byteen = 4'b1111;
|
||||
assign icache_req_if.core_req_addr = fe_inst_meta_fi.inst_pc[31:2];
|
||||
assign icache_req_if.core_req_data = 0;
|
||||
|
||||
// Can't accept new request
|
||||
assign icache_stage_delay = mrq_full || ~icache_req_if.core_req_ready;
|
||||
assign icache_stage_delay = mrq_full || !icache_req_if.core_req_ready;
|
||||
|
||||
`ifdef DBG_CORE_REQ_INFO
|
||||
assign icache_req_if.core_req_tag = {fe_inst_meta_fi.inst_pc, 2'b1, 5'b0, fe_inst_meta_fi.warp_num, mrq_write_addr};
|
||||
@@ -78,7 +78,7 @@ module VX_icache_stage #(
|
||||
assign icache_stage_wid = fe_inst_meta_id.warp_num;
|
||||
|
||||
// Can't accept new response
|
||||
assign icache_rsp_if.core_rsp_ready = ~total_freeze;
|
||||
assign icache_rsp_if.core_rsp_ready = !total_freeze;
|
||||
|
||||
`SCOPE_ASSIGN(scope_icache_req_valid, icache_req_if.core_req_valid);
|
||||
`SCOPE_ASSIGN(scope_icache_req_warp_num, fe_inst_meta_fi.warp_num);
|
||||
|
||||
@@ -132,7 +132,7 @@ module VX_lsu_unit #(
|
||||
`endif
|
||||
|
||||
// Can't accept new request
|
||||
assign delay = mrq_full || ~dcache_req_if.core_req_ready;
|
||||
assign delay = mrq_full || !dcache_req_if.core_req_ready;
|
||||
|
||||
// Core Response
|
||||
|
||||
@@ -156,20 +156,20 @@ module VX_lsu_unit #(
|
||||
assign mem_wb_if.data = core_rsp_data;
|
||||
|
||||
// Can't accept new response
|
||||
assign dcache_rsp_if.core_rsp_ready = ~no_slot_mem;
|
||||
assign dcache_rsp_if.core_rsp_ready = !no_slot_mem;
|
||||
|
||||
`SCOPE_ASSIGN(scope_dcache_req_valid, dcache_req_if.core_req_valid);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_warp_num, use_warp_num);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_curr_PC, use_pc);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_addr, {dcache_req_if.core_req_addr[0], 2'b0});
|
||||
`SCOPE_ASSIGN(scope_dcache_req_addr, use_address);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_rw, core_req_rw);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_byteen,dcache_req_if.core_req_byteen[0]);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_data, dcache_req_if.core_req_data[0]);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_byteen,dcache_req_if.core_req_byteen);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_data, dcache_req_if.core_req_data);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_tag, dcache_req_if.core_req_tag);
|
||||
`SCOPE_ASSIGN(scope_dcache_req_ready, dcache_req_if.core_req_ready);
|
||||
|
||||
`SCOPE_ASSIGN(scope_dcache_rsp_valid, dcache_rsp_if.core_rsp_valid);
|
||||
`SCOPE_ASSIGN(scope_dcache_rsp_data, dcache_rsp_if.core_rsp_data[0]);
|
||||
`SCOPE_ASSIGN(scope_dcache_rsp_data, dcache_rsp_if.core_rsp_data);
|
||||
`SCOPE_ASSIGN(scope_dcache_rsp_tag, dcache_rsp_if.core_rsp_tag);
|
||||
`SCOPE_ASSIGN(scope_dcache_rsp_ready, dcache_rsp_if.core_rsp_ready);
|
||||
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
module VX_mem_unit # (
|
||||
parameter CORE_ID = 0
|
||||
) (
|
||||
`SCOPE_SIGNALS_ICACHE_IO
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
@@ -78,6 +78,8 @@ module VX_mem_unit # (
|
||||
.CORE_TAG_ID_BITS (`DCORE_TAG_ID_BITS),
|
||||
.DRAM_TAG_WIDTH (`SDRAM_TAG_WIDTH)
|
||||
) smem (
|
||||
`SCOPE_SIGNALS_CACHE_UNBIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -161,6 +163,8 @@ module VX_mem_unit # (
|
||||
.DRAM_TAG_WIDTH (`DDRAM_TAG_WIDTH),
|
||||
.SNP_REQ_TAG_WIDTH (`DSNP_TAG_WIDTH)
|
||||
) dcache (
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -243,7 +247,7 @@ module VX_mem_unit # (
|
||||
.CORE_TAG_ID_BITS (`DCORE_TAG_ID_BITS),
|
||||
.DRAM_TAG_WIDTH (`IDRAM_TAG_WIDTH)
|
||||
) icache (
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_UNBIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -55,7 +55,7 @@ module VX_scheduler (
|
||||
|
||||
wire [`NUM_THREADS-1:0] valid_wb_new_mask = rename_table[writeback_if.warp_num][writeback_if.rd] & ~writeback_if.valid;
|
||||
|
||||
reg [CTVW-1:0] count_valid_next = (acquire_rd && ~(release_rd && (0 == valid_wb_new_mask))) ? (count_valid + 1) :
|
||||
reg [CTVW-1:0] count_valid_next = (acquire_rd && !(release_rd && (0 == valid_wb_new_mask))) ? (count_valid + 1) :
|
||||
(~acquire_rd && (release_rd && (0 == valid_wb_new_mask))) ? (count_valid - 1) :
|
||||
count_valid;
|
||||
|
||||
|
||||
@@ -35,13 +35,24 @@
|
||||
scope_decode_rs1, \
|
||||
scope_decode_rs2, \
|
||||
scope_execute_warp_num, \
|
||||
scope_execute_curr_PC, \
|
||||
scope_execute_rd, \
|
||||
scope_execute_a, \
|
||||
scope_execute_b, \
|
||||
scope_writeback_warp_num, \
|
||||
scope_writeback_curr_PC, \
|
||||
scope_writeback_wb, \
|
||||
scope_writeback_rd, \
|
||||
scope_writeback_data,
|
||||
scope_writeback_data, \
|
||||
scope_bank_addr_st0, \
|
||||
scope_bank_addr_st1, \
|
||||
scope_bank_addr_st2, \
|
||||
scope_bank_is_mrvq_st1, \
|
||||
scope_bank_miss_st1, \
|
||||
scope_bank_dirty_st1, \
|
||||
scope_bank_tag_valid_st1, \
|
||||
scope_bank_tag_match_st1, \
|
||||
scope_bank_force_miss_st1,
|
||||
|
||||
|
||||
`define SCOPE_SIGNALS_UPD_LIST \
|
||||
@@ -68,18 +79,22 @@
|
||||
scope_memory_delay, \
|
||||
scope_exec_delay, \
|
||||
scope_gpr_stage_delay, \
|
||||
scope_busy
|
||||
scope_busy, \
|
||||
scope_bank_valid_st0, \
|
||||
scope_bank_valid_st1, \
|
||||
scope_bank_valid_st2, \
|
||||
scope_bank_stall_pipe
|
||||
|
||||
`define SCOPE_SIGNALS_DECL \
|
||||
wire scope_dram_req_valid; \
|
||||
wire [31:0] scope_dram_req_addr; \
|
||||
wire scope_dram_req_rw; \
|
||||
wire [15:0] scope_dram_req_byteen; \
|
||||
wire [31:0] scope_dram_req_data; \
|
||||
wire [127:0] scope_dram_req_data; \
|
||||
wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_req_tag; \
|
||||
wire scope_dram_req_ready; \
|
||||
wire scope_dram_rsp_valid; \
|
||||
wire [31:0] scope_dram_rsp_data; \
|
||||
wire [127:0] scope_dram_rsp_data; \
|
||||
wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_rsp_tag; \
|
||||
wire scope_dram_rsp_ready; \
|
||||
wire scope_snp_req_valid; \
|
||||
@@ -101,14 +116,14 @@
|
||||
wire [`NUM_THREADS-1:0] scope_dcache_req_valid; \
|
||||
wire [`NW_BITS-1:0] scope_dcache_req_warp_num; \
|
||||
wire [31:0] scope_dcache_req_curr_PC; \
|
||||
wire [31:0] scope_dcache_req_addr; \
|
||||
wire [63:0] scope_dcache_req_addr; \
|
||||
wire scope_dcache_req_rw; \
|
||||
wire [3:0] scope_dcache_req_byteen; \
|
||||
wire [31:0] scope_dcache_req_data; \
|
||||
wire [7:0] scope_dcache_req_byteen; \
|
||||
wire [63:0] scope_dcache_req_data; \
|
||||
wire [`DCORE_TAG_WIDTH-1:0] scope_dcache_req_tag; \
