tex_unit memory partial

This commit is contained in:
Blaise Tine
2021-03-22 10:25:17 -04:00
parent 0909ad3201
commit f499ce6c95
7 changed files with 271 additions and 335 deletions

View File

@@ -281,13 +281,19 @@
// Word size in bytes // Word size in bytes
`define DWORD_SIZE 4 `define DWORD_SIZE 4
// TAG sharing enable // TAG sharing enable
`define DCORE_TAG_ID_BITS `LOG2UP(`LSUQ_SIZE) `define LSUQ_ADDR_BITS `LOG2UP(`LSUQ_SIZE)
`ifdef EXT_TEX_ENABLE
`define DCORE_TAG_ID_BITS (`LSUQ_ADDR_BITS + 2)
`else
`define DCORE_TAG_ID_BITS `LSUQ_ADDR_BITS
`endif
// Core request tag bits // Core request tag bits
`ifdef EXT_TEX_ENABLE `ifdef EXT_TEX_ENABLE
`define LSU_TAG_WIDTH (`DBG_CACHE_REQ_MDATAW + `DCORE_TAG_ID_BITS) `define LSU_DACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `LSUQ_ADDR_BITS)
`define DCORE_TAG_WIDTH (`LSU_TAG_WIDTH+1) `define TEX_DACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + 2 + `LSUQ_ADDR_BITS)
`define DCORE_TAG_WIDTH (`MAX(`LSU_DACHE_TAG_BITS, `TEX_DACHE_TAG_BITS) + 1)
`else `else
`define DCORE_TAG_WIDTH (`DBG_CACHE_REQ_MDATAW + `DCORE_TAG_ID_BITS) `define DCORE_TAG_WIDTH (`DBG_CACHE_REQ_MDATAW + `DCORE_TAG_ID_BITS)
`endif `endif

View File

@@ -51,34 +51,37 @@ module VX_execute #(
VX_dcache_core_req_if #( VX_dcache_core_req_if #(
.LANES(`NUM_THREADS), .LANES(`NUM_THREADS),
.WORD_SIZE(4), .WORD_SIZE(4),
.CORE_TAG_WIDTH(`LSU_TAG_WIDTH) .CORE_TAG_WIDTH(`LSU_DACHE_TAG_BITS)
) tex_dcache_req_if();
VX_dcache_core_rsp_if #(
.LANES(`NUM_THREADS),
.WORD_SIZE(4),
.CORE_TAG_WIDTH(`LSU_TAG_WIDTH)
) tex_dcache_rsp_if();
VX_dcache_core_req_if #(
.LANES(`NUM_THREADS),
.WORD_SIZE(4),
.CORE_TAG_WIDTH(`LSU_TAG_WIDTH)
) lsu_dcache_req_if(); ) lsu_dcache_req_if();
VX_dcache_core_rsp_if #( VX_dcache_core_rsp_if #(
.LANES(`NUM_THREADS), .LANES(`NUM_THREADS),
.WORD_SIZE(4), .WORD_SIZE(4),
.CORE_TAG_WIDTH(`LSU_TAG_WIDTH) .CORE_TAG_WIDTH(`LSU_DACHE_TAG_BITS)
) lsu_dcache_rsp_if(); ) lsu_dcache_rsp_if();
VX_dcache_core_req_if #(
.LANES(`NUM_THREADS),
.WORD_SIZE(4),
.CORE_TAG_WIDTH(`TEX_DACHE_TAG_BITS)
) tex_dcache_req_if();
VX_dcache_core_rsp_if #(
.LANES(`NUM_THREADS),
.WORD_SIZE(4),
.CORE_TAG_WIDTH(`TEX_DACHE_TAG_BITS)
) tex_dcache_rsp_if();
VX_tex_csr_if tex_csr_if(); VX_tex_csr_if tex_csr_if();
wire [1:0] tmp;
`UNUSED_VAR (tmp)
VX_tex_lsu_arb #( VX_tex_lsu_arb #(
.NUM_REQS (2), .NUM_REQS (2),
.LANES (`NUM_THREADS), .LANES (`NUM_THREADS),
.WORD_SIZE (4), .WORD_SIZE (4),
.TAG_IN_WIDTH (`LSU_TAG_WIDTH), .TAG_IN_WIDTH (`MAX(`LSU_DACHE_TAG_BITS, `TEX_DACHE_TAG_BITS)),
.TAG_OUT_WIDTH (`DCORE_TAG_WIDTH) .TAG_OUT_WIDTH (`DCORE_TAG_WIDTH)
) tex_lsu_arb ( ) tex_lsu_arb (
.clk (clk), .clk (clk),
@@ -90,7 +93,7 @@ module VX_execute #(
.req_byteen_in ({tex_dcache_req_if.byteen, lsu_dcache_req_if.byteen}), .req_byteen_in ({tex_dcache_req_if.byteen, lsu_dcache_req_if.byteen}),
.req_addr_in ({tex_dcache_req_if.addr, lsu_dcache_req_if.addr}), .req_addr_in ({tex_dcache_req_if.addr, lsu_dcache_req_if.addr}),
.req_data_in ({tex_dcache_req_if.data, lsu_dcache_req_if.data}), .req_data_in ({tex_dcache_req_if.data, lsu_dcache_req_if.data}),
.req_tag_in ({tex_dcache_req_if.tag, lsu_dcache_req_if.tag}), .req_tag_in ({tex_dcache_req_if.tag, {2'b0, lsu_dcache_req_if.tag}}),
.req_ready_in ({tex_dcache_req_if.ready, lsu_dcache_req_if.ready}), .req_ready_in ({tex_dcache_req_if.ready, lsu_dcache_req_if.ready}),
// Dcache request // Dcache request
@@ -105,7 +108,7 @@ module VX_execute #(
// Tex/LSU response // Tex/LSU response
.rsp_valid_out ({tex_dcache_rsp_if.valid, lsu_dcache_rsp_if.valid}), .rsp_valid_out ({tex_dcache_rsp_if.valid, lsu_dcache_rsp_if.valid}),
.rsp_data_out ({tex_dcache_rsp_if.data, lsu_dcache_rsp_if.data}), .rsp_data_out ({tex_dcache_rsp_if.data, lsu_dcache_rsp_if.data}),
.rsp_tag_out ({tex_dcache_rsp_if.tag, lsu_dcache_rsp_if.tag}), .rsp_tag_out ({tex_dcache_rsp_if.tag, {tmp, lsu_dcache_rsp_if.tag}}),
.rsp_ready_out ({tex_dcache_rsp_if.ready, lsu_dcache_rsp_if.ready}), .rsp_ready_out ({tex_dcache_rsp_if.ready, lsu_dcache_rsp_if.ready}),
// Dcache response // Dcache response

View File

@@ -45,7 +45,7 @@ module VX_lsu_unit #(
wire is_dup_load = lsu_req_if.wb && lsu_req_if.tmask[0] && (& addr_matches); wire is_dup_load = lsu_req_if.wb && lsu_req_if.tmask[0] && (& addr_matches);
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
reg [`LSUQ_SIZE-1:0][`DCORE_TAG_ID_BITS-1:0] pending_tags; reg [`LSUQ_SIZE-1:0][`LSUQ_ADDR_BITS-1:0] pending_tags;
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
wire ready_in; wire ready_in;
@@ -75,12 +75,12 @@ module VX_lsu_unit #(
`UNUSED_VAR (rsp_type) `UNUSED_VAR (rsp_type)
reg [`LSUQ_SIZE-1:0][`NUM_THREADS-1:0] rsp_rem_mask; reg [`LSUQ_SIZE-1:0][`NUM_THREADS-1:0] rsp_rem_mask;
reg [`NUM_THREADS-1:0] rsp_rem_mask_n; wire [`NUM_THREADS-1:0] rsp_rem_mask_n;
reg [`NUM_THREADS-1:0] req_sent_mask; reg [`NUM_THREADS-1:0] req_sent_mask;
wire req_sent_all; wire req_sent_all;
wire [`DCORE_TAG_ID_BITS-1:0] mbuf_waddr, mbuf_raddr; wire [`LSUQ_ADDR_BITS-1:0] mbuf_waddr, mbuf_raddr;
