texture unit critical path optimization

This commit is contained in:
Blaise Tine
2021-08-05 02:47:26 -07:00
parent deed327890
commit 43ad188ccb
8 changed files with 100 additions and 63 deletions

View File

@@ -91,6 +91,8 @@
`define LTRIM(x, s) x[s-1:0] `define LTRIM(x, s) x[s-1:0]
`define ADDER_CARRY_WIDTH(x, y) `MAX(x, `MIN(x, y)+1);
`define PRINT_ARRAY1D(a, m) \ `define PRINT_ARRAY1D(a, m) \
$write("{"); \ $write("{"); \
for (integer i = (m-1); i >= 0; --i) begin \ for (integer i = (m-1); i >= 0; --i) begin \

View File

@@ -36,6 +36,8 @@ module VX_tex_addr #(
`UNUSED_PARAM (CORE_ID) `UNUSED_PARAM (CORE_ID)
localparam PITCH_BITS = `ADDER_CARRY_WIDTH(`TEX_DIM_BITS, `TEX_STRIDE_BITS);
wire valid_s0; wire valid_s0;
wire [NUM_REQS-1:0] tmask_s0; wire [NUM_REQS-1:0] tmask_s0;
wire [`TEX_FILTER_BITS-1:0] filter_s0; wire [`TEX_FILTER_BITS-1:0] filter_s0;
@@ -44,7 +46,8 @@ module VX_tex_addr #(
wire [NUM_REQS-1:0][1:0][`FIXED_FRAC-1:0] clamped_hi, clamped_hi_s0; wire [NUM_REQS-1:0][1:0][`FIXED_FRAC-1:0] clamped_hi, clamped_hi_s0;
wire [`TEX_STRIDE_BITS-1:0] log_stride, log_stride_s0; wire [`TEX_STRIDE_BITS-1:0] log_stride, log_stride_s0;
wire [NUM_REQS-1:0][31:0] mip_addr, mip_addr_s0; wire [NUM_REQS-1:0][31:0] mip_addr, mip_addr_s0;
wire [NUM_REQS-1:0][1:0][`TEX_DIM_BITS-1:0] log_dims_s0; wire [NUM_REQS-1:0][PITCH_BITS-1:0] log_pitch, log_pitch_s0;
wire [NUM_REQS-1:0][`TEX_DIM_BITS-1:0] log_height, log_height_s0;
wire stall_out; wire stall_out;
@@ -61,8 +64,9 @@ module VX_tex_addr #(
for (genvar i = 0; i < NUM_REQS; ++i) begin for (genvar i = 0; i < NUM_REQS; ++i) begin
for (genvar j = 0; j < 2; ++j) begin for (genvar j = 0; j < 2; ++j) begin
wire [31:0] coord_lo = req_filter ? (req_coords[j][i] - (`FIXED_HALF >> req_logdims[i][j])) : req_coords[j][i]; wire [`FIXED_FRAC-1:0] delta = (`FIXED_HALF >> req_logdims[i][j]);
wire [31:0] coord_hi = req_filter ? (req_coords[j][i] + (`FIXED_HALF >> req_logdims[i][j])) : req_coords[j][i]; wire [31:0] coord_lo = req_filter ? (req_coords[j][i] - 32'(delta)) : req_coords[j][i];
wire [31:0] coord_hi = req_filter ? (req_coords[j][i] + 32'(delta)) : req_coords[j][i];
VX_tex_wrap #( VX_tex_wrap #(
.CORE_ID (CORE_ID) .CORE_ID (CORE_ID)
@@ -79,41 +83,45 @@ module VX_tex_addr #(
.coord_i (coord_hi), .coord_i (coord_hi),
.coord_o (clamped_hi[i][j]) .coord_o (clamped_hi[i][j])
); );
end end