|
||||
wire scope_dcache_req_ready; \
|
||||
wire [`NUM_THREADS-1:0] scope_dcache_rsp_valid; \
|
||||
wire [31:0] scope_dcache_rsp_data; \
|
||||
wire [63:0] scope_dcache_rsp_data; \
|
||||
wire [`DCORE_TAG_WIDTH-1:0] scope_dcache_rsp_tag; \
|
||||
wire scope_dcache_rsp_ready; \
|
||||
wire scope_busy; \
|
||||
@@ -125,14 +140,29 @@
|
||||
wire [4:0] scope_decode_rs2; \
|
||||
wire [`NUM_THREADS-1:0] scope_execute_valid; \
|
||||
wire [`NW_BITS-1:0] scope_execute_warp_num; \
|
||||
wire [31:0] scope_execute_curr_PC; \
|
||||
wire [4:0] scope_execute_rd; \
|
||||
wire [31:0] scope_execute_a; \
|
||||
wire [31:0] scope_execute_b; \
|
||||
wire [63:0] scope_execute_a; \
|
||||
wire [63:0] scope_execute_b; \
|
||||
wire [`NUM_THREADS-1:0] scope_writeback_valid; \
|
||||
wire [`NW_BITS-1:0] scope_writeback_warp_num; \
|
||||
wire [31:0] scope_writeback_curr_PC; \
|
||||
wire [1:0] scope_writeback_wb; \
|
||||
wire [4:0] scope_writeback_rd; \
|
||||
wire [31:0] scope_writeback_data;
|
||||
wire [63:0] scope_writeback_data; \
|
||||
wire scope_bank_valid_st0; \
|
||||
wire scope_bank_valid_st1; \
|
||||
wire scope_bank_valid_st2; \
|
||||
wire [31:0] scope_bank_addr_st0; \
|
||||
wire [31:0] scope_bank_addr_st1; \
|
||||
wire [31:0] scope_bank_addr_st2; \
|
||||
wire scope_bank_is_mrvq_st1; \
|
||||
wire scope_bank_miss_st1; \
|
||||
wire scope_bank_dirty_st1; \
|
||||
wire scope_bank_tag_valid_st1; \
|
||||
wire scope_bank_tag_match_st1; \
|
||||
wire scope_bank_force_miss_st1; \
|
||||
wire scope_bank_stall_pipe;
|
||||
|
||||
`define SCOPE_SIGNALS_ISTAGE_IO \
|
||||
output wire scope_icache_req_valid, \
|
||||
@@ -149,20 +179,33 @@
|
||||
output wire [`NUM_THREADS-1:0] scope_dcache_req_valid, \
|
||||
output wire [`NW_BITS-1:0] scope_dcache_req_warp_num, \
|
||||
output wire [31:0] scope_dcache_req_curr_PC, \
|
||||
output wire [31:0] scope_dcache_req_addr, \
|
||||
output wire [63:0] scope_dcache_req_addr, \
|
||||
output wire scope_dcache_req_rw, \
|
||||
output wire [3:0] scope_dcache_req_byteen, \
|
||||
output wire [31:0] scope_dcache_req_data, \
|
||||
output wire [7:0] scope_dcache_req_byteen, \
|
||||
output wire [63:0] scope_dcache_req_data, \
|
||||
output wire [`DCORE_TAG_WIDTH-1:0] scope_dcache_req_tag, \
|
||||
output wire scope_dcache_req_ready, \
|
||||
output wire [`NUM_THREADS-1:0] scope_dcache_rsp_valid, \
|
||||
output wire [31:0] scope_dcache_rsp_data, \
|
||||
output wire [63:0] scope_dcache_rsp_data, \
|
||||
output wire [`DCORE_TAG_WIDTH-1:0] scope_dcache_rsp_tag, \
|
||||
output wire scope_dcache_rsp_ready,
|
||||
|
||||
`define SCOPE_SIGNALS_CORE_IO \
|
||||
|
||||
`define SCOPE_SIGNALS_ICACHE_IO \
|
||||
`define SCOPE_SIGNALS_CACHE_IO \
|
||||
output wire scope_bank_valid_st0, \
|
||||
output wire scope_bank_valid_st1, \
|
||||
output wire scope_bank_valid_st2, \
|
||||
output wire [31:0] scope_bank_addr_st0, \
|
||||
output wire [31:0] scope_bank_addr_st1, \
|
||||
output wire [31:0] scope_bank_addr_st2, \
|
||||
output wire scope_bank_is_mrvq_st1, \
|
||||
output wire scope_bank_miss_st1, \
|
||||
output wire scope_bank_dirty_st1, \
|
||||
output wire scope_bank_tag_valid_st1, \
|
||||
output wire scope_bank_tag_match_st1, \
|
||||
output wire scope_bank_force_miss_st1, \
|
||||
output wire scope_bank_stall_pipe,
|
||||
|
||||
`define SCOPE_SIGNALS_PIPELINE_IO \
|
||||
output wire scope_busy, \
|
||||
@@ -172,22 +215,24 @@
|
||||
output wire scope_gpr_stage_delay,
|
||||
|
||||
`define SCOPE_SIGNALS_BE_IO \
|
||||
output wire [`NUM_THREADS-1:0] scope_decode_valid, \
|
||||
output wire [`NW_BITS-1:0] scope_decode_warp_num, \
|
||||
output wire [`NUM_THREADS-1:0] scope_decode_valid, \
|
||||
output wire [`NW_BITS-1:0] scope_decode_warp_num, \
|
||||
output wire [31:0] scope_decode_curr_PC, \
|
||||
output wire scope_decode_is_jal, \
|
||||
output wire [4:0] scope_decode_rs1, \
|
||||
output wire [4:0] scope_decode_rs2, \
|
||||
output wire [`NUM_THREADS-1:0] scope_execute_valid, \
|
||||
output wire [`NW_BITS-1:0] scope_execute_warp_num, \
|
||||
output wire [31:0] scope_execute_curr_PC, \
|
||||
output wire [4:0] scope_execute_rd, \
|
||||
output wire [31:0] scope_execute_a, \
|
||||
output wire [31:0] scope_execute_b, \
|
||||
output wire [63:0] scope_execute_a, \
|
||||
output wire [63:0] scope_execute_b, \
|
||||
output wire [`NUM_THREADS-1:0] scope_writeback_valid, \
|
||||
output wire [`NW_BITS-1:0] scope_writeback_warp_num, \
|
||||
output wire [31:0] scope_writeback_curr_PC, \
|
||||
output wire [1:0] scope_writeback_wb, \
|
||||
output wire [4:0] scope_writeback_rd, \
|
||||
output wire [31:0] scope_writeback_data,
|
||||
output wire [63:0] scope_writeback_data,
|
||||
|
||||
`define SCOPE_SIGNALS_ISTAGE_BIND \
|
||||
.scope_icache_req_valid (scope_icache_req_valid), \
|
||||
@@ -217,7 +262,82 @@
|
||||
|
||||
`define SCOPE_SIGNALS_CORE_BIND \
|
||||
|
||||
`define SCOPE_SIGNALS_ICACHE_BIND \
|
||||
`define SCOPE_SIGNALS_CACHE_BIND \
|
||||
.scope_bank_valid_st0 (scope_bank_valid_st0), \
|
||||
.scope_bank_valid_st1 (scope_bank_valid_st1), \
|
||||
.scope_bank_valid_st2 (scope_bank_valid_st2), \
|
||||
.scope_bank_addr_st0 (scope_bank_addr_st0), \
|
||||
.scope_bank_addr_st1 (scope_bank_addr_st1), \
|
||||
.scope_bank_addr_st2 (scope_bank_addr_st2), \
|
||||
.scope_bank_is_mrvq_st1 (scope_bank_is_mrvq_st1), \
|
||||
.scope_bank_miss_st1 (scope_bank_miss_st1), \
|
||||
.scope_bank_dirty_st1 (scope_bank_dirty_st1), \
|
||||
.scope_bank_tag_valid_st1 (scope_bank_tag_valid_st1), \
|
||||
.scope_bank_tag_match_st1 (scope_bank_tag_match_st1), \
|
||||
.scope_bank_force_miss_st1 (scope_bank_force_miss_st1), \
|
||||
.scope_bank_stall_pipe (scope_bank_stall_pipe),
|
||||
|
||||
`define SCOPE_SIGNALS_CACHE_UNBIND \
|
||||
/* verilator lint_off PINCONNECTEMPTY */ \
|
||||
.scope_bank_valid_st0 (), \
|
||||
.scope_bank_valid_st1 (), \
|
||||
.scope_bank_valid_st2 (), \
|
||||
.scope_bank_addr_st0 (), \
|
||||
.scope_bank_addr_st1 (), \
|
||||
.scope_bank_addr_st2 (), \
|
||||
.scope_bank_is_mrvq_st1 (), \
|
||||
.scope_bank_miss_st1 (), \
|
||||
.scope_bank_dirty_st1 (), \
|
||||
.scope_bank_tag_valid_st1 (), \
|
||||
.scope_bank_tag_match_st1 (), \
|
||||
.scope_bank_force_miss_st1 (), \
|
||||
.scope_bank_stall_pipe (), \
|
||||
/* verilator lint_on PINCONNECTEMPTY */
|
||||
|
||||
`define SCOPE_SIGNALS_CACHE_BANK_SELECT \
|
||||
/* verilator lint_off UNUSED */ \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_valid_st0; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_valid_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_valid_st2; \
