working more on Bayer version of the MCLT (632ab621) · Commits · Elphel / x393

dsp/mclt16x16.v

+2 −6

Original line number	Diff line number	Diff line
		@@ -453,11 +453,8 @@ D11 - negate for mode 3 (SS)

		// frequency domain, high address bit - page, 2 next - mode, 6 LSBs - transposed FD data (vertical first)
		wire [8:0] dtt_out_ram_wa = {dtt_out_ram_wah,dtt_out_wa16};


		reg [7:0] dtt_dly_cntr;
		reg [8:0] dtt_rd_cntr; // counter for dtt readout to rotator
		// wire [8:0] dtt_rd_ra = {dtt_rd_cntr[8],dtt_rd_cntr[1:0],dtt_rd_cntr[7:2]}; // page, mode, frequency
		wire [8:0] dtt_rd_ra = {dtt_rd_cntr[8],dtt_rd_cntr[0],dtt_rd_cntr[1],dtt_rd_cntr[7:2]}; // page, mode, frequency
		reg [2:0] dtt_rd_regen_dv; // dtt output buffer mem read, register enable, data valid
		wire [35:0] dtt_rd_data_w; // high bits are not used
		@@ -466,7 +463,6 @@ D11 - negate for mode 3 (SS)

		wire dtt_first_quad_out = ~dtt_out_ram_cntr[3] & ~dtt_out_ram_cntr[2];

		// reg [9:0] dtt_out_ram_cntr;
		always @(posedge clk) begin
		if (rst) dtt_out_ram_cntr <= 0;
		else if (dtt_inc16) dtt_out_ram_cntr <= dtt_out_ram_cntr + 1;
		@@ -507,7 +503,6 @@ D11 - negate for mode 3 (SS)
		.clk (clk), // input
		.rst (rst), // input
		.start (dtt_start), // input
		// .mode (dtt_mode), // input[1:0]
		.mode ({dtt_mode[0],dtt_mode[1]}), // input[1:0]
		.xin (dtt_r_data), // input[24:0] signed
		.pre_last_in (), // output reg
		@@ -555,6 +550,7 @@ D11 - negate for mode 3 (SS)
		// are these shift OK? Will need to be valis only @ dtt_start_out
		.shift_h (x_shft_r4), // input[6:0] signed
		.shift_v (y_shft_r4), // input[6:0] signed
		.inv_checker (1'b0), // input only used for Bayer mosaic data
		.fd_din (dtt_rd_data), // input[24:0] signed. Expected latency = 3 from start
		.fd_out (dout), // output[24:0] reg signed
		.pre_first_out (pre_first_out), // output reg

dsp/mclt16x16_bayer.v

+244 −15

Original line number	Diff line number	Diff line
		@@ -68,9 +68,18 @@ module mclt16x16_bayer#(
		output [PIX_ADDR_WIDTH-1:0] pix_addr, //!< external pixel buffer address
		output pix_re, //!< pixel read enable (sync with mpixel_a)
		output pix_page, //!< copy pixel page (should be externally combined with first color)
		input [PIXEL_WIDTH-1:0] pix_d //!< pixel data, latency = 2 from pixel address
		input [PIXEL_WIDTH-1:0] pix_d, //!< pixel data, latency = 2 from pixel address
		output pre_busy, //!< start should come each 256-th cycle (next after pre_last_in), and not after pre_busy)
		output pre_last_in, //!< may increment page
		output pre_first_out,//!< next will output first of DCT/DCT coefficients
		output pre_last_out, //!< next will be last output of DST/DST coefficients
		output [7:0] out_addr, //!< address to save coefficients, 2 MSBs - mode (CC,SC,CS,SS), others - down first
		output dv, //!< output data valid
		output signed [OUT_WIDTH - 1 : 0] dout0, //!<frequency domain data output (even samples)
		output signed [OUT_WIDTH - 1 : 0] dout1 //!<frequency domain data output (odd s amples)

		);
		localparam DTT_OUT_DELAY = 99; // 191; // start output to sin/cos rotator, ~=3/4 of 256
		localparam DTT_OUT_DELAY = 128; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
		localparam DTT_IN_DELAY = 62; // 69; // wa -ra min = 1
		reg [ 1:0] start_r;