assign mip_addr[i] = req_baseaddr + 32'(req_mipoff[i]); assign log_pitch[i] = PITCH_BITS'(req_logdims[i][0]) + PITCH_BITS'(log_stride);
assign log_height[i] = req_logdims[i][1];
assign mip_addr[i] = req_baseaddr + 32'(req_mipoff[i]);
end end
VX_pipe_register #( VX_pipe_register #(
.DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_STRIDE_BITS + REQ_INFO_WIDTH + NUM_REQS * (2 * `TEX_DIM_BITS + 32 + 2 * 2 * `FIXED_FRAC)), .DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_STRIDE_BITS + REQ_INFO_WIDTH + NUM_REQS * (PITCH_BITS + `TEX_DIM_BITS + 32 + 2 * 2 * `FIXED_FRAC)),
.RESETW (1) .RESETW (1)
) pipe_reg0 ( ) pipe_reg0 (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.enable (~stall_out), .enable (~stall_out),
.data_in ({req_valid, req_tmask, req_filter, log_stride, req_info, req_logdims, mip_addr, clamped_lo, clamped_hi}), .data_in ({req_valid, req_tmask, req_filter, log_stride, req_info, log_pitch, log_height, mip_addr, clamped_lo, clamped_hi}),
.data_out ({valid_s0, tmask_s0, filter_s0, log_stride_s0, req_info_s0, log_dims_s0, mip_addr_s0, clamped_lo_s0, clamped_hi_s0}) .data_out ({valid_s0, tmask_s0, filter_s0, log_stride_s0, req_info_s0, log_pitch_s0, log_height_s0, mip_addr_s0, clamped_lo_s0, clamped_hi_s0})
); );
// addresses generation // addresses generation
wire [NUM_REQS-1:0][1:0][`FIXED_INT-1:0] scaled_lo; wire [NUM_REQS-1:0][(`FIXED_INT+`TEX_STRIDE_BITS)-1:0] scaled_u_lo, scaled_u_hi;
wire [NUM_REQS-1:0][1:0][`FIXED_INT-1:0] scaled_hi; wire [NUM_REQS-1:0][`FIXED_INT-1:0] scaled_v_lo, scaled_v_hi;
wire [NUM_REQS-1:0][1:0][`BLEND_FRAC-1:0] blends; wire [NUM_REQS-1:0][1:0][`BLEND_FRAC-1:0] blends;
wire [NUM_REQS-1:0][3:0][31:0] addr; wire [NUM_REQS-1:0][3:0][31:0] addr;
for (genvar i = 0; i < NUM_REQS; ++i) begin for (genvar i = 0; i < NUM_REQS; ++i) begin
for (genvar j = 0; j < 2; ++j) begin assign scaled_u_lo[i] = scale_to_pitch(clamped_lo_s0[i][0], log_pitch_s0[i]);
assign scaled_lo[i][j] = `FIXED_INT'(clamped_lo_s0[i][j] >> ((`FIXED_FRAC) - log_dims_s0[i][j])); assign scaled_u_hi[i] = scale_to_pitch(clamped_hi_s0[i][0], log_pitch_s0[i]);
assign scaled_hi[i][j] = `FIXED_INT'(clamped_hi_s0[i][j] >> ((`FIXED_FRAC) - log_dims_s0[i][j])); assign scaled_v_lo[i] = scale_to_height(clamped_lo_s0[i][1], log_height_s0[i]);
assign blends[i][j] = filter_s0 ? clamped_lo_s0[i][j][`BLEND_FRAC-1:0] : `BLEND_FRAC'(0); assign scaled_v_hi[i] = scale_to_height(clamped_hi_s0[i][1], log_height_s0[i]);