wire mbuf_full; wire mbuf_full;
wire [`NUM_THREADS-1:0][1:0] req_offset, rsp_offset; wire [`NUM_THREADS-1:0][1:0] req_offset, rsp_offset;
@@ -88,15 +88,17 @@ module VX_lsu_unit #(
assign req_offset[i] = req_addr[i][1:0]; assign req_offset[i] = req_addr[i][1:0];
end end
wire mbuf_push = (| (dcache_req_if.valid & dcache_req_if.ready)) wire [`NUM_THREADS-1:0] dcache_req_fire = dcache_req_if.valid & dcache_req_if.ready;
wire dcache_rsp_fire = (| dcache_rsp_if.valid) && dcache_rsp_if.ready;
wire mbuf_push = (| dcache_req_fire)
&& (0 == req_sent_mask) // first submission only && (0 == req_sent_mask) // first submission only
&& req_wb; // loads only && req_wb; // loads only
wire mbuf_pop_part = (| dcache_rsp_if.valid) && dcache_rsp_if.ready; wire mbuf_pop = dcache_rsp_fire && (rsp_rem_mask_n == 0 || rsp_is_dup);
wire mbuf_pop = mbuf_pop_part && (rsp_rem_mask_n == 0 || rsp_is_dup);
assign mbuf_raddr = dcache_rsp_if.tag[`DCORE_TAG_ID_BITS-1:0]; assign mbuf_raddr = dcache_rsp_if.tag[`LSUQ_ADDR_BITS-1:0];
VX_index_buffer #( VX_index_buffer #(
.DATAW (`NW_BITS + 32 + `NR_BITS + 1 + `LSU_BITS + (`NUM_THREADS * 2) + 1), .DATAW (`NW_BITS + 32 + `NR_BITS + 1 + `LSU_BITS + (`NUM_THREADS * 2) + 1),
@@ -114,23 +116,20 @@ module VX_lsu_unit #(
.full (mbuf_full) .full (mbuf_full)
); );
assign req_sent_all = (&(dcache_req_if.ready | req_sent_mask | ~req_tmask)) assign req_sent_all = (&(dcache_req_fire | req_sent_mask | ~req_tmask))
|| (req_is_dup && dcache_req_if.ready[0]); || (req_is_dup & dcache_req_if.valid[0] & dcache_req_if.ready[0]);
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset || req_sent_all) begin
req_sent_mask <= 0; req_sent_mask <= 0;
end else begin end else if (!req_sent_all) begin
if (req_sent_all) req_sent_mask <= req_sent_mask | dcache_req_fire;
req_sent_mask <= 0;
else
req_sent_mask <= req_sent_mask | (dcache_req_if.valid & dcache_req_if.ready);
end end
end end
// need to hold the acquired tag index until the full request is submitted // need to hold the acquired tag index until the full request is submitted
reg [`DCORE_TAG_ID_BITS-1:0] req_tag_hold; reg [`LSUQ_ADDR_BITS-1:0] req_tag_hold;
wire [`DCORE_TAG_ID_BITS-1:0] req_tag = (0 == req_sent_mask) ? mbuf_waddr : req_tag_hold; wire [`LSUQ_ADDR_BITS-1:0] req_tag = (0 == req_sent_mask) ? mbuf_waddr : req_tag_hold;
always @(posedge clk) begin always @(posedge clk) begin
if (mbuf_push) if (mbuf_push)
req_tag_hold <= mbuf_waddr; req_tag_hold <= mbuf_waddr;
@@ -142,12 +141,13 @@ module VX_lsu_unit #(
rsp_rem_mask[mbuf_waddr] <= req_tmask; rsp_rem_mask[mbuf_waddr] <= req_tmask;
pending_tags[mbuf_waddr] <= req_tag; pending_tags[mbuf_waddr] <= req_tag;
end end
if (mbuf_pop_part) begin if (dcache_rsp_fire) begin
rsp_rem_mask[mbuf_raddr] <= rsp_rem_mask_n; rsp_rem_mask[mbuf_raddr] <= rsp_rem_mask_n;
end end
end end
wire req_ready_dep = (req_wb && ~mbuf_full) || (~req_wb && st_commit_if.ready); wire req_ready_dep = (req_wb && ~mbuf_full)
|| (~req_wb && st_commit_if.ready);
wire [`NUM_THREADS-1:0] dup_mask = {{(`NUM_THREADS-1){~req_is_dup}}, 1'b1}; wire [`NUM_THREADS-1:0] dup_mask = {{(`NUM_THREADS-1){~req_is_dup}}, 1'b1};
@@ -188,7 +188,7 @@ module VX_lsu_unit #(
assign dcache_req_if.data = mem_req_data; assign dcache_req_if.data = mem_req_data;
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
assign dcache_req_if.tag = {`NUM_THREADS{{req_pc, req_wid, req_tag}}}; assign dcache_req_if.tag = {`NUM_THREADS{req_pc, req_wid, req_tag}};
`else `else
assign dcache_req_if.tag = {`NUM_THREADS{req_tag}}; assign dcache_req_if.tag = {`NUM_THREADS{req_tag}};
`endif `endif
@@ -255,7 +255,7 @@ module VX_lsu_unit #(
assign dcache_rsp_if.ready = ~load_rsp_stall; assign dcache_rsp_if.ready = ~load_rsp_stall;
// scope registration // scope registration
`SCOPE_ASSIGN (dcache_req_fire, dcache_req_if.valid & dcache_req_if.ready); `SCOPE_ASSIGN (dcache_req_fire, dcache_req_fire);
`SCOPE_ASSIGN (dcache_req_wid, req_wid); `SCOPE_ASSIGN (dcache_req_wid, req_wid);
`SCOPE_ASSIGN (dcache_req_pc, req_pc); `SCOPE_ASSIGN (dcache_req_pc, req_pc);
`SCOPE_ASSIGN (dcache_req_addr, req_addr); `SCOPE_ASSIGN (dcache_req_addr, req_addr);
@@ -269,15 +269,15 @@ module VX_lsu_unit #(
`ifdef DBG_PRINT_CORE_DCACHE `ifdef DBG_PRINT_CORE_DCACHE
always @(posedge clk) begin always @(posedge clk) begin
if ((| (dcache_req_if.valid & dcache_req_if.ready))) begin if ((| dcache_req_fire)) begin
if ((| dcache_req_if.rw)) if ((| dcache_req_if.rw))
$display("%t: D$%0d Wr Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, data=%0h", $display("%t: D$%0d Wr Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, data=%0h",
$time, CORE_ID, req_wid, req_pc, (dcache_req_if.valid & dcache_req_if.ready), req_addr, dcache_req_if.tag, dcache_req_if.byteen, dcache_req_if.data); $time, CORE_ID, req_wid, req_pc, dcache_req_fire, req_addr, dcache_req_if.tag, dcache_req_if.byteen, dcache_req_if.data);
else else
$display("%t: D$%0d Rd Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, rd=%0d, is_dup=%b", $display("%t: D$%0d Rd Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, rd=%0d, is_dup=%b",