		@@ -83,6 +92,10 @@ module mclt16x16_bayer#(
		reg [SHIFT_WIDTH-1:0] y_shft_r3; // registered @ start_dtt
		reg [SHIFT_WIDTH-1:0] x_shft_r4; // registered @ dtt_start_first_fill
		reg [SHIFT_WIDTH-1:0] y_shft_r4; // registered @ dtt_start_first_fill
		reg inv_checker_r;
		reg inv_checker_r2;
		reg inv_checker_r3;
		reg inv_checker_r4;

		// wire signed [WND_WIDTH-1:0] window; //!< msb==0, always positive
		wire [1:0] signs; //!< bit 0: sign to add to dtt-cc input, bit 1: sign to add to dtt-cs input
		@@ -133,7 +146,12 @@ module mclt16x16_bayer#(
		wire signed [35:0] dtt_r_data_w; // high bits are not used
		wire signed [DTT_IN_WIDTH-1:0] dtt_r_data = dtt_r_data_w[DTT_IN_WIDTH-1:0];

		reg pre_last_out_r;
		wire pre_last_in_w;

		assign pre_last_out = pre_last_out_r;
		assign pre_busy = pre_busy_r \|\| start \|\| (!pre_last_in_w && phases[0]);
		assign pre_last_in = pre_last_in_w;


		always @ (posedge clk) begin
		@@ -141,23 +159,26 @@ module mclt16x16_bayer#(
		if (start) begin
		x_shft_r <= x_shft;
		y_shft_r <= y_shft;
		inv_checker_r <= inv_checker;
		end
		start_r <= {start_r[0], start};
		if (start_r[1]) begin // same latency as mpix_a_w
		x_shft_r2 <= x_shft_r; // use for the window
		y_shft_r2 <= y_shft_r;
		inv_checker_r2 <= inv_checker_r;
		end

		if (start_dtt) begin
		x_shft_r3 <= x_shft_r2;
		y_shft_r3 <= y_shft_r2;
		inv_checker_r3 <= inv_checker_r2;
		end
		/*

		if (dtt_start_first_fill) begin
		x_shft_r4 <= x_shft_r3;
		y_shft_r4 <= y_shft_r3;
		inv_checker_r4 <= inv_checker_r3;
		end
		*/
		if (phases[8]) begin
		pix_d_r <= pix_d;
		window_r <= window_w;
		@@ -249,7 +270,8 @@ module mclt16x16_bayer#(
		.window (window_w), // output[17:0] signed
		.signs (signs), // output[1:0]
		.phases (phases), // output[7:0]
		.var_first (var_first) // output reg
		.var_first (var_first), // output reg
		.pre_last_in (pre_last_in_w)// output reg
		);

		dly_var #(
		@@ -280,10 +302,9 @@ module mclt16x16_bayer#(
		.web (4'hf), // input[3:0]
		.data_in ({{(36-DTT_IN_WIDTH){1'b0}}, data_dtt_in}) // input[35:0]
		);
		wire [8:0] dbgt_diff_wara = dtt_in_wa-dtt_r_ra;
		wire [8:0] dbg_diff_wara_dtt_in = dtt_in_wa-dtt_r_ra; // SuppressThisWarning VEditor : debug only signal


		/*
		wire signed [OUT_WIDTH-1:0] dtt_out_wd;
		wire [3:0] dtt_out_wa16;
		wire dtt_out_we;
		@@ -291,10 +312,218 @@ module mclt16x16_bayer#(
		wire dtt_inc16;
		reg [4:0] dtt_out_ram_cntr;
		reg [4:0] dtt_out_ram_wah;
		reg [1:0] dtt_out_ram_wpage; // one of 4 pages (128 samples long) being written to
		wire dtt_start_fill; // some data available in DTT output buffer, OK to start consecutive readout
		reg dtt_start_first_fill;
		reg dtt_start_out; // start read out to sin/cos rotator
		*/


		// frequency domain, high address bit - page, 2 next - mode, 6 LSBs - transposed FD data (vertical first)
		// will change: 2 address bits - page, 1 - mode, 6 LSBs - transposed FD data (vertical first)
		// here delay should start after full 2 modes (128 samples) written, as readout to rotators starts from both ends

		wire [8:0] dtt_out_ram_wa = {dtt_out_ram_wah,dtt_out_wa16};
		reg [7:0] dtt_dly_cntr;
		// reg [8:0] dtt_rd_cntr; // counter for dtt readout to rotator
		reg [8:0] dtt_rd_cntr_pre; // 1 ahead of the former counter for dtt readout to rotator