|
||||
wire [NUM_BANKS-1:0][31:0] scope_per_bank_addr_st0; \
|
||||
wire [NUM_BANKS-1:0][31:0] scope_per_bank_addr_st1; \
|
||||
wire [NUM_BANKS-1:0][31:0] scope_per_bank_addr_st2; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_is_mrvq_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_miss_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_dirty_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_tag_valid_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_tag_match_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_force_miss_st1; \
|
||||
wire [NUM_BANKS-1:0] scope_per_bank_stall_pipe; \
|
||||
/* verilator lint_on UNUSED */ \
|
||||
assign scope_bank_valid_st0 = scope_per_bank_valid_st0[0]; \
|
||||
assign scope_bank_valid_st1 = scope_per_bank_valid_st1[0]; \
|
||||
assign scope_bank_valid_st2 = scope_per_bank_valid_st2[0]; \
|
||||
assign scope_bank_addr_st0 = scope_per_bank_addr_st0[0]; \
|
||||
assign scope_bank_addr_st1 = scope_per_bank_addr_st1[0]; \
|
||||
assign scope_bank_addr_st2 = scope_per_bank_addr_st2[0]; \
|
||||
assign scope_bank_is_mrvq_st1 = scope_per_bank_is_mrvq_st1[0]; \
|
||||
assign scope_bank_miss_st1 = scope_per_bank_miss_st1[0]; \
|
||||
assign scope_bank_dirty_st1 = scope_per_bank_dirty_st1[0]; \
|
||||
assign scope_bank_tag_valid_st1 = scope_per_bank_tag_valid_st1[0]; \
|
||||
assign scope_bank_tag_match_st1 = scope_per_bank_tag_match_st1[0]; \
|
||||
assign scope_bank_force_miss_st1 = scope_per_bank_force_miss_st1[0]; \
|
||||
assign scope_bank_stall_pipe = scope_per_bank_stall_pipe[0];
|
||||
|
||||
`define SCOPE_SIGNALS_CACHE_BANK_BIND \
|
||||
.scope_bank_valid_st0 (scope_per_bank_valid_st0[i]), \
|
||||
.scope_bank_valid_st1 (scope_per_bank_valid_st1[i]), \
|
||||
.scope_bank_valid_st2 (scope_per_bank_valid_st2[i]), \
|
||||
.scope_bank_addr_st0 (scope_per_bank_addr_st0[i]), \
|
||||
.scope_bank_addr_st1 (scope_per_bank_addr_st1[i]), \
|
||||
.scope_bank_addr_st2 (scope_per_bank_addr_st2[i]), \
|
||||
.scope_bank_is_mrvq_st1 (scope_per_bank_is_mrvq_st1[i]), \
|
||||
.scope_bank_miss_st1 (scope_per_bank_miss_st1[i]), \
|
||||
.scope_bank_dirty_st1 (scope_per_bank_dirty_st1[i]), \
|
||||
.scope_bank_tag_valid_st1 (scope_per_bank_tag_valid_st1[i]), \
|
||||
.scope_bank_tag_match_st1 (scope_per_bank_tag_match_st1[i]), \
|
||||
.scope_bank_force_miss_st1 (scope_per_bank_force_miss_st1[i]), \
|
||||
.scope_bank_stall_pipe (scope_per_bank_stall_pipe[i]),
|
||||
|
||||
`define SCOPE_SIGNALS_PIPELINE_BIND \
|
||||
.scope_busy (scope_busy), \
|
||||
@@ -235,11 +355,13 @@
|
||||
.scope_decode_rs2 (scope_decode_rs2), \
|
||||
.scope_execute_valid (scope_execute_valid), \
|
||||
.scope_execute_warp_num (scope_execute_warp_num), \
|
||||
.scope_execute_curr_PC (scope_execute_curr_PC), \
|
||||
.scope_execute_rd (scope_execute_rd), \
|
||||
.scope_execute_a (scope_execute_a), \
|
||||
.scope_execute_b (scope_execute_b), \
|
||||
.scope_writeback_valid (scope_writeback_valid), \
|
||||
.scope_writeback_warp_num (scope_writeback_warp_num), \
|
||||
.scope_writeback_curr_PC(scope_writeback_curr_PC), \
|
||||
.scope_writeback_wb (scope_writeback_wb), \
|
||||
.scope_writeback_rd (scope_writeback_rd), \
|
||||
.scope_writeback_data (scope_writeback_data),
|
||||
@@ -249,17 +371,21 @@
|
||||
`define SCOPE_SIGNALS_ISTAGE_IO
|
||||
`define SCOPE_SIGNALS_LSU_IO
|
||||
`define SCOPE_SIGNALS_CORE_IO
|
||||
`define SCOPE_SIGNALS_ICACHE_IO
|
||||
`define SCOPE_SIGNALS_CACHE_IO
|
||||
`define SCOPE_SIGNALS_PIPELINE_IO
|
||||
`define SCOPE_SIGNALS_BE_IO
|
||||
|
||||
`define SCOPE_SIGNALS_ISTAGE_BIND
|
||||
`define SCOPE_SIGNALS_LSU_BIND
|
||||
`define SCOPE_SIGNALS_CORE_BIND
|
||||
`define SCOPE_SIGNALS_ICACHE_BIND
|
||||
`define SCOPE_SIGNALS_CACHE_BIND
|
||||
`define SCOPE_SIGNALS_PIPELINE_BIND
|
||||
`define SCOPE_SIGNALS_BE_BIND
|
||||
|
||||
`define SCOPE_SIGNALS_CACHE_UNBIND
|
||||
`define SCOPE_SIGNALS_CACHE_BANK_SELECT
|
||||
`define SCOPE_SIGNALS_CACHE_BANK_BIND
|
||||
|
||||
`define SCOPE_ASSIGN(d,s)
|
||||
`endif
|
||||
|
||||
|
||||
@@ -4,7 +4,7 @@ module Vortex (
|
||||
`SCOPE_SIGNALS_ISTAGE_IO
|
||||
`SCOPE_SIGNALS_LSU_IO
|
||||
`SCOPE_SIGNALS_CORE_IO
|
||||
`SCOPE_SIGNALS_ICACHE_IO
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
`SCOPE_SIGNALS_PIPELINE_IO
|
||||
`SCOPE_SIGNALS_BE_IO
|
||||
|
||||
@@ -66,7 +66,7 @@ module Vortex (
|
||||
`SCOPE_SIGNALS_ISTAGE_BIND
|
||||
`SCOPE_SIGNALS_LSU_BIND
|
||||
`SCOPE_SIGNALS_CORE_BIND
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
`SCOPE_SIGNALS_PIPELINE_BIND
|
||||
`SCOPE_SIGNALS_BE_BIND
|
||||
|
||||
@@ -162,7 +162,7 @@ module Vortex (
|
||||
`SCOPE_SIGNALS_ISTAGE_BIND
|
||||
`SCOPE_SIGNALS_LSU_BIND
|
||||
`SCOPE_SIGNALS_CORE_BIND
|
||||
`SCOPE_SIGNALS_ICACHE_BIND
|
||||
`SCOPE_SIGNALS_CACHE_BIND
|
||||
`SCOPE_SIGNALS_PIPELINE_BIND
|
||||
`SCOPE_SIGNALS_BE_BIND
|
||||
|
||||
@@ -332,7 +332,9 @@ module Vortex (
|
||||
.NUM_SNP_REQUESTS (`NUM_CLUSTERS),
|
||||
.SNP_REQ_TAG_WIDTH (`L3SNP_TAG_WIDTH),
|
||||
.SNP_FWD_TAG_WIDTH (`L2SNP_TAG_WIDTH)
|
||||
) gpu_l3cache (
|
||||
) l3cache (
|
||||
`SCOPE_SIGNALS_CACHE_UNBIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
|
||||
96
hw/rtl/cache/VX_bank.v
vendored
96
hw/rtl/cache/VX_bank.v
vendored
@@ -52,6 +52,8 @@ module VX_bank #(
|
||||
// Snooping request tag width
|
||||
parameter SNP_REQ_TAG_WIDTH = 0
|
||||
) (
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
@@ -153,7 +155,7 @@ module VX_bank #(
|
||||
`UNUSED_PIN (size)
|
||||
);
|
||||
|
||||
assign snp_req_ready = ~snrq_full;
|
||||
assign snp_req_ready = !snrq_full;
|
||||
|
||||
wire dfpq_pop;
|
||||
wire dfpq_empty;
|
||||
@@ -223,7 +225,7 @@ module VX_bank #(
|
||||
.reqq_full (reqq_full)
|
||||
);
|
||||
|
||||
assign core_req_ready = ~reqq_full;
|
||||
assign core_req_ready = !reqq_full;
|
||||
assign reqq_push = (| core_req_valid) && core_req_ready;
|
||||
|
||||
wire mrvq_pop;
|
||||
@@ -291,7 +293,7 @@ module VX_bank #(
|
||||
wire qual_valid_st0;
|
||||
wire [`LINE_ADDR_WIDTH-1:0] qual_addr_st0;
|
||||
wire [`UP(`WORD_SELECT_WIDTH)-1:0] qual_wsel_st0;
|
||||
wire qual_from_mrvq_st0;
|
||||
wire qual_is_mrvq_st0;
|
||||
|
||||
wire [`WORD_WIDTH-1:0] qual_writeword_st0;
|
||||
wire [`BANK_LINE_WIDTH-1:0] qual_writedata_st0;
|
||||
@@ -308,7 +310,7 @@ module VX_bank #(
|
||||