$time, CORE_ID, req_wid, req_pc, (dcache_req_if.valid & dcache_req_if.ready), req_addr, dcache_req_if.tag, dcache_req_if.byteen, req_rd, req_is_dup); $time, CORE_ID, req_wid, req_pc, dcache_req_fire, req_addr, dcache_req_if.tag, dcache_req_if.byteen, req_rd, req_is_dup);
end end
if ((| dcache_rsp_if.valid) && dcache_rsp_if.ready) begin if (dcache_rsp_fire) begin
$display("%t: D$%0d Rsp: valid=%b, wid=%0d, PC=%0h, tag=%0h, rd=%0d, data=%0h, is_dup=%b", $display("%t: D$%0d Rsp: valid=%b, wid=%0d, PC=%0h, tag=%0h, rd=%0d, data=%0h, is_dup=%b",
$time, CORE_ID, dcache_rsp_if.valid, rsp_wid, rsp_pc, dcache_rsp_if.tag, rsp_rd, dcache_rsp_if.data, rsp_is_dup); $time, CORE_ID, dcache_rsp_if.valid, rsp_wid, rsp_pc, dcache_rsp_if.tag, rsp_rd, dcache_rsp_if.data, rsp_is_dup);
end end

View File

@@ -2,7 +2,7 @@
module VX_tex_addr_gen #( module VX_tex_addr_gen #(
parameter CORE_ID = 0, parameter CORE_ID = 0,
parameter REQ_TAG_WIDTH = 1 parameter REQ_INFO_WIDTH = 1
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,
@@ -14,17 +14,19 @@ module VX_tex_addr_gen #(
// inputs // inputs
input wire [`NW_BITS-1:0] req_wid,
input wire [`NUM_THREADS-1:0] req_tmask, input wire [`NUM_THREADS-1:0] req_tmask,
input wire [REQ_TAG_WIDTH-1:0] req_tag, input wire [31:0] req_PC,
input wire [REQ_INFO_WIDTH-1:0] req_info,
input wire [`TEX_FILTER_BITS-1:0] filter, input wire [`TEX_FILTER_BITS-1:0] filter,
input wire [`TEX_WRAP_BITS-1:0] wrap_u, input wire [`TEX_WRAP_BITS-1:0] wrap_u,
input wire [`TEX_WRAP_BITS-1:0] wrap_v, input wire [`TEX_WRAP_BITS-1:0] wrap_v,
input wire [`TEX_ADDR_BITS-1:0] base_addr, input wire [`TEX_ADDR_BITS-1:0] base_addr,
input wire [`TEX_STRIDE_BITS-1:0] log2_stride, input wire [`TEX_STRIDE_BITS-1:0] log_stride,
input wire [`TEX_WIDTH_BITS-1:0] log2_width, input wire [`TEX_WIDTH_BITS-1:0] log_width,
input wire [`TEX_HEIGHT_BITS-1:0] log2_height, input wire [`TEX_HEIGHT_BITS-1:0] log_height,
input wire [`NUM_THREADS-1:0][31:0] coord_u, input wire [`NUM_THREADS-1:0][31:0] coord_u,
input wire [`NUM_THREADS-1:0][31:0] coord_v, input wire [`NUM_THREADS-1:0][31:0] coord_v,
@@ -32,14 +34,17 @@ module VX_tex_addr_gen #(
// outputs // outputs
output wire mem_req_valid, output wire mem_req_valid,
output wire [`NUM_THREADS-1:0] mem_req_tmask, output wire [`NW_BITS-1:0] mem_req_wid,
output wire [`NUM_THREADS-1:0] mem_req_tmask,
output wire [31:0] mem_req_PC,
output wire [`TEX_FILTER_BITS-1:0] mem_req_filter, output wire [`TEX_FILTER_BITS-1:0] mem_req_filter,
output wire [`TEX_STRIDE_BITS-1:0] mem_req_stride,
output wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_u, output wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_u,
output wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_v, output wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_v,
output wire [REQ_TAG_WIDTH-1:0] mem_req_tag, output wire [REQ_INFO_WIDTH-1:0] mem_req_info,
output wire [`NUM_THREADS-1:0][3:0][31:0] mem_req_addr, output wire [`NUM_THREADS-1:0][3:0][31:0] mem_req_addr,
input wire mem_req_ready input wire mem_req_ready
); );
`UNUSED_PARAM (CORE_ID) `UNUSED_PARAM (CORE_ID)
@@ -55,10 +60,10 @@ module VX_tex_addr_gen #(
wire [31:0] fu[1:0]; wire [31:0] fu[1:0];
wire [31:0] fv[1:0]; wire [31:0] fv[1:0];
assign fu[0] = coord_u[i] - (filter ? (`FIXED_HALF >> log2_width) : 0); assign fu[0] = coord_u[i] - (filter ? (`FIXED_HALF >> log_width) : 0);
assign fv[0] = coord_v[i] - (filter ? (`FIXED_HALF >> log2_height) : 0); assign fv[0] = coord_v[i] - (filter ? (`FIXED_HALF >> log_height) : 0);
assign fu[1] = coord_u[i] + (filter ? (`FIXED_HALF >> log2_width) : 0); assign fu[1] = coord_u[i] + (filter ? (`FIXED_HALF >> log_width) : 0);
assign fv[1] = coord_v[i] + (filter ? (`FIXED_HALF >> log2_height) : 0); assign fv[1] = coord_v[i] + (filter ? (`FIXED_HALF >> log_height) : 0);
VX_tex_wrap #( VX_tex_wrap #(
.CORE_ID (CORE_ID) .CORE_ID (CORE_ID)
@@ -102,28 +107,28 @@ module VX_tex_addr_gen #(
wire [`FIXED_FRAC-1:0] x [1:0]; wire [`FIXED_FRAC-1:0] x [1:0];
wire [`FIXED_FRAC-1:0] y [1:0]; wire [`FIXED_FRAC-1:0] y [1:0];
assign x[0] = u[0][i] >> ((`FIXED_FRAC) - log2_width); assign x[0] = u[0][i] >> ((`FIXED_FRAC) - log_width);
assign x[1] = u[1][i] >> ((`FIXED_FRAC) - log2_width); assign x[1] = u[1][i] >> ((`FIXED_FRAC) - log_width);
assign y[0] = v[0][i] >> ((`FIXED_FRAC) - log2_height); assign y[0] = v[0][i] >> ((`FIXED_FRAC) - log_height);
assign y[1] = v[1][i] >> ((`FIXED_FRAC) - log2_height); assign y[1] = v[1][i] >> ((`FIXED_FRAC) - log_height);
assign addr[i][0] = base_addr + (32'(x[0]) + (32'(y[0]) << log2_width)) << log2_stride; assign addr[i][0] = base_addr + (32'(x[0]) + (32'(y[0]) << log_width)) << log_stride;
assign addr[i][1] = base_addr + (32'(x[1]) + (32'(y[0]) << log2_width)) << log2_stride; assign addr[i][1] = base_addr + (32'(x[1]) + (32'(y[0]) << log_width)) << log_stride;