		// TODO: fix rd addresses

		// wire [8:0] dtt_rd_ra = {dtt_rd_cntr[8],dtt_rd_cntr[0],dtt_rd_cntr[1],dtt_rd_cntr[7:2]}; // page, mode, frequency
		reg [8:0] dtt_rd_ra0;
		reg [8:0] dtt_rd_ra1;

		reg [3:0] dtt_rd_regen_dv; // dtt output buffer mem read, register enable, data valid
		wire [35:0] dtt_rd_data0_w; // high bits are not used
		wire [35:0] dtt_rd_data1_w; // high bits are not used
		// data to be input to phase rotator
		wire signed [OUT_WIDTH-1:0] dtt_rd_data0 = dtt_rd_data0_w[OUT_WIDTH-1:0]; // valid with dtt_rd_regen_dv[3]
		wire signed [OUT_WIDTH-1:0] dtt_rd_data1 = dtt_rd_data1_w[OUT_WIDTH-1:0]; // valid with dtt_rd_regen_dv[3]

		// wire dtt_first_quad_out = ~dtt_out_ram_cntr[3] & ~dtt_out_ram_cntr[2];
		wire dtt_first_quad_out = ~dtt_out_ram_cntr[2];

		always @(posedge clk) begin
		if (rst) dtt_out_ram_cntr <= 0;
		else if (dtt_inc16) dtt_out_ram_cntr <= dtt_out_ram_cntr + 1;
		dtt_out_ram_wah <= dtt_out_ram_cntr - dtt_sub16;

		dtt_start_first_fill <= dtt_start_fill & dtt_first_quad_out;

		if (dtt_start_first_fill) dtt_out_ram_wpage <= dtt_out_ram_wah[4:3];

		if (rst) dtt_dly_cntr <= 0;
		else if (dtt_start_first_fill) dtt_dly_cntr <= DTT_OUT_DELAY;
		else if (\|dtt_dly_cntr) dtt_dly_cntr <= dtt_dly_cntr - 1;

		dtt_start_out <= dtt_dly_cntr == 1;

		// if (rst) dtt_rd_regen_dv[0] <= 0;
		// else if (dtt_start_out) dtt_rd_regen_dv[0] <= 1;
		// else if (&dtt_rd_cntr[7:0]) dtt_rd_regen_dv[0] <= 0;

		if (rst) dtt_rd_regen_dv[0] <= 0;
		else if (dtt_start_out) dtt_rd_regen_dv[0] <= 1;
		else if (&dtt_rd_cntr_pre[6:0]) dtt_rd_regen_dv[0] <= 0;

		if (rst) dtt_rd_regen_dv[3:1] <= 0;
		else dtt_rd_regen_dv[3:1] <= dtt_rd_regen_dv[2:0];

		// if (dtt_start_out) dtt_rd_cntr_pre <= {dtt_out_ram_wah[4], 8'b0}; //copy page number
		if (dtt_start_out) dtt_rd_cntr_pre <= {dtt_out_ram_wpage, 7'b0}; //copy page number

		else if (dtt_rd_regen_dv[0]) dtt_rd_cntr_pre <= dtt_rd_cntr_pre + 1;
		//// wire [8:0] dtt_rd_ra = {dtt_rd_cntr[8],dtt_rd_cntr[0],dtt_rd_cntr[1],dtt_rd_cntr[7:2]}; // page, mode, frequency

		dtt_rd_ra0 <= {dtt_rd_cntr_pre[8:7],
		dtt_rd_cntr_pre[6] ^ dtt_rd_cntr_pre[5],
		dtt_rd_cntr_pre[5]? (~dtt_rd_cntr_pre[4:0]) : dtt_rd_cntr_pre[4:0],
		dtt_rd_cntr_pre[5]};
		dtt_rd_ra1 <= {dtt_rd_cntr_pre[8:7],
		dtt_rd_cntr_pre[6] ^ dtt_rd_cntr_pre[5],
		dtt_rd_cntr_pre[5]? (~dtt_rd_cntr_pre[4:0]) : dtt_rd_cntr_pre[4:0],
		~dtt_rd_cntr_pre[5]};