wire [`BANK_LINE_WIDTH-1:0] writedata_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire is_snp_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire snp_invalidate_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire from_mrvq_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire is_mrvq_st1 [STAGE_1_CYCLES-1:0];
|
||||
|
||||
assign qual_is_fill_st0 = dfpq_pop_unqual;
|
||||
|
||||
@@ -352,7 +354,7 @@ module VX_bank #(
|
||||
reqq_pop_unqual ? reqq_req_writeword_st0 :
|
||||
0;
|
||||
|
||||
assign qual_from_mrvq_st0 = mrvq_pop_unqual;
|
||||
assign qual_is_mrvq_st0 = mrvq_pop_unqual;
|
||||
|
||||
`ifdef DBG_CORE_REQ_INFO
|
||||
if (WORD_SIZE != `GLOBAL_BLOCK_SIZE) begin
|
||||
@@ -367,8 +369,8 @@ module VX_bank #(
|
||||
.reset (reset),
|
||||
.stall (stall_bank_pipe),
|
||||
.flush (1'b0),
|
||||
.in ({qual_from_mrvq_st0, qual_is_snp_st0, qual_snp_invalidate_st0, qual_going_to_write_st0, qual_valid_st0, qual_addr_st0, qual_wsel_st0, qual_writeword_st0, qual_inst_meta_st0, qual_is_fill_st0, qual_writedata_st0}),
|
||||
.out ({from_mrvq_st1[0] , is_snp_st1[0], snp_invalidate_st1[0], going_to_write_st1[0], valid_st1[0], addr_st1[0], wsel_st1[0], writeword_st1[0], inst_meta_st1[0], is_fill_st1[0], writedata_st1[0]})
|
||||
.in ({qual_is_mrvq_st0, qual_is_snp_st0, qual_snp_invalidate_st0, qual_going_to_write_st0, qual_valid_st0, qual_addr_st0, qual_wsel_st0, qual_writeword_st0, qual_inst_meta_st0, qual_is_fill_st0, qual_writedata_st0}),
|
||||
.out ({is_mrvq_st1[0] , is_snp_st1[0], snp_invalidate_st1[0], going_to_write_st1[0], valid_st1[0], addr_st1[0], wsel_st1[0], writeword_st1[0], inst_meta_st1[0], is_fill_st1[0], writedata_st1[0]})
|
||||
);
|
||||
|
||||
genvar i;
|
||||
@@ -380,8 +382,8 @@ module VX_bank #(
|
||||
.reset (reset),
|
||||
.stall (stall_bank_pipe),
|
||||
.flush (1'b0),
|
||||
.in ({from_mrvq_st1[i-1], is_snp_st1[i-1], snp_invalidate_st1[i-1], going_to_write_st1[i-1], valid_st1[i-1], addr_st1[i-1], wsel_st1[i-1], writeword_st1[i-1], inst_meta_st1[i-1], is_fill_st1[i-1], writedata_st1[i-1]}),
|
||||
.out ({from_mrvq_st1[i] , is_snp_st1[i], snp_invalidate_st1[i], going_to_write_st1[i], valid_st1[i], addr_st1[i], wsel_st1[i], writeword_st1[i], inst_meta_st1[i], is_fill_st1[i], writedata_st1[i]})
|
||||
.in ({is_mrvq_st1[i-1], is_snp_st1[i-1], snp_invalidate_st1[i-1], going_to_write_st1[i-1], valid_st1[i-1], addr_st1[i-1], wsel_st1[i-1], writeword_st1[i-1], inst_meta_st1[i-1], is_fill_st1[i-1], writedata_st1[i-1]}),
|
||||
.out ({is_mrvq_st1[i] , is_snp_st1[i], snp_invalidate_st1[i], going_to_write_st1[i], valid_st1[i], addr_st1[i], wsel_st1[i], writeword_st1[i], inst_meta_st1[i], is_fill_st1[i], writedata_st1[i]})
|
||||
);
|
||||
end
|
||||
|
||||
@@ -404,26 +406,31 @@ module VX_bank #(
|
||||
wire mrvq_init_ready_state_st1e;
|
||||
wire miss_add_because_miss;
|
||||
wire valid_st1e;
|
||||
wire from_mrvq_st1e;
|
||||
wire is_mrvq_st1e;
|
||||
wire mrvq_recover_ready_state_st1e;
|
||||
wire[`LINE_ADDR_WIDTH-1:0] addr_st1e;
|
||||
|
||||
assign from_mrvq_st1e = from_mrvq_st1[STAGE_1_CYCLES-1];
|
||||
wire tag_valid_st1e;
|
||||
wire tag_match_st1e;
|
||||
|
||||
assign is_mrvq_st1e = is_mrvq_st1[STAGE_1_CYCLES-1];
|
||||
assign valid_st1e = valid_st1 [STAGE_1_CYCLES-1];
|
||||
assign is_snp_st1e = is_snp_st1 [STAGE_1_CYCLES-1];
|
||||
assign snp_invalidate_st1e = snp_invalidate_st1 [STAGE_1_CYCLES-1];
|
||||
assign addr_st1e = addr_st1[STAGE_1_CYCLES-1];
|
||||
|
||||
assign {tag_st1e, mem_rw_st1e, mem_byteen_st1e, tid_st1e} = inst_meta_st1[STAGE_1_CYCLES-1];
|
||||
|
||||
assign st2_pending_hazard_st1e = (miss_add_because_miss)
|
||||
&& ((addr_st2 == addr_st1[STAGE_1_CYCLES-1]) && !is_fill_st2);
|
||||
&& ((addr_st2 == addr_st1e) && !is_fill_st2);
|
||||
|
||||
assign force_request_miss_st1e = (valid_st1e && !from_mrvq_st1e && (mrvq_pending_hazard_st1e || st2_pending_hazard_st1e))
|
||||
|| (valid_st1e && from_mrvq_st1e && recover_mrvq_state_st2);
|
||||
assign force_request_miss_st1e = (valid_st1e && !is_mrvq_st1e && (mrvq_pending_hazard_st1e || st2_pending_hazard_st1e))
|
||||
|| (valid_st1e && is_mrvq_st1e && recover_mrvq_state_st2);
|
||||
|
||||
assign mrvq_recover_ready_state_st1e = valid_st1e
|
||||
&& from_mrvq_st1e
|
||||
&& is_mrvq_st1e
|
||||
&& recover_mrvq_state_st2
|
||||
&& (addr_st2 == addr_st1[STAGE_1_CYCLES-1]);
|
||||
&& (addr_st2 == addr_st1e);
|
||||
|
||||
VX_tag_data_access #(
|
||||
.CACHE_SIZE (CACHE_SIZE),
|
||||
@@ -447,7 +454,7 @@ module VX_bank #(
|
||||
// Actual Read/Write
|
||||
.valid_req_st1e (valid_st1e),
|
||||
.writefill_st1e (is_fill_st1[STAGE_1_CYCLES-1]),
|
||||
.writeaddr_st1e (addr_st1[STAGE_1_CYCLES-1]),
|
||||
.writeaddr_st1e (addr_st1e),
|
||||
.wordsel_st1e (wsel_st1[STAGE_1_CYCLES-1]),
|
||||
.writeword_st1e (writeword_st1[STAGE_1_CYCLES-1]),
|
||||
.writedata_st1e (writedata_st1[STAGE_1_CYCLES-1]),
|
||||
@@ -467,7 +474,10 @@ module VX_bank #(
|
||||
.dirtyb_st1e (dirtyb_st1e),
|
||||
.fill_saw_dirty_st1e (fill_saw_dirty_st1e),
|
||||
.snp_to_mrvq_st1e (snp_to_mrvq_st1e),
|
||||
.mrvq_init_ready_state_st1e(mrvq_init_ready_state_st1e)
|
||||
.mrvq_init_ready_state_st1e(mrvq_init_ready_state_st1e),
|
||||
|
||||
.tag_valid_st1e (tag_valid_st1e),
|
||||
.tag_match_st1e (tag_match_st1e)
|
||||
);
|
||||
|
||||
`ifdef DBG_CORE_REQ_INFO
|
||||
@@ -476,8 +486,8 @@ module VX_bank #(
|
||||
end
|
||||
`endif
|
||||
|
||||
wire qual_valid_st1e_2 = valid_st1e && !is_fill_st1[STAGE_1_CYCLES-1];
|
||||
wire from_mrvq_st1e_st2 = from_mrvq_st1e;
|
||||
wire qual_valid_st1e_2 = valid_st1e && !is_fill_st1[STAGE_1_CYCLES-1];
|
||||
wire is_mrvq_st1e_st2 = is_mrvq_st1e;
|
||||
|
||||
wire valid_st2;
|
||||
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st2;
|
||||
@@ -493,7 +503,7 @@ module VX_bank #(
|
||||
wire is_snp_st2;
|
||||
wire snp_invalidate_st2;
|
||||
wire snp_to_mrvq_st2;
|
||||
wire from_mrvq_st2;
|
||||
wire is_mrvq_st2;
|
||||
wire mrvq_init_ready_state_st2;
|
||||
wire mrvq_recover_ready_state_st2;
|
||||
wire mrvq_init_ready_state_unqual_st2;
|
||||
@@ -507,8 +517,8 @@ module VX_bank #(
|
||||
.reset (reset),
|
||||
.stall (stall_bank_pipe),
|
||||
.flush (1'b0),
|
||||
.in ({mrvq_recover_ready_state_st1e, from_mrvq_st1e_st2, mrvq_init_ready_state_st1e , snp_to_mrvq_st1e, is_snp_st1e, snp_invalidate_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1] , qual_valid_st1e_2, addr_st1[STAGE_1_CYCLES-1], wsel_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, dirtyb_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
|
||||
