texture unit hardware optimizations
This commit is contained in:
@@ -17,6 +17,7 @@ module VX_tex_addr #(
|
||||
input wire [`TEX_FILTER_BITS-1:0] req_filter,
|
||||
input wire [1:0][`TEX_WRAP_BITS-1:0] req_wraps,
|
||||
input wire [`TEX_ADDR_BITS-1:0] req_baseaddr,
|
||||
input wire [NUM_REQS-1:0][`TEX_LOD_BITS-1:0] mip_level,
|
||||
input wire [NUM_REQS-1:0][`TEX_MIPOFF_BITS-1:0] req_mipoff,
|
||||
input wire [NUM_REQS-1:0][1:0][`TEX_LOD_BITS-1:0] req_logdims,
|
||||
input wire [REQ_INFOW-1:0] req_info,
|
||||
@@ -28,6 +29,7 @@ module VX_tex_addr #(
|
||||
output wire [NUM_REQS-1:0] rsp_tmask,
|
||||
output wire [`TEX_FILTER_BITS-1:0] rsp_filter,
|
||||
output wire [`TEX_LGSTRIDE_BITS-1:0] rsp_lgstride,
|
||||
output wire [NUM_REQS-1:0][31:0] rsp_baseaddr,
|
||||
output wire [NUM_REQS-1:0][3:0][31:0] rsp_addr,
|
||||
output wire [NUM_REQS-1:0][1:0][`TEX_BLEND_FRAC-1:0] rsp_blends,
|
||||
output wire [REQ_INFOW-1:0] rsp_info,
|
||||
@@ -38,6 +40,7 @@ module VX_tex_addr #(
|
||||
|
||||
localparam SHIFT_BITS = $clog2(`TEX_FXD_FRAC+1);
|
||||
localparam PITCH_BITS = `MAX(`TEX_LOD_BITS, `TEX_LGSTRIDE_BITS) + 1;
|
||||
localparam SCALED_DIM = `TEX_FXD_FRAC + `TEX_DIM_BITS;
|
||||
localparam SCALED_X_W = `TEX_DIM_BITS + `TEX_BLEND_FRAC;
|
||||
localparam OFFSET_U_W = `TEX_DIM_BITS + `TEX_LGSTRIDE_MAX;
|
||||
localparam OFFSET_V_W = `TEX_DIM_BITS + `TEX_DIM_BITS + `TEX_LGSTRIDE_MAX;
|
||||
@@ -69,7 +72,7 @@ module VX_tex_addr #(
|
||||
|
||||
for (genvar i = 0; i < NUM_REQS; ++i) begin
|
||||
for (genvar j = 0; j < 2; ++j) begin
|
||||
wire [`TEX_FXD_FRAC-1:0] delta = (`TEX_FXD_HALF >> req_logdims[i][j]);
|
||||
wire [`TEX_FXD_FRAC-1:0] delta = `TEX_FXD_FRAC'((SCALED_DIM'(`TEX_FXD_HALF) << mip_level[i]) >> req_logdims[i][j]);
|
||||
wire [`TEX_FXD_BITS-1:0] coord_lo = req_filter ? (req_coords[j][i] - `TEX_FXD_BITS'(delta)) : req_coords[j][i];
|
||||
wire [`TEX_FXD_BITS-1:0] coord_hi = req_filter ? (req_coords[j][i] + `TEX_FXD_BITS'(delta)) : req_coords[j][i];
|
||||
|
||||
@@ -89,14 +92,14 @@ module VX_tex_addr #(
|
||||
.coord_o (clamped_hi[i][j])
|
||||
);
|
||||
|
||||
assign dim_shift[i][j] = (`TEX_FXD_FRAC - `TEX_BLEND_FRAC - req_logdims[i][j]);
|
||||
assign dim_shift[i][j] = (`TEX_FXD_FRAC - `TEX_BLEND_FRAC - (req_logdims[i][j] - mip_level[i]));
|
||||
end
|
||||
assign log_pitch[i] = PITCH_BITS'(req_logdims[i][0]) + PITCH_BITS'(log_stride);
|
||||
assign mip_addr[i] = req_baseaddr + 32'(req_mipoff[i]);
|
||||
assign log_pitch[i] = PITCH_BITS'(req_logdims[i][0] - mip_level[i]) + PITCH_BITS'(log_stride);
|
||||
assign mip_addr[i] = req_baseaddr + `TEX_ADDR_BITS'(req_mipoff[i]);
|
||||
end
|
||||
|
||||
VX_pipe_register #(
|
||||
.DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_LGSTRIDE_BITS + REQ_INFOW + NUM_REQS * (PITCH_BITS + 2 * SHIFT_BITS + 32 + 2 * 2 * `TEX_FXD_FRAC)),
|
||||
.DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_LGSTRIDE_BITS + REQ_INFOW + NUM_REQS * (PITCH_BITS + 2 * SHIFT_BITS + `TEX_ADDR_BITS + 2 * 2 * `TEX_FXD_FRAC)),
|
||||
.RESETW (1)
|
||||
) pipe_reg0 (
|
||||
.clk (clk),
|
||||
@@ -114,8 +117,6 @@ module VX_tex_addr #(