for (genvar j = 0; j < 2; ++j) begin
assign blends[i][j] = filter_s0 ? clamped_lo_s0[i][j][`BLEND_FRAC-1:0] : `BLEND_FRAC'(0);
end end
end end
for (genvar i = 0; i < NUM_REQS; ++i) begin for (genvar i = 0; i < NUM_REQS; ++i) begin
assign addr[i][0] = mip_addr_s0[i] + (32'(scaled_lo[i][0]) + (32'(scaled_lo[i][1]) << log_dims_s0[i][0])) << log_stride_s0; assign addr[i][0] = mip_addr_s0[i] + 32'(scaled_u_lo[i]) + (32'(scaled_v_lo[i]) << log_pitch_s0[i]);
assign addr[i][1] = mip_addr_s0[i] + (32'(scaled_hi[i][0]) + (32'(scaled_lo[i][1]) << log_dims_s0[i][0])) << log_stride_s0; assign addr[i][1] = mip_addr_s0[i] + 32'(scaled_u_hi[i]) + (32'(scaled_v_lo[i]) << log_pitch_s0[i]);
assign addr[i][2] = mip_addr_s0[i] + (32'(scaled_lo[i][0]) + (32'(scaled_hi[i][1]) << log_dims_s0[i][0])) << log_stride_s0; assign addr[i][2] = mip_addr_s0[i] + 32'(scaled_u_lo[i]) + (32'(scaled_v_hi[i]) << log_pitch_s0[i]);
assign addr[i][3] = mip_addr_s0[i] + (32'(scaled_hi[i][0]) + (32'(scaled_hi[i][1]) << log_dims_s0[i][0])) << log_stride_s0; assign addr[i][3] = mip_addr_s0[i] + 32'(scaled_u_hi[i]) + (32'(scaled_v_hi[i]) << log_pitch_s0[i]);
end end
assign stall_out = rsp_valid && ~rsp_ready; assign stall_out = rsp_valid && ~rsp_ready;
@@ -131,9 +139,10 @@ module VX_tex_addr #(
assign req_ready = ~stall_out; assign req_ready = ~stall_out;
`ifdef DBG_PRINT_TEX `ifdef DBG_PRINT_TEX
wire [`NW_BITS-1:0] rsp_wid; wire [`NW_BITS-1:0] rsp_wid;
wire [31:0] rsp_PC; wire [31:0] rsp_PC;
assign {rsp_wid, rsp_PC} = rsp_info[`NW_BITS+32-1:0]; assign {rsp_wid, rsp_PC} = rsp_info[`NW_BITS+32-1:0];
always @(posedge clk) begin always @(posedge clk) begin
@@ -146,4 +155,22 @@ module VX_tex_addr #(
end end
`endif `endif
function logic [(`FIXED_INT+`TEX_STRIDE_BITS)-1:0] scale_to_pitch (input logic [`FIXED_FRAC-1:0] src,
input logic [PITCH_BITS-1:0] dim);
`IGNORE_UNUSED_BEGIN
logic [(`FIXED_BITS+`TEX_STRIDE_BITS)-1:0] out;
`IGNORE_UNUSED_END
out = (`FIXED_BITS+`TEX_STRIDE_BITS)'(src) << dim;
return out[`FIXED_FRAC +: (`FIXED_INT+`TEX_STRIDE_BITS)];
endfunction
function logic [`FIXED_INT-1:0] scale_to_height (input logic [`FIXED_FRAC-1:0] src,
input logic [`TEX_DIM_BITS-1:0] dim);
`IGNORE_UNUSED_BEGIN
logic [`FIXED_BITS-1:0] out;
`IGNORE_UNUSED_END
out = `FIXED_BITS'(src) << dim;
return out[`FIXED_FRAC +: `FIXED_INT];
endfunction
endmodule endmodule