assign addr[i][2] = base_addr + (32'(x[0]) + (32'(y[1]) << log2_width)) << log2_stride; assign addr[i][2] = base_addr + (32'(x[0]) + (32'(y[1]) << log_width)) << log_stride;
assign addr[i][3] = base_addr + (32'(x[1]) + (32'(y[1]) << log2_width)) << log2_stride; assign addr[i][3] = base_addr + (32'(x[1]) + (32'(y[1]) << log_width)) << log_stride;
end end
wire stall_out = mem_req_valid && ~mem_req_ready; wire stall_out = mem_req_valid && ~mem_req_ready;
VX_pipe_register #( VX_pipe_register #(
.DATAW (1 + `NUM_THREADS + `TEX_FILTER_BITS + REQ_TAG_WIDTH + (`NUM_THREADS * 4 * 32) + (2*`NUM_THREADS * `FIXED_FRAC)), .DATAW (1 + `NW_BITS + `NUM_THREADS + 32 + `TEX_FILTER_BITS + `TEX_STRIDE_BITS + REQ_INFO_WIDTH + (`NUM_THREADS * 4 * 32) + (2*`NUM_THREADS * `FIXED_FRAC)),
.RESETW (1) .RESETW (1)
) pipe_reg ( ) pipe_reg (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.enable (~stall_out), .enable (~stall_out),
.data_in ({valid_in, req_tmask, filter, req_tag, addr, u[0], v[0]}), .data_in ({valid_in, req_wid, req_tmask, req_PC, filter, log_stride, req_info, addr, u[0], v[0]}),
.data_out ({mem_req_valid, mem_req_tmask, mem_req_filter, mem_req_tag, mem_req_addr, mem_req_u, mem_req_v}) .data_out ({mem_req_valid, mem_req_wid, mem_req_tmask, mem_req_PC, mem_req_filter, mem_req_stride, mem_req_info, mem_req_addr, mem_req_u, mem_req_v})
); );
assign ready_in = ~stall_out; assign ready_in = ~stall_out;

View File

@@ -1,11 +1,9 @@
`include "VX_tex_define.vh" `include "VX_tex_define.vh"
module VX_tex_memory #( module VX_tex_memory #(
parameter CORE_ID = 0, parameter CORE_ID = 0,
parameter REQ_TAG_WIDTH = 1 parameter REQ_INFO_WIDTH = 1
) ( ) (
`SCOPE_IO_VX_lsu_unit
input wire clk, input wire clk,
input wire reset, input wire reset,
@@ -14,102 +12,66 @@ module VX_tex_memory #(
VX_dcache_core_rsp_if dcache_rsp_if, VX_dcache_core_rsp_if dcache_rsp_if,
// inputs // inputs
input wire req_valid, input wire req_valid,
input wire [`NUM_THREADS-1:0] req_tmask, input wire [`NW_BITS-1:0] req_wid,
input wire [`TEX_FILTER_BITS-1:0] req_filter, input wire [`NUM_THREADS-1:0] req_tmask,
input wire [31:0] req_PC,
input wire [`TEX_FILTER_BITS-1:0] req_filter,
input wire [`TEX_STRIDE_BITS-1:0] req_stride,
input wire [`NUM_THREADS-1:0][3:0][31:0] req_addr, input wire [`NUM_THREADS-1:0][3:0][31:0] req_addr,
input wire [REQ_TAG_WIDTH-1:0] req_tag, input wire [REQ_INFO_WIDTH-1:0] req_info,
output wire req_ready, output wire req_ready,
// outputs // outputs
output wire rsp_valid, output wire rsp_valid,
output wire [`NUM_THREADS-1:0] rsp_tmask, output wire [`NW_BITS-1:0] rsp_wid,
output wire [`TEX_FILTER_BITS-1:0] rsp_filter, output wire [`NUM_THREADS-1:0] rsp_tmask,
output wire [31:0] rsp_PC,
output wire [`TEX_FILTER_BITS-1:0] rsp_filter,
output wire [`NUM_THREADS-1:0][3:0][31:0] rsp_data, output wire [`NUM_THREADS-1:0][3:0][31:0] rsp_data,
output wire [REQ_TAG_WIDTH-1:0] rsp_tag, output wire [REQ_INFO_WIDTH-1:0] rsp_info,
input wire rsp_ready input wire rsp_ready
); );
`UNUSED_PARAM (CORE_ID) `UNUSED_PARAM (CORE_ID)
/*wire req_valid;
wire [`NUM_THREADS-1:0] req_tmask;
wire [`NUM_THREADS-1:0][31:0] req_addr;
wire [`LSU_BITS-1:0] req_type;
wire [`NUM_THREADS-1:0][31:0] req_data;
wire [`NR_BITS-1:0] req_rd;
wire req_wb;
wire [`NW_BITS-1:0] req_wid;
wire [31:0] req_pc;
wire req_is_dup;
wire [`NUM_THREADS-1:0][31:0] full_address; wire [3:0] dup_reqs;
for (genvar i = 0; i < `NUM_THREADS; i++) begin wire [3:0][`NUM_THREADS-1:0][29:0] req_addr_w;
assign full_address[i] = lsu_req_if.base_addr[i] + lsu_req_if.offset; wire [3:0][`NUM_THREADS-1:0][1:0] align_offs;
// reorder address into quads
for (genvar i = 0; i < `NUM_THREADS; ++i) begin
for (genvar j = 0; j < 4; ++j) begin
assign req_addr_w[j][i] = req_addr[i][j][31:2];
assign align_offs[j][i] = req_addr[i][j][1:0];
end
end end
wire [`NUM_THREADS-1:0] addr_matches; // find duplicate addresses
for (genvar i = 0; i < `NUM_THREADS; i++) begin
assign addr_matches[i] = (full_address[0][31:2] == full_address[i][31:2]) || ~lsu_req_if.tmask[i];
end
wire is_dup_load = lsu_req_if.wb && lsu_req_if.tmask[0] && (& addr_matches);
`IGNORE_WARNINGS_BEGIN
reg [`LSUQ_SIZE-1:0][`DCORE_TAG_ID_BITS-1:0] pending_tags;
`IGNORE_WARNINGS_END
wire ready_in; for (genvar i = 0; i < 4; ++i) begin
wire stall_in = ~ready_in && req_valid; wire [`NUM_THREADS-1:0] addr_matches;
for (genvar j = 0; j < `NUM_THREADS; j++) begin
VX_pipe_register #( assign addr_matches[j] = (req_addr_w[i][0] == req_addr_w[i][j]) || ~req_tmask[j];
.DATAW (1 + 1 + `NW_BITS + `NUM_THREADS + 32 + (`NUM_THREADS * 32) + `LSU_BITS + `NR_BITS + 1 + (`NUM_THREADS * 32)), end
.RESETW (1) assign dup_reqs[i] = req_tmask[0] && (& addr_matches);
) req_pipe_reg (
.clk (clk),
.reset (reset),
.enable (!stall_in),
.data_in ({lsu_req_if.valid, is_dup_load, lsu_req_if.wid, lsu_req_if.tmask, lsu_req_if.PC, full_address, lsu_req_if.op_type, lsu_req_if.rd, lsu_req_if.wb, lsu_req_if.store_data}),
.data_out ({req_valid, req_is_dup, req_wid, req_tmask, req_pc, req_addr, req_type, req_rd, req_wb, req_data})
);
// Can accept new request?