		end


		dtt_iv_8x8_ad #(
		.INPUT_WIDTH (DTT_IN_WIDTH),
		.OUT_WIDTH (OUT_WIDTH),
		.OUT_RSHIFT1 (OUT_RSHIFT),
		.OUT_RSHIFT2 (OUT_RSHIFT2),
		.TRANSPOSE_WIDTH (TRANSPOSE_WIDTH),
		.DSP_B_WIDTH (DSP_B_WIDTH),
		.DSP_A_WIDTH (DSP_A_WIDTH),
		.DSP_P_WIDTH (DSP_P_WIDTH)
		) dtt_iv_8x8_ad_i (
		.clk (clk), // input
		.rst (rst), // input
		.start (dtt_start), // input
		.mode ({dtt_mode, 1'b0}), // input[1:0] for checker-board: only 2 of 4 modes (CC, SC)
		.xin (dtt_r_data), // input[24:0] signed
		.pre_last_in (), // output reg
		.mode_out (), // dtt_mode_out), // output[1:0] reg
		.pre_busy (), // output reg
		.out_wd (dtt_out_wd), // output[24:0] reg
		.out_wa (dtt_out_wa16), // output[3:0] reg
		.out_we (dtt_out_we), // output reg
		.sub16 (dtt_sub16), // output reg
		.inc16 (dtt_inc16), // output reg
		.start_out (dtt_start_fill) // output[24:0] signed
		);
		// still incorrectly shows difference when filled nto sequentially
		reg [8:0] dbg_last_dtt_out_ram_wa; // SuppressThisWarning VEditor : debug only signal
		wire [8:0] dbg_diff_wara_dtt_out0 = dbg_last_dtt_out_ram_wa-dtt_rd_ra0; // SuppressThisWarning VEditor : debug only signal
		wire [8:0] dbg_diff_wara_dtt_out1 = dbg_last_dtt_out_ram_wa-dtt_rd_ra1; // SuppressThisWarning VEditor : debug only signal
		always @ (posedge clk) begin
		if (dtt_out_we) dbg_last_dtt_out_ram_wa <= dtt_out_ram_wa;
		end

		// Two of 2 page buffers after dtt (feeding two phase rotators), address MSB is not needed
		ram18p_var_w_var_r #(
		.REGISTERS(1),
		.LOG2WIDTH_WR(5),
		.LOG2WIDTH_RD(5)
		) ram18p_var_w_var_r_dtt_out0_i (
		.rclk (clk), // input
		.raddr (dtt_rd_ra0), // input[8:0]
		.ren (dtt_rd_regen_dv[1]), // input
		.regen (dtt_rd_regen_dv[2]), // input
		.data_out (dtt_rd_data0_w), // output[35:0]
		.wclk (clk), // input
		.waddr (dtt_out_ram_wa), // input[8:0]
		.we (dtt_out_we), // input
		.web (4'hf), // input[3:0]
		.data_in ({{(36-DTT_IN_WIDTH){1'b0}}, dtt_out_wd}) // input[35:0]
		);

		ram18p_var_w_var_r #(
		.REGISTERS(1),
		.LOG2WIDTH_WR(5),
		.LOG2WIDTH_RD(5)
		) ram18p_var_w_var_r_dtt_out1_i (
		.rclk (clk), // input
		.raddr (dtt_rd_ra1), // input[8:0]
		.ren (dtt_rd_regen_dv[1]), // input
		.regen (dtt_rd_regen_dv[2]), // input
		.data_out (dtt_rd_data1_w), // output[35:0]
		.wclk (clk), // input
		.waddr (dtt_out_ram_wa), // input[8:0]
		.we (dtt_out_we), // input
		.web (4'hf), // input[3:0]
		.data_in ({{(36-DTT_IN_WIDTH){1'b0}}, dtt_out_wd}) // input[35:0]
		);

		// Two rotators in parallel
		// TODO: add negating for antisymmetrical (inverted checkerboard)
		// Rotate phase (equvalent to pixel domain shift)
		phase_rotator #(
		.FD_WIDTH(OUT_WIDTH),
		.SHIFT_WIDTH(SHIFT_WIDTH), // should be exactly 7
		.DSP_B_WIDTH(DSP_B_WIDTH),
		.DSP_A_WIDTH(DSP_A_WIDTH),
		.DSP_P_WIDTH(DSP_P_WIDTH)
		) phase_rotator0_i (
		.clk (clk), // input
		.rst (rst), // input
		.start (dtt_start_out), // input
		// are these shift OK? Will need to be valis only @ dtt_start_out
		.shift_h (x_shft_r4), // input[6:0] signed
		.shift_v (y_shft_r4), // input[6:0] signed
		.inv_checker (inv_checker_r4),// input only used for Bayer mosaic data
		.fd_din (dtt_rd_data0), // input[24:0] signed. Expected latency = 3 from start
		.fd_out (dout0), // output[24:0] reg signed
		.pre_first_out (pre_first_out), // output reg
		.fd_dv (dv) // output reg
		);

