rtl/gfx: implement matrix transpose

3 files changed, 27 insertions, 3 deletions

diff --git a/gfx_hw.tcl b/gfx_hw.tcl
index 570d948..5d56f44 100644
--- a/gfx_hw.tcl
+++ b/gfx_hw.tcl
@@ -51,6 +51,7 @@ add_fileset_file fold_flow.sv SYSTEM_VERILOG PATH rtl/gfx/fold_flow.sv
 add_fileset_file skid_flow.sv SYSTEM_VERILOG PATH rtl/gfx/skid_flow.sv
 add_fileset_file skid_buf.sv SYSTEM_VERILOG PATH rtl/gfx/skid_buf.sv
 add_fileset_file vec_dot.sv SYSTEM_VERILOG PATH rtl/gfx/vec_dot.sv
+add_fileset_file transpose.sv SYSTEM_VERILOG PATH rtl/gfx/transpose.sv
 
 
 # 
diff --git a/rtl/gfx/mat_mat_mul.sv b/rtl/gfx/mat_mat_mul.sv
index cf4bd35..7c21249 100644
--- a/rtl/gfx/mat_mat_mul.sv
+++ b/rtl/gfx/mat_mat_mul.sv
@@ -15,7 +15,7 @@ module mat_mat_mul
 	             out_valid
 );
 
-	mat4 a_hold, b_hold, q_hold, mul_b;
+	mat4 a_hold, b_hold, b_transpose, q_hold, mul_b;
 	vec4 mul_q;
 	logic mul_in_ready, mul_in_valid, mul_out_ready, mul_out_valid;
 	index4 in_index, out_index;
@@ -26,6 +26,12 @@ module mat_mat_mul
 	assign mul_in_valid = in_valid || in_index != `INDEX4_MIN;
 	assign mul_out_ready = out_ready || out_index != `INDEX4_MAX;
 
+	transpose transpose
+	(
+		.in(b),
+		.out(b_transpose)
+	);
+
 	mat_vec_mul mul
 	(
 		.a(in_index == `INDEX4_MIN ? a : a_hold),
@@ -40,7 +46,7 @@ module mat_mat_mul
 
 	always_comb begin
 		mul_b = b_hold;
-		mul_b[0] = b[0];
+		mul_b[0] = b_transpose[0];
 
 		q = q_hold;
 		q[`VECS_PER_MAT - 1] = mul_q;
@@ -61,7 +67,7 @@ module mat_mat_mul
 	always_ff @(posedge clk) begin
 		if (in_ready) begin
 			a_hold <= a;
-			b_hold <= b;
+			b_hold <= b_transpose;
 		end
 
 		if (mul_out_ready && mul_out_valid)
diff --git a/rtl/gfx/transpose.sv b/rtl/gfx/transpose.sv
new file mode 100644
index 0000000..1df68d5
--- /dev/null
+++ b/rtl/gfx/transpose.sv
@@ -0,0 +1,17 @@
+`include "gfx/gfx_defs.sv"
+
+module transpose
+(
+	input  mat4 in,
+	output mat4 out
+);
+
+	integer i, j;
+
+	// Esto no tiene costo en hardware, es un renombramiento de señales
+	always_comb
+		for (i = 0; i < `VECS_PER_MAT; ++i)
+			for (j = 0; j < `FLOATS_PER_VEC; ++j)
+				out[i][j] = in[j][i];
+
+endmodule