assign lsu_req_if.ready = ~stall_in;
wire [`NW_BITS-1:0] rsp_wid;
wire [31:0] rsp_pc;
wire [`NR_BITS-1:0] rsp_rd;
wire rsp_wb;
wire [`LSU_BITS-1:0] rsp_type;
wire rsp_is_dup;
`UNUSED_VAR (rsp_type)
reg [`LSUQ_SIZE-1:0][`NUM_THREADS-1:0] rsp_rem_mask;
reg [`NUM_THREADS-1:0] rsp_rem_mask_n;
reg [`NUM_THREADS-1:0] req_sent_mask;
wire req_sent_all;
wire [`DCORE_TAG_ID_BITS-1:0] mbuf_waddr, mbuf_raddr;
wire mbuf_full;
wire [`NUM_THREADS-1:0][1:0] req_offset, rsp_offset;
for (genvar i = 0; i < `NUM_THREADS; i++) begin
assign req_offset[i] = req_addr[i][1:0];
end end
wire mbuf_push = (| (dcache_req_if.valid & dcache_req_if.ready)) // save requet metadata into index buffer
&& (0 == req_sent_mask) // first submission only
&& req_wb; // loads only
wire mbuf_pop_part = (| dcache_rsp_if.valid) && dcache_rsp_if.ready; wire [`LSUQ_ADDR_BITS-1:0] mbuf_waddr, mbuf_raddr;
wire mbuf_push, mbuf_pop, mbuf_full;
wire mbuf_pop = mbuf_pop_part && (rsp_rem_mask_n == 0 || rsp_is_dup); wire [REQ_INFO_WIDTH-1:0] ib_req_info;
wire [`TEX_FILTER_BITS-1:0] ib_req_filter;
assign mbuf_raddr = dcache_rsp_if.tag[`DCORE_TAG_ID_BITS-1:0]; wire [`TEX_STRIDE_BITS-1:0] ib_stride;
wire [3:0][`NUM_THREADS-1:0][1:0] ib_align_offs;
wire [3:0] ib_dup_reqs;
assign mbuf_push = req_valid && req_ready;
VX_index_buffer #( VX_index_buffer #(
.DATAW (`NW_BITS + 32 + `NR_BITS + 1 + `LSU_BITS + (`NUM_THREADS * 2) + 1), .DATAW (REQ_INFO_WIDTH + `TEX_FILTER_BITS + `TEX_STRIDE_BITS + (4 * `NUM_THREADS * 2) + 4),
.SIZE (`LSUQ_SIZE) .SIZE (`LSUQ_SIZE)
) req_metadata ( ) req_metadata (
.clk (clk), .clk (clk),
@@ -117,189 +79,135 @@ module VX_tex_memory #(
.write_addr (mbuf_waddr), .write_addr (mbuf_waddr),
.acquire_slot (mbuf_push), .acquire_slot (mbuf_push),
.read_addr (mbuf_raddr), .read_addr (mbuf_raddr),
.write_data ({req_wid, req_pc, req_rd, req_wb, req_type, req_offset, req_is_dup}), .write_data ({req_info, req_filter, req_stride, align_offs, dup_reqs}),
.read_data ({rsp_wid, rsp_pc, rsp_rd, rsp_wb, rsp_type, rsp_offset, rsp_is_dup}), .read_data ({ib_req_info, ib_req_filter, ib_stride, ib_align_offs, ib_dup_reqs}),
.release_addr (mbuf_raddr), .release_addr (mbuf_raddr),
.release_slot (mbuf_pop), .release_slot (mbuf_pop),
.full (mbuf_full) .full (mbuf_full)
); );
assign req_sent_all = (&(dcache_req_if.ready | req_sent_mask | ~req_tmask)) // can take more requests?
