ip: downgrade to 16-bit floats

4 files changed, 16 insertions, 9 deletions

diff --git a/rtl/gfx/fp_add.sv b/rtl/gfx/fp_add.sv
index 7f7e2c6..6cf4874 100644
--- a/rtl/gfx/fp_add.sv
+++ b/rtl/gfx/fp_add.sv
@@ -21,6 +21,7 @@ module fp_add
 `ifndef VERILATOR
 	ip_fp_add ip_add
 	(
+		.en(1),
 		.areset(0),
 		.*
 	);
diff --git a/rtl/gfx/fp_mul.sv b/rtl/gfx/fp_mul.sv
index c2f4e2a..c5aa56a 100644
--- a/rtl/gfx/fp_mul.sv
+++ b/rtl/gfx/fp_mul.sv
@@ -21,6 +21,7 @@ module fp_mul
 `ifndef VERILATOR
 	ip_fp_mul ip_mul
 	(
+		.en(1),
 		.areset(0),
 		.*
 	);
diff --git a/rtl/gfx/gfx.sv b/rtl/gfx/gfx.sv
index ab2bcce..ec2fb13 100644
--- a/rtl/gfx/gfx.sv
+++ b/rtl/gfx/gfx.sv
@@ -12,10 +12,13 @@ module gfx
 	output logic[31:0] cmd_readdata
 );
 
+	fp readdata, writedata;
 	mat4 a, b, q, hold_q;
 	logic start, done;
 
-	assign cmd_readdata = hold_q[cmd_address[3:2]][cmd_address[1:0]];
+	assign readdata = hold_q[cmd_address[3:2]][cmd_address[1:0]];
+	assign writedata = cmd_writedata[`FLOAT_BITS - 1:0];
+	assign cmd_readdata = {{($bits(cmd_readdata) - `FLOAT_BITS){1'b0}}, readdata};
 
 	mat_mat_mul mul
 	(
@@ -25,9 +28,9 @@ module gfx
 	always_ff @(posedge clk) begin
 		if (cmd_write) begin
 			if (cmd_address[4])
-				a[cmd_address[3:2]][cmd_address[1:0]] <= cmd_writedata;
+				a[cmd_address[3:2]][cmd_address[1:0]] <= writedata;
 			else
-				b[cmd_address[3:2]][cmd_address[1:0]] <= cmd_writedata;
+				b[cmd_address[3:2]][cmd_address[1:0]] <= writedata;
 		end
 
 		if (done)
diff --git a/rtl/gfx/gfx_defs.sv b/rtl/gfx/gfx_defs.sv
index e52f243..d01822a 100644
--- a/rtl/gfx/gfx_defs.sv
+++ b/rtl/gfx/gfx_defs.sv
@@ -1,14 +1,16 @@
 `ifndef GFX_DEFS_SV
 `define GFX_DEFS_SV
 
-`define FP_ADD_STAGES 13
-`define FP_MUL_STAGES 6
-
+`define FLOAT_BITS     16
 `define FLOATS_PER_VEC 4
 `define VECS_PER_MAT   4
 
-typedef logic[31:0] fp;
-typedef fp          vec4[`FLOATS_PER_VEC];
-typedef vec4        mat4[`VECS_PER_MAT];
+// Target de 100MHz con float16, rounding aproximado
+`define FP_ADD_STAGES 4 // ~325 LUTs
+`define FP_MUL_STAGES 3 // ~119 LUTs ~1 bloque DSP
+
+typedef logic[`FLOAT_BITS - 1:0] fp;
+typedef fp                       vec4[`FLOATS_PER_VEC];
+typedef vec4                     mat4[`VECS_PER_MAT];
 
 `endif