View File

@@ -3,11 +3,12 @@
`include "VX_define.vh" `include "VX_define.vh"
`define FIXED_FRAC 20 `define FIXED_BITS 32
`define FIXED_INT (32 - `FIXED_FRAC) `define FIXED_FRAC 20
`define FIXED_ONE (2 ** `FIXED_FRAC) `define FIXED_INT (`FIXED_BITS - `FIXED_FRAC)
`define FIXED_HALF (`FIXED_ONE >> 1) `define FIXED_ONE (2 ** `FIXED_FRAC)
`define FIXED_MASK (`FIXED_ONE - 1) `define FIXED_HALF (`FIXED_ONE >> 1)
`define FIXED_MASK (`FIXED_ONE - 1)
`define TEX_ADDR_BITS 32 `define TEX_ADDR_BITS 32
`define TEX_FORMAT_BITS 3 `define TEX_FORMAT_BITS 3

View File

@@ -13,6 +13,12 @@ module VX_tex_format #(
always @(*) begin always @(*) begin
case (format) case (format)
`TEX_FORMAT_R8G8B8A8: begin
texel_out_r[07:00] = texel_in[7:0];
texel_out_r[15:08] = texel_in[15:8];
texel_out_r[23:16] = texel_in[23:16];
texel_out_r[31:24] = texel_in[31:24];
end
`TEX_FORMAT_R5G6B5: begin `TEX_FORMAT_R5G6B5: begin
texel_out_r[07:00] = {texel_in[15:11], texel_in[15:13]}; texel_out_r[07:00] = {texel_in[15:11], texel_in[15:13]};
texel_out_r[15:08] = {texel_in[10:5], texel_in[10:9]}; texel_out_r[15:08] = {texel_in[10:5], texel_in[10:9]};
@@ -31,24 +37,18 @@ module VX_tex_format #(
texel_out_r[23:16] = texel_in[7:0]; texel_out_r[23:16] = texel_in[7:0];
texel_out_r[31:24] = texel_in[15:8]; texel_out_r[31:24] = texel_in[15:8];
end end
`TEX_FORMAT_A8: begin
texel_out_r[07:00] = 0;
texel_out_r[15:08] = 0;
texel_out_r[23:16] = 0;
texel_out_r[31:24] = texel_in[7:0];
end
`TEX_FORMAT_L8: begin `TEX_FORMAT_L8: begin
texel_out_r[07:00] = texel_in[7:0]; texel_out_r[07:00] = texel_in[7:0];
texel_out_r[15:08] = texel_in[7:0]; texel_out_r[15:08] = texel_in[7:0];
texel_out_r[23:16] = texel_in[7:0]; texel_out_r[23:16] = texel_in[7:0];
texel_out_r[31:24] = 8'hff; texel_out_r[31:24] = 8'hff;
end end
// `TEX_FORMAT_R8G8B8A8 //`TEX_FORMAT_A8
default: begin default: begin
texel_out_r[07:00] = texel_in[7:0]; texel_out_r[07:00] = 0;
texel_out_r[15:08] = texel_in[15:8]; texel_out_r[15:08] = 0;
texel_out_r[23:16] = texel_in[23:16]; texel_out_r[23:16] = 0;
texel_out_r[31:24] = texel_in[31:24]; texel_out_r[31:24] = texel_in[7:0];
end end
endcase endcase
end end

View File

@@ -1,17 +1,16 @@
`include "VX_tex_define.vh" `include "VX_tex_define.vh"
module VX_tex_lerp #( module VX_tex_lerp (
) ( input wire [3:0][7:0] in1,
input wire [`BLEND_FRAC-1:0] blend, input wire [3:0][7:0] in2,
input wire [31:0] in1, input wire [8:0] alpha,
input wire [31:0] in2, input wire [7:0] beta,
output wire [31:0] out output wire [3:0][7:0] out
); );
for (genvar i = 0; i < 4; ++i) begin for (genvar i = 0; i < 4; ++i) begin
wire [8:0] blend_m1 = `BLEND_ONE - blend; wire [16:0] sum = in1[i] * alpha + in2[i] * beta;
wire [16:0] sum = in1[i*8+:8] * blend_m1 + in2[i*8+:8] * blend;
`UNUSED_VAR (sum) `UNUSED_VAR (sum)
assign out[i*8+:8] = sum[15:8]; assign out[i] = sum[15:8];
end end
endmodule endmodule