|| (req_is_dup && dcache_req_if.ready[0]); assign req_ready = ~mbuf_full;
// save request addresses into fifo
wire reqq_empty, reqq_full;
wire reqq_push, reqq_pop;
wire [3:0][`NUM_THREADS-1:0][29:0] q_req_addr;
wire [`LSUQ_ADDR_BITS-1:0] q_ib_waddr;
wire [`NW_BITS-1:0] q_req_wid;
wire [`NUM_THREADS-1:0] q_req_tmask;
wire [31:0] q_req_PC;
wire [`TEX_FILTER_BITS-1:0] q_req_filter;
wire [3:0] q_dup_reqs;
assign reqq_push = mbuf_push;
VX_fifo_queue #(
.DATAW (`NUM_THREADS * 4 * 30 + `LSUQ_ADDR_BITS + `NW_BITS + `NUM_THREADS + 32 + `TEX_FILTER_BITS + 4),
.SIZE (`LSUQ_SIZE),
.BUFFERED (1)
) req_queue (
.clk (clk),
.reset (reset),
.push (reqq_push),
.pop (reqq_pop),
.data_in ({req_addr_w, mbuf_waddr, req_wid, req_tmask, req_PC, req_filter, dup_reqs}),
.data_out ({q_req_addr, q_ib_waddr, q_req_wid, q_req_tmask, q_req_PC, q_req_filter, q_dup_reqs}),
.empty (reqq_empty),
`UNUSED_PIN (full),
`UNUSED_PIN (alm_full),
`UNUSED_PIN (alm_empty),
`UNUSED_PIN (size)
);
///////////////////////////////////////////////////////////////////////////
wire [`NUM_THREADS-1:0][29:0] texel_addr;
wire texel_valid, texel_sent, last_texel_sent;
wire texel_is_dup;
reg [1:0] texel_idx;
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset || last_texel_sent) begin
req_sent_mask <= 0; texel_idx <= 0;
end else begin end else if (texel_sent) begin
if (req_sent_all) texel_idx <= texel_idx + 1;
req_sent_mask <= 0;
else
req_sent_mask <= req_sent_mask | (dcache_req_if.valid & dcache_req_if.ready);
end
end
// need to hold the acquired tag index until the full request is submitted
reg [`DCORE_TAG_ID_BITS-1:0] req_tag_hold;
wire [`DCORE_TAG_ID_BITS-1:0] req_tag = (0 == req_sent_mask) ? mbuf_waddr : req_tag_hold;
always @(posedge clk) begin
if (mbuf_push)
req_tag_hold <= mbuf_waddr;
end
assign rsp_rem_mask_n = rsp_rem_mask[mbuf_raddr] & ~dcache_rsp_if.valid;
always @(posedge clk) begin
if (mbuf_push) begin
rsp_rem_mask[mbuf_waddr] <= req_tmask;
pending_tags[mbuf_waddr] <= req_tag;
end
if (mbuf_pop_part) begin
rsp_rem_mask[mbuf_raddr] <= rsp_rem_mask_n;
end end
end end
// wire req_ready_dep = (req_wb && ~mbuf_full) || (~req_wb && st_commit_if.ready); assign texel_valid = ~reqq_empty;
wire req_ready_dep = (req_wb && ~mbuf_full); assign texel_addr = q_req_addr[texel_idx];
assign texel_is_dup = q_dup_reqs[texel_idx];
wire [`NUM_THREADS-1:0] dup_mask = {{(`NUM_THREADS-1){~req_is_dup}}, 1'b1}; wire is_last_texel = (texel_idx == (q_req_filter ? 3 : 0));
assign last_texel_sent = texel_sent && is_last_texel;
assign reqq_pop = last_texel_sent;
// DCache Request // DCache Request
reg [`NUM_THREADS-1:0][29:0] mem_req_addr; reg [`NUM_THREADS-1:0] texel_sent_mask;
reg [`NUM_THREADS-1:0][3:0] mem_req_byteen; wire [`NUM_THREADS-1:0] dcache_req_fire;
reg [`NUM_THREADS-1:0][31:0] mem_req_data;
always @(*) begin assign dcache_req_fire = dcache_req_if.valid & dcache_req_if.ready;
for (integer i = 0; i < `NUM_THREADS; i++) begin
mem_req_byteen[i] = {4{req_wb}};
case (`LSU_WSIZE(req_type))
0: mem_req_byteen[i][req_offset[i]] = 1;
1: begin
mem_req_byteen[i][req_offset[i]] = 1;
mem_req_byteen[i][{req_addr[i][1], 1'b1}] = 1;
end
default : mem_req_byteen[i] = {4{1'b1}};
endcase
mem_req_data[i] = 'x; assign texel_sent = (&(dcache_req_fire | texel_sent_mask | ~q_req_tmask))
case (req_offset[i]) || (texel_is_dup & dcache_req_if.valid[0] & dcache_req_if.ready[0]);
1: mem_req_data[i][31:8] = req_data[i][23:0];
2: mem_req_data[i][31:16] = req_data[i][15:0];
3: mem_req_data[i][31:24] = req_data[i][7:0];
default: mem_req_data[i] = req_data[i];
endcase
mem_req_addr[i] = req_addr[i][31:2]; always @(posedge clk) begin
if (reset) begin
texel_sent_mask <= 0;
end else begin
if (texel_sent)
texel_sent_mask <= 0;
else
texel_sent_mask <= texel_sent_mask | (dcache_req_if.valid & dcache_req_if.ready);
end end
end end
assign dcache_req_if.valid = {`NUM_THREADS{req_valid && req_ready_dep}} & req_tmask & dup_mask & ~req_sent_mask; wire [`NUM_THREADS-1:0] dup_mask = {{(`NUM_THREADS-1){~texel_is_dup}}, 1'b1};
assign dcache_req_if.rw = {`NUM_THREADS{~req_wb}};
assign dcache_req_if.addr = mem_req_addr; assign dcache_req_if.valid = {`NUM_THREADS{texel_valid}} & q_req_tmask & dup_mask & ~texel_sent_mask;
assign dcache_req_if.byteen = mem_req_byteen; assign dcache_req_if.rw = {`NUM_THREADS{1'b0}};
assign dcache_req_if.data = mem_req_data; assign dcache_req_if.addr = texel_addr;
assign dcache_req_if.byteen = {`NUM_THREADS{4'b1111}};
assign dcache_req_if.data = 'x;
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
assign dcache_req_if.tag = {`NUM_THREADS{{req_pc, req_wid, req_tag}}}; assign dcache_req_if.tag = {`NUM_THREADS{q_req_PC, q_req_wid, texel_idx, q_ib_waddr}};
`else `else
assign dcache_req_if.tag = {`NUM_THREADS{req_tag}}; assign dcache_req_if.tag = {`NUM_THREADS{q_ib_waddr}};
`endif `endif
assign ready_in = req_ready_dep && req_sent_all;
// send store commit // Dcache Response
//wire is_store_rsp = req_valid && ~req_wb && req_sent_all; reg [3:0][`NUM_THREADS-1:0][31:0] rsp_texels;
reg [`LSUQ_SIZE-1:0][3:0][`NUM_THREADS-1:0] rsp_rem_mask;
wire dcache_rsp_fire;
wire [1:0] rsp_texel_idx;
wire rsp_is_dup;
// assign st_commit_if.valid = is_store_rsp; assign dcache_rsp_fire = (| dcache_rsp_if.valid) && dcache_rsp_if.ready;
// assign st_commit_if.wid = req_wid;
// assign st_commit_if.tmask = req_tmask;
// assign st_commit_if.PC = req_pc;
// assign st_commit_if.rd = 0;
// assign st_commit_if.wb = 0;
// assign st_commit_if.eop = 1'b1;
// assign st_commit_if.data = 0;
// load response formatting wire [`NUM_THREADS-1:0] rsp_rem_mask_n = rsp_rem_mask[mbuf_raddr][rsp_texel_idx] & ~dcache_rsp_if.valid;
always @(posedge clk) begin
reg [`NUM_THREADS-1:0][31:0] rsp_data; if ((|dcache_req_fire) && (0 == texel_sent_mask)) begin
wire [`NUM_THREADS-1:0] rsp_tmask; rsp_rem_mask[q_ib_waddr][rsp_texel_idx] <= q_req_tmask;
end
for (genvar i = 0; i < `NUM_THREADS; i++) begin if (dcache_rsp_fire) begin
wire [31:0] src_data = (i == 0 || rsp_is_dup) ? dcache_rsp_if.data[0] : dcache_rsp_if.data[i]; rsp_rem_mask[mbuf_raddr][rsp_texel_idx] <= rsp_rem_mask_n;
reg [31:0] rsp_data_shifted;
always @(*) begin
rsp_data_shifted[31:16] = src_data[31:16];
rsp_data_shifted[15:0] = rsp_offset[i][1] ? src_data[31:16] : src_data[15:0];
rsp_data_shifted[7:0] = rsp_offset[i][0] ? rsp_data_shifted[15:8] : rsp_data_shifted[7:0];
end
always @(*) begin
case (`LSU_FMT(rsp_type))
`FMT_B: rsp_data[i] = 32'(signed'(rsp_data_shifted[7:0]));
`FMT_H: rsp_data[i] = 32'(signed'(rsp_data_shifted[15:0]));
`FMT_BU: rsp_data[i] = 32'(unsigned'(rsp_data_shifted[7:0]));
`FMT_HU: rsp_data[i] = 32'(unsigned'(rsp_data_shifted[15:0]));
default: rsp_data[i] = rsp_data_shifted;
endcase
end
end
assign rsp_tmask = rsp_is_dup ? rsp_rem_mask[mbuf_raddr] : dcache_rsp_if.valid;
// send load commit
wire load_rsp_stall = ~ld_commit_if.ready && ld_commit_if.valid;
VX_pipe_register #(
.DATAW (1 + `NW_BITS + `NUM_THREADS + 32 + `NR_BITS + 1 + (`NUM_THREADS * 32) + 1),
.RESETW (1)
) rsp_pipe_reg (
.clk (clk),
.reset (reset),
.enable (!load_rsp_stall),
.data_in ({(| dcache_rsp_if.valid), rsp_wid, rsp_tmask, rsp_pc, rsp_rd, rsp_wb, rsp_data, mbuf_pop}),
.data_out ({ld_commit_if.valid, ld_commit_if.wid, ld_commit_if.tmask, ld_commit_if.PC, ld_commit_if.rd, ld_commit_if.wb, ld_commit_if.data, ld_commit_if.eop})
);
// Can accept new cache response?
