rtl/gfx: reimplement multiplier as a much smaller mat-vec pipeline

1 files changed, 22 insertions, 7 deletions

diff --git a/rtl/gfx/mat_vec_mul.sv b/rtl/gfx/mat_vec_mul.sv
index 43860c9..9f5dcae 100644
--- a/rtl/gfx/mat_vec_mul.sv
+++ b/rtl/gfx/mat_vec_mul.sv
@@ -5,27 +5,42 @@ module mat_vec_mul
 	input  logic clk,
 	             rst_n,
 
-	input  logic start,
 	input  mat4  a,
 	input  vec4  x,
+	input  logic in_valid,
+	             out_ready,
 
-	output logic done,
-	output vec4  q
+	output vec4  q,
+	output logic in_ready,
+	             out_valid
 );
 
-	logic dones[`FLOATS_PER_VEC];
+	logic stall_mul, stall_fold, mul_ready, mul_valid, feedback, feedback_last;
 
-	assign done = dones[0];
+	pipeline_flow #(.STAGES(`FP_MUL_STAGES)) mul
+	(
+		.stall(stall_mul),
+		.out_ready(mul_ready),
+		.out_valid(mul_valid),
+		.*
+	);
+
+	fold_flow fold
+	(
+		.stall(stall_fold),
+		.in_ready(mul_ready),
+		.in_valid(mul_valid),
+		.*
+	);
 
 	genvar i;
 	generate
-		for (i = 0; i < `FLOATS_PER_VEC; ++i) begin: dots
+		for (i = 0; i < `VECS_PER_MAT; ++i) begin: dots
 			vec_dot dot_i
 			(
 				.a(a[i]),
 				.b(x),
 				.q(q[i]),
-				.done(dones[i]),
 				.*
 			);
 		end