View File

@@ -51,20 +51,25 @@ module VX_tex_sampler #(
); );
end end
wire [7:0] beta = req_blends[i][0];
wire [8:0] alpha = `BLEND_ONE - beta;
VX_tex_lerp #( VX_tex_lerp #(
) tex_lerp_ul ( ) tex_lerp_ul (
.blend (req_blends[i][0]), .in1 (fmt_texels[0]),
.in1 (fmt_texels[0]), .in2 (fmt_texels[1]),
.in2 (fmt_texels[1]), .alpha (alpha),
.out (texel_ul[i]) .beta (beta),
.out (texel_ul[i])
); );
VX_tex_lerp #( VX_tex_lerp #(
) tex_lerp_uh ( ) tex_lerp_uh (
.blend (req_blends[i][0]), .in1 (fmt_texels[2]),
.in1 (fmt_texels[2]), .in2 (fmt_texels[3]),
.in2 (fmt_texels[3]), .alpha (alpha),
.out (texel_uh[i]) .beta (beta),
.out (texel_uh[i])
); );
assign blend_v[i] = req_blends[i][1]; assign blend_v[i] = req_blends[i][1];
@@ -82,12 +87,16 @@ module VX_tex_sampler #(
); );
for (genvar i = 0; i < NUM_REQS; i++) begin for (genvar i = 0; i < NUM_REQS; i++) begin
wire [7:0] beta = blend_v_s0[i];
wire [8:0] alpha = `BLEND_ONE - beta;
VX_tex_lerp #( VX_tex_lerp #(
) tex_lerp_v ( ) tex_lerp_v (
.blend (blend_v_s0[i]), .in1 (texel_ul_s0[i]),
.in1 (texel_ul_s0[i]), .in2 (texel_uh_s0[i]),
.in2 (texel_uh_s0[i]), .alpha (alpha),
.out (texel_v[i]) .beta (beta),
.out (texel_v[i])
); );
end end
@@ -108,7 +117,6 @@ module VX_tex_sampler #(
assign req_ready = ~stall_out; assign req_ready = ~stall_out;
`ifdef DBG_PRINT_TEX `ifdef DBG_PRINT_TEX
wire [`NW_BITS-1:0] req_wid, rsp_wid; wire [`NW_BITS-1:0] req_wid, rsp_wid;
wire [31:0] req_PC, rsp_PC; wire [31:0] req_PC, rsp_PC;

View File

@@ -1,6 +1,6 @@
`include "VX_platform.vh" `include "VX_platform.vh"
module VX_sat_fx #( module VX_tex_sat #(
parameter IN_W = 1, parameter IN_W = 1,
parameter OUT_W = 1, parameter OUT_W = 1,
parameter MODEL = 1 parameter MODEL = 1
@@ -11,11 +11,11 @@ module VX_sat_fx #(
`STATIC_ASSERT(((OUT_W+1) < IN_W), ("invalid parameter")) `STATIC_ASSERT(((OUT_W+1) < IN_W), ("invalid parameter"))
if (MODEL == 1) begin if (MODEL == 1) begin
assign data_out = data_in[IN_W-1] ? OUT_W'(0) : ((data_in > {OUT_W{1'b1}}) ? {OUT_W{1'b1}} : OUT_W'(data_in));
end else begin
wire [OUT_W-1:0] underflow_mask = {OUT_W{~data_in[IN_W-1]}}; wire [OUT_W-1:0] underflow_mask = {OUT_W{~data_in[IN_W-1]}};
wire [OUT_W-1:0] overflow_mask = {OUT_W{(| data_in[IN_W-2:OUT_W])}}; wire [OUT_W-1:0] overflow_mask = {OUT_W{(| data_in[IN_W-2:OUT_W])}};
assign data_out = (data_in[OUT_W-1:0] | overflow_mask) & underflow_mask; assign data_out = (data_in[OUT_W-1:0] | overflow_mask) & underflow_mask;
end else begin
assign data_out = data_in[IN_W-1] ? OUT_W'(0) : ((data_in > {OUT_W{1'b1}}) ? {OUT_W{1'b1}} : OUT_W'(data_in));
end end
endmodule endmodule

View File

@@ -14,7 +14,7 @@ module VX_tex_wrap #(
wire [`FIXED_FRAC-1:0] clamp; wire [`FIXED_FRAC-1:0] clamp;
VX_sat_fx #( VX_tex_sat #(
.IN_W (32), .IN_W (32),
.OUT_W (`FIXED_FRAC) .OUT_W (`FIXED_FRAC)
) sat_fx ( ) sat_fx (