rtl/gfx: implement SP combiner

4 files changed, 80 insertions, 2 deletions

diff --git a/rtl/gfx/gfx_defs.sv b/rtl/gfx/gfx_defs.sv
index 0a4a5da..260b17a 100644
--- a/rtl/gfx/gfx_defs.sv
+++ b/rtl/gfx/gfx_defs.sv
@@ -255,4 +255,6 @@ typedef struct packed
 	mat4     data;
 } wb_op;
 
+`define GFX_SP_COMBINER_FIFO_DEPTH 4 // TODO: optimizar esto
+
 `endif
diff --git a/rtl/gfx/gfx_sp.sv b/rtl/gfx/gfx_sp.sv
index a65bd6e..acb1dc2 100644
--- a/rtl/gfx/gfx_sp.sv
+++ b/rtl/gfx/gfx_sp.sv
@@ -64,6 +64,19 @@ module gfx_sp
 		.*
 	);
 
+	gfx_sp_combiner combiner
+	(
+		.a(),
+		.b(),
+		.wb(),
+		.deco(),
+		.in_ready(),
+		.in_valid(0),
+		.wb_ready(1),
+		.wb_valid(),
+		.*
+	);
+
 	logic batch_end, deco_ready;
 	assign deco_ready = 1;
 
diff --git a/rtl/gfx/gfx_sp_combiner.sv b/rtl/gfx/gfx_sp_combiner.sv
new file mode 100644
index 0000000..e5e64c3
--- /dev/null
+++ b/rtl/gfx/gfx_sp_combiner.sv
@@ -0,0 +1,63 @@
+`include "gfx/gfx_defs.sv"
+
+module gfx_sp_combiner
+(
+	input  logic     clk,
+	                 rst_n,
+
+	input  insn_deco deco,
+	input  mat4      a,
+	                 b,
+	input  logic     in_valid,
+	output logic     in_ready,
+
+	output wb_op     wb,
+	input  logic     wb_ready,
+	output logic     wb_valid
+);
+
+	wb_op wb_out;
+	logic mul_ready, mul_valid, fifo_ready, fifo_valid, skid_ready, out_stall;
+
+	assign in_ready = mul_ready && fifo_ready;
+
+	gfx_mat_mat mul
+	(
+		.q(wb_out.data),
+		.in_ready(mul_ready),
+		.in_valid(in_valid && fifo_ready),
+		.out_ready(skid_ready && fifo_valid),
+		.out_valid(mul_valid),
+		.*
+	);
+
+	gfx_fifo #(.WIDTH($bits(vreg_num)), .DEPTH(`GFX_SP_COMBINER_FIFO_DEPTH)) depth
+	(
+		.in(deco.dst),
+		.out(wb_out.dst),
+		.in_ready(fifo_ready),
+		.in_valid(in_valid && mul_ready),
+		.out_ready(skid_ready && mul_valid),
+		.out_valid(fifo_valid),
+		.*
+	);
+
+	gfx_skid_flow out_flow
+	(
+		.stall(out_stall),
+		.in_ready(skid_ready),
+		.in_valid(fifo_valid && mul_valid),
+		.out_ready(wb_ready),
+		.out_valid(wb_valid),
+		.*
+	);
+
+	gfx_skid_buf #(.WIDTH($bits(wb))) out_skid
+	(
+		.in(wb_out),
+		.out(wb),
+		.stall(out_stall),
+		.*
+	);
+
+endmodule
diff --git a/rtl/gfx/gfx_sp_shuffler.sv b/rtl/gfx/gfx_sp_shuffler.sv
index e4c119d..627bf8a 100644
--- a/rtl/gfx/gfx_sp_shuffler.sv
+++ b/rtl/gfx/gfx_sp_shuffler.sv
@@ -37,7 +37,7 @@ module gfx_sp_shuffler
 
 	genvar gen_i;
 	generate
-		for (gen_i = 0; gen_i < `VECS_PER_MAT; ++gen_i) begin: lanes
+		for (gen_i = 0; gen_i < `GFX_SP_LANES; ++gen_i) begin: lanes
 			gfx_sp_select select
 			(
 				.a(a[gen_i]),
@@ -63,7 +63,7 @@ module gfx_sp_shuffler
 			is_swizzle <= deco.shuffler.is_swizzle;
 
 			wb_out.dst <= hold_dst;
-			for (integer i = 0; i < `VECS_PER_MAT; ++i)
+			for (integer i = 0; i < `GFX_SP_LANES; ++i)
 				wb_out.data[i] <= is_swizzle ? swizzle_out[i] : select_out[i];
 		end