assign dcache_rsp_if.ready = ~load_rsp_stall;
// scope registration
`SCOPE_ASSIGN (dcache_req_fire, dcache_req_if.valid & dcache_req_if.ready);
`SCOPE_ASSIGN (dcache_req_wid, req_wid);
`SCOPE_ASSIGN (dcache_req_pc, req_pc);
`SCOPE_ASSIGN (dcache_req_addr, req_addr);
`SCOPE_ASSIGN (dcache_req_rw, ~req_wb);
`SCOPE_ASSIGN (dcache_req_byteen,dcache_req_if.byteen);
`SCOPE_ASSIGN (dcache_req_data, dcache_req_if.data);
`SCOPE_ASSIGN (dcache_req_tag, req_tag);
`SCOPE_ASSIGN (dcache_rsp_fire, dcache_rsp_if.valid & {`NUM_THREADS{dcache_rsp_if.ready}});
`SCOPE_ASSIGN (dcache_rsp_data, dcache_rsp_if.data);
`SCOPE_ASSIGN (dcache_rsp_tag, mbuf_raddr);
`ifdef DBG_PRINT_CORE_DCACHE
always @(posedge clk) begin
if ((| (dcache_req_if.valid & dcache_req_if.ready))) begin
if ((| dcache_req_if.rw))
$display("%t: D$%0d Wr Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, data=%0h",
$time, CORE_ID, req_wid, req_pc, (dcache_req_if.valid & dcache_req_if.ready), req_addr, dcache_req_if.tag, dcache_req_if.byteen, dcache_req_if.data);
else
$display("%t: D$%0d Rd Req: wid=%0d, PC=%0h, tmask=%b, addr=%0h, tag=%0h, byteen=%0h, rd=%0d, is_dup=%b",
$time, CORE_ID, req_wid, req_pc, (dcache_req_if.valid & dcache_req_if.ready), req_addr, dcache_req_if.tag, dcache_req_if.byteen, req_rd, req_is_dup);
end
if ((| dcache_rsp_if.valid) && dcache_rsp_if.ready) begin
$display("%t: D$%0d Rsp: valid=%b, wid=%0d, PC=%0h, tag=%0h, rd=%0d, data=%0h, is_dup=%b",
$time, CORE_ID, dcache_rsp_if.valid, rsp_wid, rsp_pc, dcache_rsp_if.tag, rsp_rd, dcache_rsp_if.data, rsp_is_dup);
end
if (mbuf_full) begin
$write("%t: D$%0d queue-full:", $time, CORE_ID);
for (integer j = 0; j < `LSUQ_SIZE; j++) begin
$write(" tag%0d=%0h", j, pending_tags[j]);
end
$write("\n");
end end
end end
`endif*/
assign mbuf_raddr = dcache_rsp_if.tag[`LSUQ_ADDR_BITS-1:0];
assign rsp_texel_idx = dcache_rsp_if.tag[`LSUQ_ADDR_BITS-1+:2];
assign rsp_is_dup = ib_dup_reqs[rsp_texel_idx];
assign rsp_tmask = rsp_is_dup ? rsp_rem_mask[mbuf_raddr][rsp_texel_idx]: dcache_rsp_if.valid;
assign mbuf_pop = dcache_rsp_fire && (0 == rsp_rem_mask_n || rsp_is_dup);
endmodule endmodule

View File

@@ -18,8 +18,8 @@ module VX_tex_unit #(
VX_tex_rsp_if tex_rsp_if VX_tex_rsp_if tex_rsp_if
); );
localparam REQ_TAG_WIDTH_A = `TEX_FORMAT_BITS + `NW_BITS + 32 + `NR_BITS + 1; localparam REQ_INFO_WIDTH_A = `TEX_FORMAT_BITS + `NR_BITS + 1;
localparam REQ_TAG_WIDTH_M = (2 * `NUM_THREADS * `FIXED_FRAC) + REQ_TAG_WIDTH_A; localparam REQ_INFO_WIDTH_M = (2 * `NUM_THREADS * `FIXED_FRAC) + REQ_INFO_WIDTH_A;
`UNUSED_PARAM (CORE_ID) `UNUSED_PARAM (CORE_ID)
`UNUSED_VAR (reset) `UNUSED_VAR (reset)
@@ -69,23 +69,28 @@ module VX_tex_unit #(
// address generation // address generation
wire mem_req_valid; wire mem_req_valid;
wire [`NW_BITS-1:0] mem_req_wid;
wire [`NUM_THREADS-1:0] mem_req_tmask; wire [`NUM_THREADS-1:0] mem_req_tmask;
wire [31:0] mem_req_PC;
wire [`TEX_FILTER_BITS-1:0] mem_req_filter; wire [`TEX_FILTER_BITS-1:0] mem_req_filter;
wire [`TEX_STRIDE_BITS-1:0] mem_req_stride;
wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_u; wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_u;
wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_v; wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] mem_req_v;
wire [`NUM_THREADS-1:0][3:0][31:0] mem_req_addr; wire [`NUM_THREADS-1:0][3:0][31:0] mem_req_addr;
wire [REQ_TAG_WIDTH_A-1:0] mem_req_tag; wire [REQ_INFO_WIDTH_A-1:0] mem_req_info;
wire mem_req_ready; wire mem_req_ready;
wire mem_rsp_valid; wire mem_rsp_valid;
wire [`NW_BITS-1:0] mem_rsp_wid;
wire [`NUM_THREADS-1:0] mem_rsp_tmask; wire [`NUM_THREADS-1:0] mem_rsp_tmask;
wire [31:0] mem_rsp_PC;
wire [`TEX_FILTER_BITS-1:0] mem_rsp_filter; wire [`TEX_FILTER_BITS-1:0] mem_rsp_filter;
wire [`NUM_THREADS-1:0][3:0][31:0] mem_rsp_data; wire [`NUM_THREADS-1:0][3:0][31:0] mem_rsp_data;
wire [REQ_TAG_WIDTH_M-1:0] mem_rsp_tag; wire [REQ_INFO_WIDTH_M-1:0] mem_rsp_info;
wire mem_rsp_ready; wire mem_rsp_ready;
VX_tex_addr_gen #( VX_tex_addr_gen #(
.REQ_TAG_WIDTH (REQ_TAG_WIDTH_A) .REQ_INFO_WIDTH (REQ_INFO_WIDTH_A)
) tex_addr_gen ( ) tex_addr_gen (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -93,27 +98,33 @@ module VX_tex_unit #(
.valid_in (tex_req_if.valid), .valid_in (tex_req_if.valid),