.out ({mrvq_recover_ready_state_st2 , from_mrvq_st2 , mrvq_init_ready_state_unqual_st2, snp_to_mrvq_st2 , is_snp_st2 , snp_invalidate_st2, fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , wsel_st2, writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , dirtyb_st2, inst_meta_st2 })
|
||||
.in ({mrvq_recover_ready_state_st1e, is_mrvq_st1e_st2, mrvq_init_ready_state_st1e , snp_to_mrvq_st1e, is_snp_st1e, snp_invalidate_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1] , qual_valid_st1e_2, addr_st1e, wsel_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, dirtyb_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
|
||||
.out ({mrvq_recover_ready_state_st2 , is_mrvq_st2 , mrvq_init_ready_state_unqual_st2, snp_to_mrvq_st2 , is_snp_st2 , snp_invalidate_st2, fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , wsel_st2, writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , dirtyb_st2, inst_meta_st2 })
|
||||
);
|
||||
|
||||
`ifdef DBG_CORE_REQ_INFO
|
||||
@@ -530,7 +540,7 @@ module VX_bank #(
|
||||
|| dwbq_push_stall
|
||||
|| dram_fill_req_stall);
|
||||
|
||||
assign recover_mrvq_state_st2 = miss_add && from_mrvq_st2;
|
||||
assign recover_mrvq_state_st2 = miss_add && is_mrvq_st2;
|
||||
|
||||
wire [`LINE_ADDR_WIDTH-1:0] miss_add_addr = addr_st2;
|
||||
wire [`UP(`WORD_SELECT_WIDTH)-1:0] miss_add_wsel = wsel_st2;
|
||||
@@ -539,10 +549,10 @@ module VX_bank #(
|
||||
wire miss_add_is_snp = is_snp_st2;
|
||||
wire miss_add_snp_invalidate = snp_invalidate_st2;
|
||||
|
||||
wire miss_add_from_mrvq = valid_st2 && from_mrvq_st2 && !stall_bank_pipe;
|
||||
wire miss_add_is_mrvq = valid_st2 && is_mrvq_st2 && !stall_bank_pipe;
|
||||
|
||||
assign mrvq_init_ready_state_hazard_st0_st1 = miss_add_unqual && qual_is_fill_st0 && (miss_add_addr == qual_addr_st0 );
|
||||
assign mrvq_init_ready_state_hazard_st1e_st1 = miss_add_unqual && is_fill_st1[STAGE_1_CYCLES-1] && (miss_add_addr == addr_st1[STAGE_1_CYCLES-1]);
|
||||
assign mrvq_init_ready_state_hazard_st0_st1 = miss_add_unqual && qual_is_fill_st0 && (miss_add_addr == qual_addr_st0);
|
||||
assign mrvq_init_ready_state_hazard_st1e_st1 = miss_add_unqual && is_fill_st1[STAGE_1_CYCLES-1] && (miss_add_addr == addr_st1e);
|
||||
|
||||
assign mrvq_init_ready_state_st2 = mrvq_init_ready_state_unqual_st2
|
||||
|| mrvq_init_ready_state_hazard_st0_st1
|
||||
@@ -564,7 +574,7 @@ module VX_bank #(
|
||||
|
||||
// Enqueue
|
||||
.miss_add (miss_add),
|
||||
.from_mrvq (miss_add_from_mrvq),
|
||||
.is_mrvq (miss_add_is_mrvq),
|
||||
.miss_add_addr (miss_add_addr),
|
||||
.miss_add_wsel (miss_add_wsel),
|
||||
.miss_add_data (miss_add_data),
|
||||
@@ -580,7 +590,7 @@ module VX_bank #(
|
||||
|
||||
// Broadcast
|
||||
.is_fill_st1 (is_fill_st1[STAGE_1_CYCLES-1]),
|
||||
.fill_addr_st1 (addr_st1[STAGE_1_CYCLES-1]),
|
||||
.fill_addr_st1 (addr_st1e),
|
||||
.pending_hazard (mrvq_pending_hazard_st1e),
|
||||
|
||||
// Dequeue
|
||||
@@ -641,7 +651,7 @@ module VX_bank #(
|
||||
|
||||
wire dram_fill_req_unqual = miss_add_unqual
|
||||
&& (!mrvq_init_ready_state_st2
|
||||
|| (from_mrvq_st2 && !mrvq_recover_ready_state_st2));
|
||||
|| (is_mrvq_st2 && !mrvq_recover_ready_state_st2));
|
||||
|
||||
assign dram_fill_req_valid = dram_fill_req_unqual
|
||||
&& !(dwbq_push_stall
|
||||
@@ -649,7 +659,7 @@ module VX_bank #(
|
||||
|| cwbq_push_stall);
|
||||
|
||||
assign dram_fill_req_addr = addr_st2;
|
||||
assign dram_fill_req_stall = dram_fill_req_unqual && ~dram_fill_req_ready;
|
||||
assign dram_fill_req_stall = dram_fill_req_unqual && !dram_fill_req_ready;
|
||||
|
||||
// Enqueue DRAM writeback request
|
||||
|
||||
@@ -706,11 +716,11 @@ module VX_bank #(
|
||||
end
|
||||
|
||||
// when both dwb and snp are asserted, first release the cwb, then release the snp.
|
||||
assign dram_wb_req_valid = ~dwbq_empty && dwbq_is_dwb_out && (~dwbq_is_snp_out || dwbq_dual_valid_sel == 0);
|
||||
assign snp_rsp_valid = ~dwbq_empty && dwbq_is_snp_out && (~dwbq_is_dwb_out || dwbq_dual_valid_sel == 1);
|
||||
assign dram_wb_req_valid = !dwbq_empty && dwbq_is_dwb_out && (~dwbq_is_snp_out || dwbq_dual_valid_sel == 0);
|
||||
assign snp_rsp_valid = !dwbq_empty && dwbq_is_snp_out && (~dwbq_is_dwb_out || dwbq_dual_valid_sel == 1);
|
||||
|
||||
assign dwbq_pop = (dwbq_is_dwb_out && ~dwbq_is_snp_out && dram_wb_req_fire)
|
||||
|| (dwbq_is_snp_out && ~dwbq_is_dwb_out && snp_rsp_fire)
|
||||
assign dwbq_pop = (dwbq_is_dwb_out && !dwbq_is_snp_out && dram_wb_req_fire)
|
||||
|| (dwbq_is_snp_out && !dwbq_is_dwb_out && snp_rsp_fire)
|
||||
|| (dwbq_is_dwb_out && dwbq_is_snp_out && snp_rsp_fire);
|
||||
|
||||
// bank pipeline stall
|
||||
@@ -745,4 +755,20 @@ module VX_bank #(
|
||||
end
|
||||
`endif
|
||||
|
||||
`SCOPE_ASSIGN(scope_bank_valid_st0, qual_valid_st0);
|
||||
`SCOPE_ASSIGN(scope_bank_valid_st1, valid_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_valid_st2, valid_st2);
|
||||
|
||||
`SCOPE_ASSIGN(scope_bank_is_mrvq_st1, is_mrvq_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_miss_st1, miss_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_dirty_st1, dirty_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_tag_valid_st1, tag_valid_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_tag_match_st1, tag_match_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_force_miss_st1, force_request_miss_st1e);
|
||||
`SCOPE_ASSIGN(scope_bank_stall_pipe, stall_bank_pipe);
|
||||
|
||||
`SCOPE_ASSIGN(scope_bank_addr_st0, `LINE_TO_BYTE_ADDR(qual_addr_st0, BANK_ID));
|
||||
`SCOPE_ASSIGN(scope_bank_addr_st1, `LINE_TO_BYTE_ADDR(addr_st1e, BANK_ID));
|
||||
`SCOPE_ASSIGN(scope_bank_addr_st2, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID));
|
||||
|
||||
endmodule : VX_bank
|
||||
58
hw/rtl/cache/VX_cache.v
vendored
58
hw/rtl/cache/VX_cache.v
vendored
@@ -65,7 +65,7 @@ module VX_cache #(
|
||||
// Snooping forward tag width
|
||||
parameter SNP_FWD_TAG_WIDTH = 1
|
||||
) (
|
||||
`SCOPE_SIGNALS_ICACHE_IO
|
||||
`SCOPE_SIGNALS_CACHE_IO
|
||||
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
@@ -167,6 +167,8 @@ module VX_cache #(
|
||||
wire [NUM_BANKS-1:0][SNP_REQ_TAG_WIDTH-1:0] per_bank_snp_rsp_tag;
|
||||
wire [NUM_BANKS-1:0] per_bank_snp_rsp_ready;
|
||||
|
||||
`SCOPE_SIGNALS_CACHE_BANK_SELECT
|
||||
|
||||
wire snp_req_valid_qual;
|
||||
wire [`DRAM_ADDR_WIDTH-1:0] snp_req_addr_qual;
|
||||
wire snp_req_invalidate_qual;
|
||||
@@ -352,28 +354,30 @@ module VX_cache #(
|
||||
assign curr_bank_snp_rsp_ready = per_bank_snp_rsp_ready[i];
|
||||
|
||||
VX_bank #(
|
||||
.BANK_ID (i),
|
||||
.CACHE_ID (CACHE_ID),
|
||||