|
||||
wire [NUM_REQS-1:0][OFFSET_U_W-1:0] offset_u_hi;
|
||||
wire [NUM_REQS-1:0][OFFSET_V_W-1:0] offset_v_lo;
|
||||
wire [NUM_REQS-1:0][OFFSET_V_W-1:0] offset_v_hi;
|
||||
wire [NUM_REQS-1:0][31:0] base_addr_lo;
|
||||
wire [NUM_REQS-1:0][31:0] base_addr_hi;
|
||||
wire [NUM_REQS-1:0][1:0][`TEX_BLEND_FRAC-1:0] blends;
|
||||
wire [NUM_REQS-1:0][3:0][31:0] addr;
|
||||
|
||||
@@ -134,26 +135,23 @@ module VX_tex_addr #(
|
||||
assign offset_v_lo[i] = OFFSET_V_W'(scaled_lo[i][1][`TEX_BLEND_FRAC +: `TEX_DIM_BITS]) << log_pitch_s0[i];
|
||||
assign offset_v_hi[i] = OFFSET_V_W'(scaled_hi[i][1][`TEX_BLEND_FRAC +: `TEX_DIM_BITS]) << log_pitch_s0[i];
|
||||
|
||||
assign base_addr_lo[i] = mip_addr_s0[i] + 32'(offset_v_lo[i]);
|
||||
assign base_addr_hi[i] = mip_addr_s0[i] + 32'(offset_v_hi[i]);
|
||||
|
||||
assign addr[i][0] = base_addr_lo[i] + 32'(offset_u_lo[i]);
|
||||
assign addr[i][1] = base_addr_lo[i] + 32'(offset_u_hi[i]);
|
||||
assign addr[i][2] = base_addr_hi[i] + 32'(offset_u_lo[i]);
|
||||
assign addr[i][3] = base_addr_hi[i] + 32'(offset_u_hi[i]);
|
||||
assign addr[i][0] = 32'(offset_v_lo[i]) + 32'(offset_u_lo[i]);
|
||||
assign addr[i][1] = 32'(offset_v_lo[i]) + 32'(offset_u_hi[i]);
|
||||
assign addr[i][2] = 32'(offset_v_hi[i]) + 32'(offset_u_lo[i]);
|
||||
assign addr[i][3] = 32'(offset_v_hi[i]) + 32'(offset_u_hi[i]);
|
||||
end
|
||||
|
||||
assign stall_out = rsp_valid && ~rsp_ready;
|
||||
|
||||
VX_pipe_register #(
|
||||
.DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_LGSTRIDE_BITS + (NUM_REQS * 4 * 32) + (2 * NUM_REQS * `TEX_BLEND_FRAC) + REQ_INFOW),
|
||||
.DATAW (1 + NUM_REQS + `TEX_FILTER_BITS + `TEX_LGSTRIDE_BITS + (NUM_REQS * 32) + (NUM_REQS * 4 * 32) + (2 * NUM_REQS * `TEX_BLEND_FRAC) + REQ_INFOW),
|
||||
.RESETW (1)
|
||||
) pipe_reg1 (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.enable (~stall_out),
|
||||
.data_in ({valid_s0, tmask_s0, filter_s0, log_stride_s0, addr, blends, req_info_s0}),
|
||||
.data_out ({rsp_valid, rsp_tmask, rsp_filter, rsp_lgstride, rsp_addr, rsp_blends, rsp_info})
|
||||
.data_in ({valid_s0, tmask_s0, filter_s0, log_stride_s0, mip_addr_s0, addr, blends, req_info_s0}),
|
||||
.data_out ({rsp_valid, rsp_tmask, rsp_filter, rsp_lgstride, rsp_baseaddr, rsp_addr, rsp_blends, rsp_info})
|
||||
);
|
||||
|
||||
assign req_ready = ~stall_out;
|
||||
@@ -176,6 +174,8 @@ module VX_tex_addr #(
|
||||
`TRACE_ARRAY2D(clamped_lo, 2, NUM_REQS);
|
||||
dpi_trace(", clamped_hi=");
|
||||
`TRACE_ARRAY2D(clamped_hi, 2, NUM_REQS);
|
||||
dpi_trace(", mip_addr=");
|
||||
`TRACE_ARRAY1D(mip_addr, NUM_REQS);
|
||||
dpi_trace("\n");
|
||||
end
|
||||
|
||||
@@ -192,10 +192,6 @@ module VX_tex_addr #(
|
||||
`TRACE_ARRAY1D(offset_v_lo, NUM_REQS);
|
||||
dpi_trace(", offset_v_hi=");
|
||||
`TRACE_ARRAY1D(offset_v_hi, NUM_REQS);
|
||||
dpi_trace(", base_addr_lo=");
|
||||
`TRACE_ARRAY1D(base_addr_lo, NUM_REQS);
|
||||
dpi_trace(", base_addr_hi=");
|
||||
`TRACE_ARRAY1D(base_addr_hi, NUM_REQS);
|
||||
dpi_trace("\n");
|
||||
end
|
||||
|
||||
|
||||
Reference in New Issue
Block a user