.ready_in (tex_req_if.ready), .ready_in (tex_req_if.ready),
.req_wid (tex_req_if.wid),
.req_tmask (tex_req_if.tmask),
.req_PC (tex_req_if.PC),
.req_info ({tex_format[tex_req_if.unit], tex_req_if.rd, tex_req_if.wb}),
.filter (tex_filter[tex_req_if.unit]), .filter (tex_filter[tex_req_if.unit]),
.wrap_u (tex_wrap_u[tex_req_if.unit]), .wrap_u (tex_wrap_u[tex_req_if.unit]),
.wrap_v (tex_wrap_v[tex_req_if.unit]), .wrap_v (tex_wrap_v[tex_req_if.unit]),
.req_tmask (tex_req_if.tmask),
.req_tag ({tex_format[tex_req_if.unit], tex_req_if.wid, tex_req_if.PC, tex_req_if.rd, tex_req_if.wb}),
.base_addr (tex_addr[tex_req_if.unit]), .base_addr (tex_addr[tex_req_if.unit]),
.log2_stride (tex_stride[tex_req_if.unit]), .log_stride (tex_stride[tex_req_if.unit]),
.log2_width (tex_width[tex_req_if.unit]), .log_width (tex_width[tex_req_if.unit]),
.log2_height (tex_height[tex_req_if.unit]), .log_height (tex_height[tex_req_if.unit]),
.coord_u (tex_req_if.u), .coord_u (tex_req_if.u),
.coord_v (tex_req_if.v), .coord_v (tex_req_if.v),
.lod (tex_req_if.lod), .lod (tex_req_if.lod),
.mem_req_valid (mem_req_valid), .mem_req_valid (mem_req_valid),
.mem_req_tmask (mem_req_tmask), .mem_req_wid (mem_req_wid),
.mem_req_tmask (mem_req_tmask),
.mem_req_PC (mem_req_PC),
.mem_req_filter (mem_req_filter), .mem_req_filter (mem_req_filter),
.mem_req_stride (mem_req_stride),
.mem_req_u (mem_req_u), .mem_req_u (mem_req_u),
.mem_req_v (mem_req_v), .mem_req_v (mem_req_v),
.mem_req_tag (mem_req_tag), .mem_req_info (mem_req_info),
.mem_req_addr (mem_req_addr), .mem_req_addr (mem_req_addr),
.mem_req_ready (mem_req_ready) .mem_req_ready (mem_req_ready)
); );
@@ -121,8 +132,8 @@ module VX_tex_unit #(
// retrieve texel values from memory // retrieve texel values from memory
VX_tex_memory #( VX_tex_memory #(
.CORE_ID (CORE_ID), .CORE_ID (CORE_ID),
.REQ_TAG_WIDTH (REQ_TAG_WIDTH_M) .REQ_INFO_WIDTH (REQ_INFO_WIDTH_M)
) tex_memory ( ) tex_memory (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -133,18 +144,23 @@ module VX_tex_unit #(
// inputs // inputs
.req_valid (mem_req_valid), .req_valid (mem_req_valid),
.req_wid (mem_req_wid),
.req_tmask (mem_req_tmask), .req_tmask (mem_req_tmask),
.req_PC (mem_req_PC),
.req_filter(mem_req_filter), .req_filter(mem_req_filter),
.req_stride(mem_req_stride),
.req_addr (mem_req_addr), .req_addr (mem_req_addr),
.req_tag ({mem_req_u, mem_req_v, mem_req_tag}), .req_info ({mem_req_u, mem_req_v, mem_req_info}),
.req_ready (mem_req_ready), .req_ready (mem_req_ready),
// outputs // outputs
.rsp_valid (mem_rsp_valid), .rsp_valid (mem_rsp_valid),
.rsp_wid (mem_rsp_wid),
.rsp_tmask (mem_rsp_tmask), .rsp_tmask (mem_rsp_tmask),
.rsp_PC (mem_rsp_PC),
.rsp_filter(mem_rsp_filter), .rsp_filter(mem_rsp_filter),
.rsp_data (mem_rsp_data), .rsp_data (mem_rsp_data),
.rsp_tag (mem_rsp_tag), .rsp_info (mem_rsp_info),
.rsp_ready (mem_rsp_ready) .rsp_ready (mem_rsp_ready)
); );
@@ -153,12 +169,10 @@ module VX_tex_unit #(
wire [`TEX_FORMAT_BITS-1:0] rsp_format; wire [`TEX_FORMAT_BITS-1:0] rsp_format;
wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] rsp_u; wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] rsp_u;
wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] rsp_v; wire [`NUM_THREADS-1:0][`FIXED_FRAC-1:0] rsp_v;
wire [`NW_BITS-1:0] rsp_wid;
wire [31:0] rsp_PC;
wire [`NR_BITS-1:0] rsp_rd; wire [`NR_BITS-1:0] rsp_rd;
wire rsp_wb; wire rsp_wb;
assign {rsp_format, rsp_u, rsp_v, rsp_wid, rsp_PC, rsp_rd, rsp_wb} = mem_rsp_tag; assign {rsp_format, rsp_u, rsp_v, rsp_rd, rsp_wb} = mem_rsp_info;
VX_tex_sampler #( VX_tex_sampler #(
.CORE_ID (CORE_ID) .CORE_ID (CORE_ID)
@@ -168,14 +182,14 @@ module VX_tex_unit #(
// inputs // inputs
.req_valid (mem_rsp_valid), .req_valid (mem_rsp_valid),
.req_wid (mem_rsp_wid),
.req_tmask (mem_rsp_tmask), .req_tmask (mem_rsp_tmask),
.req_PC (mem_rsp_PC),
.req_texels (mem_rsp_data), .req_texels (mem_rsp_data),
.req_filter (mem_rsp_filter), .req_filter (mem_rsp_filter),
.req_format (rsp_format), .req_format (rsp_format),
.req_u (rsp_u), .req_u (rsp_u),
.req_v (rsp_v), .req_v (rsp_v),
.req_wid (rsp_wid),
.req_PC (rsp_PC),
.req_rd (rsp_rd), .req_rd (rsp_rd),
.req_wb (rsp_wb), .req_wb (rsp_wb),
.req_ready (mem_rsp_ready), .req_ready (mem_rsp_ready),