.CACHE_SIZE (CACHE_SIZE),
|
||||
.BANK_LINE_SIZE (BANK_LINE_SIZE),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE (WORD_SIZE),
|
||||
.NUM_REQUESTS (NUM_REQUESTS),
|
||||
.STAGE_1_CYCLES (STAGE_1_CYCLES),
|
||||
.CREQ_SIZE (CREQ_SIZE),
|
||||
.MRVQ_SIZE (MRVQ_SIZE),
|
||||
.DFPQ_SIZE (DFPQ_SIZE),
|
||||
.SNRQ_SIZE (SNRQ_SIZE),
|
||||
.CWBQ_SIZE (CWBQ_SIZE),
|
||||
.DWBQ_SIZE (DWBQ_SIZE),
|
||||
.DFQQ_SIZE (DFQQ_SIZE),
|
||||
.DRAM_ENABLE (DRAM_ENABLE),
|
||||
.WRITE_ENABLE (WRITE_ENABLE),
|
||||
.SNOOP_FORWARDING (SNOOP_FORWARDING),
|
||||
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
|
||||
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS),
|
||||
.SNP_REQ_TAG_WIDTH (SNP_REQ_TAG_WIDTH)
|
||||
.BANK_ID (i),
|
||||
.CACHE_ID (CACHE_ID),
|
||||
.CACHE_SIZE (CACHE_SIZE),
|
||||
.BANK_LINE_SIZE (BANK_LINE_SIZE),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE (WORD_SIZE),
|
||||
.NUM_REQUESTS (NUM_REQUESTS),
|
||||
.STAGE_1_CYCLES (STAGE_1_CYCLES),
|
||||
.CREQ_SIZE (CREQ_SIZE),
|
||||
.MRVQ_SIZE (MRVQ_SIZE),
|
||||
.DFPQ_SIZE (DFPQ_SIZE),
|
||||
.SNRQ_SIZE (SNRQ_SIZE),
|
||||
.CWBQ_SIZE (CWBQ_SIZE),
|
||||
.DWBQ_SIZE (DWBQ_SIZE),
|
||||
.DFQQ_SIZE (DFQQ_SIZE),
|
||||
.DRAM_ENABLE (DRAM_ENABLE),
|
||||
.WRITE_ENABLE (WRITE_ENABLE),
|
||||
.SNOOP_FORWARDING (SNOOP_FORWARDING),
|
||||
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
|
||||
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS),
|
||||
.SNP_REQ_TAG_WIDTH (SNP_REQ_TAG_WIDTH)
|
||||
) bank (
|
||||
`SCOPE_SIGNALS_CACHE_BANK_BIND
|
||||
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
// Core request
|
||||
@@ -452,11 +456,11 @@ module VX_cache #(
|
||||
);
|
||||
|
||||
VX_cache_core_rsp_merge #(
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE (WORD_SIZE),
|
||||
.NUM_REQUESTS (NUM_REQUESTS),
|
||||
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
|
||||
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS)
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE (WORD_SIZE),
|
||||
.NUM_REQUESTS (NUM_REQUESTS),
|
||||
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
|
||||
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS)
|
||||
) cache_core_rsp_merge (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
4
hw/rtl/cache/VX_cache_dram_req_arb.v
vendored
4
hw/rtl/cache/VX_cache_dram_req_arb.v
vendored
@@ -56,7 +56,7 @@ module VX_cache_dram_req_arb #(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
.dram_req (dram_req_valid && ~dram_req_rw),
|
||||
.dram_req (dram_req_valid && !dram_req_rw),
|
||||
.dram_req_addr(dram_req_addr),
|
||||
|
||||
.pref_pop (pref_pop),
|
||||
@@ -91,7 +91,7 @@ module VX_cache_dram_req_arb #(
|
||||
.dfqq_full (dfqq_full)
|
||||
);
|
||||
|
||||
assign dram_fill_req_ready = ~dfqq_full;
|
||||
assign dram_fill_req_ready = !dfqq_full;
|
||||
|
||||
wire [`BANK_BITS-1:0] dwb_bank;
|
||||
|
||||
|
||||
8
hw/rtl/cache/VX_cache_miss_resrv.v
vendored
8
hw/rtl/cache/VX_cache_miss_resrv.v
vendored
@@ -23,7 +23,7 @@ module VX_cache_miss_resrv #(
|
||||
|
||||
// Miss enqueue
|
||||
input wire miss_add,
|
||||
input wire from_mrvq,
|
||||
input wire is_mrvq,
|
||||
input wire[`LINE_ADDR_WIDTH-1:0] miss_add_addr,
|
||||
input wire[`UP(`WORD_SELECT_WIDTH)-1:0] miss_add_wsel,
|
||||
input wire[`WORD_WIDTH-1:0] miss_add_data,
|
||||
@@ -102,11 +102,11 @@ module VX_cache_miss_resrv #(
|
||||
miss_resrv_is_snp_st0,
|
||||
miss_resrv_snp_invalidate_st0} = metadata_table[dequeue_index];
|
||||
|
||||
wire mrvq_push = miss_add && enqueue_possible && !from_mrvq;
|
||||
wire mrvq_push = miss_add && enqueue_possible && !is_mrvq;
|
||||
wire mrvq_pop = miss_resrv_pop && dequeue_possible;
|
||||
|
||||
wire recover_state = miss_add && from_mrvq;
|
||||
wire increment_head = !miss_add && from_mrvq;
|
||||
wire recover_state = miss_add && is_mrvq;
|
||||
wire increment_head = !miss_add && is_mrvq;
|
||||
|
||||
wire update_ready = (|make_ready);
|
||||
|
||||
|
||||
2
hw/rtl/cache/VX_prefetcher.v
vendored
2
hw/rtl/cache/VX_prefetcher.v
vendored
@@ -28,7 +28,7 @@ module VX_prefetcher #(
|
||||
wire current_full;
|
||||
wire current_empty;
|
||||
|
||||
assign current_valid = ~current_empty;
|
||||
assign current_valid = !current_empty;
|
||||
|
||||
wire update_use = ((use_valid == 0) || ((use_valid-1) == 0)) && current_valid;
|
||||
|
||||
|
||||
23
hw/rtl/cache/VX_tag_data_access.v
vendored
23
hw/rtl/cache/VX_tag_data_access.v
vendored
@@ -50,14 +50,17 @@ module VX_tag_data_access #(
|
||||
output wire[BANK_LINE_SIZE-1:0] dirtyb_st1e,
|
||||
output wire fill_saw_dirty_st1e,
|
||||
output wire snp_to_mrvq_st1e,
|
||||
output wire mrvq_init_ready_state_st1e
|
||||
output wire mrvq_init_ready_state_st1e,
|
||||
|
||||
output wire tag_valid_st1e,
|
||||
output wire tag_match_st1e
|
||||
);
|
||||
|
||||
reg read_valid_st1c[STAGE_1_CYCLES-1:0];
|
||||
reg read_dirty_st1c[STAGE_1_CYCLES-1:0];
|
||||
reg[BANK_LINE_SIZE-1:0] read_dirtyb_st1c[STAGE_1_CYCLES-1:0];
|
||||
reg[`TAG_SELECT_BITS-1:0] read_tag_st1c [STAGE_1_CYCLES-1:0];
|
||||
reg[`BANK_LINE_WIDTH-1:0] read_data_st1c [STAGE_1_CYCLES-1:0];
|
||||
wire read_valid_st1c[STAGE_1_CYCLES-1:0];
|
||||
wire read_dirty_st1c[STAGE_1_CYCLES-1:0];
|
||||
wire[BANK_LINE_SIZE-1:0] read_dirtyb_st1c[STAGE_1_CYCLES-1:0];
|
||||
wire[`TAG_SELECT_BITS-1:0] read_tag_st1c [STAGE_1_CYCLES-1:0];
|
||||
wire[`BANK_LINE_WIDTH-1:0] read_data_st1c [STAGE_1_CYCLES-1:0];
|
||||
|
||||
wire qual_read_valid_st1;
|
||||
wire qual_read_dirty_st1;
|
||||
@@ -135,7 +138,7 @@ module VX_tag_data_access #(
|
||||
);
|
||||
end
|
||||
|
||||
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || ~DRAM_ENABLE; // If shared memory, always valid
|
||||
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || !DRAM_ENABLE; // If shared memory, always valid
|
||||
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && DRAM_ENABLE && WRITE_ENABLE; // Dirty only applies in Dcache
|
||||
assign use_read_tag_st1e = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writetag_st1e; // Tag is always the same in SM
|
||||
assign use_read_dirtyb_st1e= read_dirtyb_st1c[STAGE_1_CYCLES-1];
|
||||
@@ -177,11 +180,8 @@ module VX_tag_data_access #(
|
||||
wire snoop_hit_no_pending = valid_req_st1e && is_snp_st1e && use_read_valid_st1e && tags_match && (use_read_dirty_st1e || snp_invalidate_st1e) && !force_request_miss_st1e;
|
||||
wire req_invalid = valid_req_st1e && !is_snp_st1e && !use_read_valid_st1e && !writefill_st1e;
|
||||