		phase_rotator #(
		.FD_WIDTH(OUT_WIDTH),
		.SHIFT_WIDTH(SHIFT_WIDTH), // should be exactly 7
		.DSP_B_WIDTH(DSP_B_WIDTH),
		.DSP_A_WIDTH(DSP_A_WIDTH),
		.DSP_P_WIDTH(DSP_P_WIDTH)
		) phase_rotator1_i (
		.clk (clk), // input
		.rst (rst), // input
		.start (dtt_start_out), // input
		// are these shift OK? Will need to be valis only @ dtt_start_out
		.shift_h (x_shft_r4), // input[6:0] signed
		.shift_v (y_shft_r4), // input[6:0] signed
		.inv_checker (inv_checker_r4),// input only used for Bayer mosaic data
		.fd_din (dtt_rd_data1), // input[24:0] signed. Expected latency = 3 from start
		.fd_out (dout1), // output[24:0] reg signed
		.pre_first_out (pre_first_out), // output reg
		.fd_dv (dv) // output reg
		);



		reg [7:0] out_addr_r;
		reg [3:0] dead_cntr;
		reg pre_busy_r;

		assign out_addr = {out_addr_r[0],out_addr_r[1],out_addr_r[4:2],out_addr_r[7:5]};

		always @ (posedge clk) begin
		if (!dv) out_addr_r <= 0;
		else out_addr_r <= out_addr_r + 1;

		pre_last_out_r <= out_addr_r == 8'hfe;

		if (rst) pre_busy_r <= 0;
		else if (pre_last_in_w) pre_busy_r <= 1;
		else if (dead_cntr == 0) pre_busy_r <= 0;


		if (~pre_busy_r) dead_cntr <= DEAD_CYCLES;
		else dead_cntr <= dead_cntr - 1;

		end




		endmodule

dsp/mclt_bayer_fold.v

+7 −2

Original line number	Diff line number	Diff line
		@@ -72,7 +72,8 @@ module mclt_bayer_fold#(
		output signed [WND_WIDTH-1:0] window, //!< msb==0, always positive
		output [1:0] signs, //!< bit 0: sign to add to dtt-cc input, bit 1: sign to add to dtt-cs input
		output [14:0] phases, //!< other signals
		output reg var_first
		output reg var_first, //!< first of 2 fold variants (4 for monochrome, 2 left for checker)
		output reg pre_last_in //!< pre last data in
		);
		reg [6:0] in_cntr; // input phase counter
		reg [14:0] run_r; // run phase
		@@ -146,6 +147,10 @@ module mclt_bayer_fold#(
		var_first <= var_first_d;
		end

		pre_last_in <= in_cntr[7:0] == 8'hfd;



		end

		ram18tp_var_w_var_r #(

dsp/phase_rotator.v

+3 −0

Original line number	Diff line number	Diff line
		@@ -51,6 +51,7 @@ module phase_rotator#(
		input start, //!< single-cycle start pulse that goes 1 cycle before first data
		input signed [SHIFT_WIDTH-1:0] shift_h, //!< subpixel shift horizontal
		input signed [SHIFT_WIDTH-1:0] shift_v, //!< subpixel shift vertical
		input inv_checker, //!< negate 2-nd and fourth samples (for handling inverted checkerboard)
		// input data CC,CS,SC,SS in column scan order (matching DTT)
		input signed [FD_WIDTH-1:0] fd_din, //!< frequency domain data in, LATENCY=3 from start
		output reg signed [FD_WIDTH-1:0] fd_out, //!< frequency domain data in
		@@ -90,6 +91,7 @@ module phase_rotator#(
		reg [SHIFT_WIDTH-1:0] shift_v0;
		reg [SHIFT_WIDTH-1:0] shift_vr;
		reg [SHIFT_WIDTH-1:0] shift_hv; // combined horizonta and vertical shifts to match cntr_mux;
		reg inv_checker_r;
		reg [4:0] sign_cs; // sign for cos / sin, feed to DSP
		wire sign_cs_d; // sign_cs delayed by 3 clocks
		reg [1:0] sign_cs_r; // sign_cs delayed by 5 clocks
		@@ -108,6 +110,7 @@ module phase_rotator#(

		if (start) shift_hr <= shift_h;
		if (start) shift_v0 <= shift_v;
		if (start) inv_checker_r <= inv_checker;
		if (start_d[3]) shift_vr <= shift_v0;

		if (rst) run_h <= 0;

Admin message