wire req_miss = valid_req_st1e && !is_snp_st1e && use_read_valid_st1e && !writefill_st1e && !tags_match;
|
||||
|
||||
wire real_miss = req_invalid || req_miss;
|
||||
|
||||
wire force_core_miss = (force_request_miss_st1e && !is_snp_st1e && !writefill_st1e && valid_req_st1e && !real_miss);
|
||||
|
||||
assign snp_to_mrvq_st1e = valid_req_st1e && is_snp_st1e && force_request_miss_st1e;
|
||||
|
||||
// The second term is basically saying always make an entry ready if there's already antoher entry waiting, even if you yourself see a miss
|
||||
@@ -197,6 +197,9 @@ module VX_tag_data_access #(
|
||||
assign fill_saw_dirty_st1e = real_writefill && dirty_st1e;
|
||||
assign invalidate_line = snoop_hit_no_pending;
|
||||
|
||||
assign tag_valid_st1e = use_read_valid_st1e;
|
||||
assign tag_match_st1e = tags_match;
|
||||
|
||||
endmodule
|
||||
|
||||
|
||||
|
||||
@@ -1,14 +1,14 @@
|
||||
`include "VX_define.vh"
|
||||
|
||||
module VX_divide #(
|
||||
parameter WIDTHN=1,
|
||||
parameter WIDTHD=1,
|
||||
parameter NREP="UNSIGNED",
|
||||
parameter DREP="UNSIGNED",
|
||||
parameter SPEED="MIXED", // "MIXED" or "HIGHEST"
|
||||
parameter PIPELINE=0
|
||||
parameter WIDTHN = 1,
|
||||
parameter WIDTHD = 1,
|
||||
parameter REP = "UNSIGNED",
|
||||
parameter PIPELINE = 0
|
||||
) (
|
||||
input clock, aclr, clken,
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire clken,
|
||||
|
||||
input [WIDTHN-1:0] numer,
|
||||
input [WIDTHD-1:0] denom,
|
||||
@@ -17,105 +17,86 @@ module VX_divide #(
|
||||
output reg [WIDTHD-1:0] remainder
|
||||
);
|
||||
|
||||
generate
|
||||
`ifdef QUARTUS
|
||||
|
||||
if (NREP != DREP) begin
|
||||
different_nrep_drep_not_yet_supported non_existing_module();
|
||||
end
|
||||
lpm_divide #(
|
||||
.LPM_WIDTHN(WIDTHN),
|
||||
.LPM_WIDTHD(WIDTHD),
|
||||
.LPM_NREPRESENTATION(REP),
|
||||
.LPM_DREPRESENTATION(REP),
|
||||
.LPM_PIPELINE(PIPELINE),
|
||||
.DSP_BLOCK_BALANCING("LOGIC ELEMENTS"),
|
||||
.MAXIMIZE_SPEED(9)
|
||||
) quartus_divider (
|
||||
.clock(clk),
|
||||
.aclr(reset),
|
||||
.clken(clken),
|
||||
.numer(numer),
|
||||
.denom(denom),
|
||||
.quotient(quotient),
|
||||
.remain(remainder)
|
||||
);
|
||||
|
||||
`ifdef QUARTUS
|
||||
`else
|
||||
|
||||
localparam lpm_speed=SPEED == "HIGHEST" ? 9 : 5;
|
||||
wire [WIDTHN-1:0] numer_pipe_end;
|
||||
wire [WIDTHD-1:0] denom_pipe_end;
|
||||
|
||||
lpm_divide #(
|
||||
.LPM_WIDTHN(WIDTHN),
|
||||
.LPM_WIDTHD(WIDTHD),
|
||||
.LPM_NREPRESENTATION(NREP),
|
||||
.LPM_DREPRESENTATION(DREP),
|
||||
.LPM_PIPELINE(PIPELINE),
|
||||
.LPM_REMAINDERPOSITIVE("FALSE"), // emulate verilog % operator
|
||||
.MAXIMIZE_SPEED(lpm_speed)
|
||||
) quartus_divider (
|
||||
.clock(clock),
|
||||
.aclr(aclr),
|
||||
.clken(clken),
|
||||
.numer(numer),
|
||||
.denom(denom),
|
||||
.quotient(quotient),
|
||||
.remain(remainder)
|
||||
);
|
||||
if (PIPELINE == 0) begin
|
||||
assign numer_pipe_end = numer;
|
||||
assign denom_pipe_end = denom;
|
||||
end else begin
|
||||
reg [WIDTHN-1:0] numer_pipe [0:PIPELINE-1];
|
||||
reg [WIDTHD-1:0] denom_pipe [0:PIPELINE-1];
|
||||
|
||||
`else
|
||||
|
||||
wire [WIDTHN-1:0] numer_pipe_end;
|
||||
wire [WIDTHD-1:0] denom_pipe_end;
|
||||
|
||||
if (PIPELINE == 0) begin
|
||||
assign numer_pipe_end = numer;
|
||||
assign denom_pipe_end = denom;
|
||||
end else begin
|
||||
reg [WIDTHN-1:0] numer_pipe [0:PIPELINE-1];
|
||||
reg [WIDTHD-1:0] denom_pipe [0:PIPELINE-1];
|
||||
|
||||
genvar i;
|
||||
for (i = 0; i < PIPELINE-1; i++) begin : pipe_stages
|
||||
always @(posedge clock or posedge aclr) begin
|
||||
if (aclr) begin
|
||||
numer_pipe[i+1] <= 0;
|
||||
denom_pipe[i+1] <= 0;
|
||||
end
|
||||
else if (clken) begin
|
||||
numer_pipe[i+1] <= numer_pipe[i];
|
||||
denom_pipe[i+1] <= denom_pipe[i];
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
always @(posedge clock or posedge aclr) begin
|
||||
if (aclr) begin
|
||||
numer_pipe[0] <= 0;
|
||||
denom_pipe[0] <= 0;
|
||||
genvar i;
|
||||
for (i = 0; i < PIPELINE; i++) begin
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
numer_pipe[i] <= 0;
|
||||
denom_pipe[i] <= 0;
|
||||
end
|
||||
else if (clken) begin
|
||||
numer_pipe[0] <= numer;
|
||||
denom_pipe[0] <= denom;
|
||||
end
|
||||
end
|
||||
|
||||
assign numer_pipe_end = numer_pipe[PIPELINE-1];
|
||||
assign denom_pipe_end = denom_pipe[PIPELINE-1];
|
||||
end
|
||||
|
||||
/* * * * * * * * * * * * * * * * * * * * * * */
|
||||
/* Do the actual fallback computation here */
|
||||
/* * * * * * * * * * * * * * * * * * * * * * */
|
||||
|
||||
if (NREP == "SIGNED") begin
|
||||
always @(*) begin
|
||||
if (denom_pipe_end == 0) begin
|
||||
quotient = 32'hffffffff;
|
||||
remainder = numer_pipe_end;
|
||||
end
|
||||
else if (denom_pipe_end == 32'hffffffff
|
||||
&& numer_pipe_end == 32'h80000000) begin
|
||||
// this edge case kills verilator in some cases by causing a division
|
||||
// overflow exception. INT_MIN / -1 (on x86)
|
||||
quotient = 0;
|
||||
remainder = 0;
|
||||
end
|
||||
else begin
|
||||
quotient = $signed(numer_pipe_end) / $signed(denom_pipe_end);
|
||||
remainder = $signed(numer_pipe_end) % $signed(denom_pipe_end);
|
||||
if (i == 0) begin
|
||||
numer_pipe[0] <= 0;
|
||||
denom_pipe[0] <= 0;
|
||||
end else begin
|
||||
numer_pipe[i] <= numer_pipe[i-1];
|
||||
denom_pipe[i] <= denom_pipe[i-1];
|
||||
end
|
||||
end
|
||||
end
|
||||
end
|
||||
else begin
|
||||
assign quotient = (denom_pipe_end == 0) ? 32'hffffffff : numer_pipe_end/denom_pipe_end;
|
||||
assign remainder = (denom_pipe_end == 0) ? numer_pipe_end : numer_pipe_end%denom_pipe_end;
|
||||
end
|
||||
|
||||
assign numer_pipe_end = numer_pipe[PIPELINE-1];
|
||||
assign denom_pipe_end = denom_pipe[PIPELINE-1];
|
||||
end
|
||||
|
||||
always @(*) begin
|
||||
if (denom_pipe_end == 0) begin
|
||||
quotient = {WIDTHN{1'b1}};
|
||||
remainder = numer_pipe_end;
|
||||
end
|
||||
`ifndef SYNTHESIS
|
||||
// this edge case kills verilator in some cases by causing a division
|
||||
// overflow exception. INT_MIN / -1 (on x86)
|
||||
else if (numer_pipe_end == {1'b1, (WIDTHN-1)'(0)}
|
||||
&& denom_pipe_end == {WIDTHD{1'b1}}) begin
|
||||
quotient = 0;
|
||||
remainder = 0;
|
||||
end
|
||||
`endif
|
||||
else begin
|
||||
if (REP == "SIGNED") begin
|
||||
quotient = $signed(numer_pipe_end) / $signed(denom_pipe_end);
|
||||
remainder = $signed(numer_pipe_end) % $signed(denom_pipe_end);
|
||||
end else begin
|
||||
quotient = numer_pipe_end / denom_pipe_end;
|
||||
remainder = numer_pipe_end % denom_pipe_end;
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
endgenerate
|
||||
`endif
|
||||
|
||||
endmodule : VX_divide
|
||||
|
||||
@@ -28,8 +28,8 @@ module VX_indexable_queue #(
|
||||
assign empty = (wr_ptr == rd_ptr);
|
||||
assign full = (wr_a == rd_a) && (wr_ptr[`LOG2UP(SIZE)] != rd_ptr[`LOG2UP(SIZE)]);
|
||||
|
||||
assign enqueue = push && ~full;
|
||||
assign dequeue = ~empty && ~valid[rd_a]; // auto-remove when head is invalid
|
||||
assign enqueue = push && !full;
|
||||
assign dequeue = !empty && !valid[rd_a]; // auto-remove when head is invalid
|
||||
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
|
||||
@@ -30,17 +30,17 @@ module VX_matrix_arbiter #(
|
||||
for (i = 0; i < N; ++i) begin
|
||||
for (j = 0; j < N; ++j) begin
|
||||
if (j > i) begin
|
||||
assign pri[j][i] = requests[i] & state[i][j];
|
||||
assign pri[j][i] = requests[i] && state[i][j];
|
||||
end
|
||||
else if (j < i) begin
|
||||
assign pri[j][i] = requests[i] & ~state[j][i];
|
||||
assign pri[j][i] = requests[i] && !state[j][i];
|
||||
end
|
||||
else begin
|
||||
assign pri[j][i] = 0;
|
||||
end
|
||||
end
|
||||
|
||||
assign grant_onehot[i] = requests[i] & ~(| pri[i]);
|
||||
assign grant_onehot[i] = requests[i] && !(| pri[i]);
|
||||
end
|
||||
|
||||
for (i = 0; i < N; ++i) begin
|
||||
@@ -50,7 +50,7 @@ module VX_matrix_arbiter #(
|
||||
state[i][j] <= 0;
|
||||
end
|
||||
else begin
|
||||
state[i][j] <= (state[i][j] || grant_onehot[j]) && ~grant_onehot[i];
|
||||
state[i][j] <= (state[i][j] || grant_onehot[j]) && !grant_onehot[i];
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
@@ -1,16 +1,14 @@
|
||||
`include "VX_define.vh"
|
||||
|
||||
module VX_mult #(
|
||||
parameter WIDTHA=1,
|
||||
parameter WIDTHB=1,
|
||||
parameter WIDTHP=1,
|
||||
parameter REP="UNSIGNED",
|
||||
parameter SPEED="MIXED", // "MIXED" or "HIGHEST"
|
||||
parameter PIPELINE=0,
|
||||
parameter FORCE_LE="NO"
|
||||
parameter WIDTHA = 1,
|
||||
parameter WIDTHB = 1,
|
||||
parameter WIDTHP = 1,
|
||||
parameter REP = "UNSIGNED",
|
||||
parameter PIPELINE = 0
|
||||
) (
|
||||
input clock,
|
||||
input aclr,
|
||||
input clk,
|
||||
input reset,
|
||||
input clken,
|
||||
|
||||
input [WIDTHA-1:0] dataa,
|
||||
@@ -19,102 +17,67 @@ module VX_mult #(
|
||||
output reg [WIDTHP-1:0] result
|
||||
);
|
||||
|
||||
generate
|
||||
`ifdef QUARTUS
|
||||
|
||||
`ifdef QUARTUS
|
||||
lpm_mult #(
|
||||
.LPM_WIDTHA(WIDTHA),
|
||||
.LPM_WIDTHB(WIDTHB),
|
||||
.LPM_WIDTHP(WIDTHP),
|
||||
.LPM_REPRESENTATION(REP),
|
||||
.LPM_PIPELINE(PIPELINE),
|
||||
.DSP_BLOCK_BALANCING("LOGIC ELEMENTS"),
|
||||
.MAXIMIZE_SPEED(9)
|
||||
) quartus_mult (
|
||||
.clock(clk),
|
||||
.aclr(reset),
|
||||
.clken(clken),
|
||||
.dataa(dataa),
|
||||
.datab(datab),
|
||||
.result(result)
|
||||
);
|
||||
|
||||
localparam lpm_speed = (SPEED == "HIGHEST") ? 10 : 5;
|
||||
`else
|
||||
|
||||
if (FORCE_LE == "YES") begin
|
||||
lpm_mult #(
|
||||
.LPM_WIDTHA(WIDTHA),
|
||||
.LPM_WIDTHB(WIDTHB),
|
||||
.LPM_WIDTHP(WIDTHP),
|
||||
.LPM_REPRESENTATION(REP),
|
||||
.LPM_PIPELINE(PIPELINE),
|
||||
.DSP_BLOCK_BALANCING("LOGIC ELEMENTS"),
|
||||
.MAXIMIZE_SPEED(lpm_speed)
|
||||
) quartus_mult (
|
||||
.clock(clock),
|
||||
.aclr(aclr),
|
||||
.clken(clken),
|
||||
.dataa(dataa),
|
||||
.datab(datab),
|
||||
.result(result)
|
||||
);
|
||||
end
|
||||
else begin
|
||||
lpm_mult#(
|
||||
.LPM_WIDTHA(WIDTHA),
|
||||
.LPM_WIDTHB(WIDTHB),
|
||||
.LPM_WIDTHP(WIDTHP),
|
||||
.LPM_REPRESENTATION(REP),
|
||||
.LPM_PIPELINE(PIPELINE),
|
||||
.MAXIMIZE_SPEED(lpm_speed)
|
||||
) quartus_mult(
|
||||
.clock(clock),
|
||||
.aclr(aclr),
|
||||
.clken(clken),
|
||||
.dataa(dataa),
|
||||
.datab(datab),
|
||||
.result(result)
|
||||
);
|
||||
end
|
||||
wire [WIDTHA-1:0] dataa_pipe_end;
|
||||
wire [WIDTHB-1:0] datab_pipe_end;
|
||||
|
||||
`else
|
||||
if (PIPELINE == 0) begin
|
||||
assign dataa_pipe_end = dataa;
|
||||
assign datab_pipe_end = datab;
|
||||
end else begin
|
||||
reg [WIDTHA-1:0] dataa_pipe [0:PIPELINE-1];
|
||||
reg [WIDTHB-1:0] datab_pipe [0:PIPELINE-1];
|
||||
|
||||
wire [WIDTHA-1:0] dataa_pipe_end;
|
||||
wire [WIDTHB-1:0] datab_pipe_end;
|
||||
|
||||
if (PIPELINE == 0) begin
|
||||
assign dataa_pipe_end = dataa;
|
||||
assign datab_pipe_end = datab;
|
||||
end else begin
|
||||
reg [WIDTHA-1:0] dataa_pipe [0:PIPELINE-1];
|
||||
reg [WIDTHB-1:0] datab_pipe [0:PIPELINE-1];
|
||||
|
||||
genvar i;
|
||||
for (i = 0; i < PIPELINE-1; i++) begin : pipe_stages
|
||||
always @(posedge clock or posedge aclr) begin
|
||||
if (aclr) begin
|
||||
dataa_pipe[i+1] <= 0;
|
||||
datab_pipe[i+1] <= 0;
|
||||
end
|
||||
else if (clken) begin
|
||||
dataa_pipe[i+1] <= dataa_pipe[i];
|
||||
datab_pipe[i+1] <= datab_pipe[i];
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
always @(posedge clock or posedge aclr) begin
|
||||
if (aclr) begin
|
||||
dataa_pipe[0] <= 0;
|
||||
datab_pipe[0] <= 0;
|
||||
genvar i;
|
||||
for (i = 0; i < PIPELINE; i++) begin
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
dataa_pipe[i] <= 0;
|
||||
datab_pipe[i] <= 0;
|
||||
end
|
||||
else if (clken) begin
|
||||
dataa_pipe[0] <= dataa;
|
||||
datab_pipe[0] <= datab;
|
||||
if (i == 0) begin
|
||||
dataa_pipe[0] <= dataa;
|
||||
datab_pipe[0] <= datab;
|
||||
end else begin
|
||||
dataa_pipe[i] <= dataa_pipe[i-1];
|
||||
datab_pipe[i] <= datab_pipe[i-1];
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
assign dataa_pipe_end = dataa_pipe[PIPELINE-1];
|
||||
assign datab_pipe_end = datab_pipe[PIPELINE-1];
|
||||
end
|
||||
|
||||
/* * * * * * * * * * * * * * * * * * * * * * */
|
||||
/* Do the actual fallback computation here */
|
||||
/* * * * * * * * * * * * * * * * * * * * * * */
|
||||
assign dataa_pipe_end = dataa_pipe[PIPELINE-1];
|
||||
assign datab_pipe_end = datab_pipe[PIPELINE-1];
|
||||
end
|
||||
|
||||
if (REP == "SIGNED") begin
|
||||
assign result = $signed(dataa_pipe_end) * $signed(datab_pipe_end);
|
||||
end
|
||||
else begin
|
||||
assign result = dataa_pipe_end * datab_pipe_end;
|
||||
end
|
||||
if (REP == "SIGNED") begin
|
||||
assign result = $signed(dataa_pipe_end) * $signed(datab_pipe_end);
|
||||
end
|
||||
else begin
|
||||
assign result = dataa_pipe_end * datab_pipe_end;
|
||||
end
|
||||
|
||||
`endif
|
||||
|
||||
endgenerate
|
||||
`endif
|
||||
|
||||
endmodule: VX_mult
|
||||
|
||||
@@ -155,6 +155,6 @@ module VX_tb_divide();
|
||||
end
|
||||
|
||||
always #1
|
||||
clk = ~clk;
|
||||
clk = !clk;
|
||||
|
||||
endmodule: VX_tb_divide
|
||||
Reference in New Issue
Block a user