rtl/gfx: implement shuffler

7 files changed, 125 insertions, 32 deletions

diff --git a/rtl/gfx/gfx_defs.sv b/rtl/gfx/gfx_defs.sv
index b65d74e..0a4a5da 100644
--- a/rtl/gfx/gfx_defs.sv
+++ b/rtl/gfx/gfx_defs.sv
@@ -12,9 +12,9 @@
 `define FP_INV_STAGES 3  // ~178 LUTs ~1 bloque DSP
 
 typedef logic[`FLOAT_BITS - 1:0]  fp;
-typedef fp                        vec2[2];
-typedef fp                        vec4[`FLOATS_PER_VEC];
-typedef vec4                      mat4[`VECS_PER_MAT];
+typedef fp[1:0]                   vec2;
+typedef fp[`FLOATS_PER_VEC - 1:0] vec4;
+typedef vec4[`VECS_PER_MAT - 1:0] mat4;
 
 `define FP_UNIT 16'h3c00
 
@@ -229,8 +229,8 @@ typedef struct packed
 
 typedef struct packed
 {
-	logic         is_imm,
-	              is_swizzle;
+	logic         is_swizzle,
+	              is_broadcast;
 	fp            imm;
 	vec_mask      select_mask;
 	swizzle_lanes swizzle_op;
@@ -249,4 +249,10 @@ typedef struct packed
 	shuffler_deco shuffler;
 } insn_deco; // "insn_decode" ya existe en core, esto es confuso pero lo hice por tiempo
 
+typedef struct packed
+{
+	vreg_num dst;
+	mat4     data;
+} wb_op;
+
 `endif
diff --git a/rtl/gfx/gfx_shuffle.sv b/rtl/gfx/gfx_shuffle.sv
deleted file mode 100644
index c52ef25..0000000
--- a/rtl/gfx/gfx_shuffle.sv
+++ /dev/null
@@ -1,20 +0,0 @@
-`include "gfx/gfx_defs.sv"
-
-module gfx_shuffle
-(
-	input  logic    clk,
-
-	input  vec4     a,
-	                b,
-	input  vec_mask select,
-	input  logic    stall,
-
-	output vec4     out
-);
-
-	always_ff @(posedge clk)
-		if (!stall)
-			for (integer i = 0; i < `FLOATS_PER_VEC; ++i)
-				out[i] <= select[i] ? b[i] : a[i];
-
-endmodule
diff --git a/rtl/gfx/gfx_sp.sv b/rtl/gfx/gfx_sp.sv
index 4459870..a65bd6e 100644
--- a/rtl/gfx/gfx_sp.sv
+++ b/rtl/gfx/gfx_sp.sv
@@ -51,6 +51,19 @@ module gfx_sp
 		.*
 	);
 
+	gfx_sp_shuffler shuffler
+	(
+		.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_decode.sv b/rtl/gfx/gfx_sp_decode.sv
index d24c608..41ce438 100644
--- a/rtl/gfx/gfx_sp_decode.sv
+++ b/rtl/gfx/gfx_sp_decode.sv
@@ -51,8 +51,8 @@ module gfx_sp_decode
 		deco_in.ex.combiner = 0;
 		deco_in.ex.shuffler = 0;
 
-		deco_in.shuffler.is_imm = 1'bx;
 		deco_in.shuffler.is_swizzle = 1'bx;
+		deco_in.shuffler.is_broadcast = 1'bx;
 
 		unique casez (insn)
 			`GFX_INSN_OP_SELECT: begin
@@ -61,8 +61,8 @@ module gfx_sp_decode
 				deco_in.read_src_b = 1;
 
 				deco_in.ex.shuffler = 1;
-				deco_in.shuffler.is_imm = 0;
 				deco_in.shuffler.is_swizzle = 0;
+				deco_in.shuffler.is_broadcast = 0;
 			end
 
 			`GFX_INSN_OP_SWIZZL: begin
@@ -70,7 +70,6 @@ module gfx_sp_decode
 				deco_in.read_src_a = 1;
 
 				deco_in.ex.shuffler = 1;
-				deco_in.shuffler.is_imm = 0;
 				deco_in.shuffler.is_swizzle = 1;
 			end
 
@@ -78,8 +77,8 @@ module gfx_sp_decode
 				deco_in.writeback = 1;
 
 				deco_in.ex.shuffler = 1;
-				deco_in.shuffler.is_imm = 1;
 				deco_in.shuffler.is_swizzle = 0;
+				deco_in.shuffler.is_broadcast = 1;
 			end
 
 			`GFX_INSN_OP_MATVEC: begin
diff --git a/rtl/gfx/gfx_sp_select.sv b/rtl/gfx/gfx_sp_select.sv
new file mode 100644
index 0000000..46b23c9
--- /dev/null
+++ b/rtl/gfx/gfx_sp_select.sv
@@ -0,0 +1,25 @@
+`include "gfx/gfx_defs.sv"
+
+module gfx_sp_select
+(
+	input  logic         clk,
+
+	input  vec4          a,
+	                     b,
+	input  shuffler_deco deco,
+	input  logic         stall,
+
+	output vec4          out
+);
+
+	always_ff @(posedge clk)
+		if (!stall)
+			for (integer i = 0; i < `FLOATS_PER_VEC; ++i)
+				if (deco.is_broadcast)
+					out[i] <= deco.imm;
+				else if (deco.select_mask[i])
+					out[i] <= b[i];
+				else
+					out[i] <= a[i];
+
+endmodule
diff --git a/rtl/gfx/gfx_sp_shuffler.sv b/rtl/gfx/gfx_sp_shuffler.sv
new file mode 100644
index 0000000..e4c119d
--- /dev/null
+++ b/rtl/gfx/gfx_sp_shuffler.sv
@@ -0,0 +1,70 @@
+`include "gfx/gfx_defs.sv"
+
+module gfx_sp_shuffler
+(
+	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
+);
+
+	mat4 select_out, swizzle_out;
+	wb_op wb_out;
+	logic stall, is_swizzle;
+	vreg_num hold_dst;
+
+	gfx_pipeline_flow #(.STAGES(2)) flow
+	(
+		.out_ready(wb_ready),
+		.out_valid(wb_valid),
+		.*
+	);
+
+	gfx_skid_buf #(.WIDTH($bits(wb))) skid
+	(
+		.in(wb_out),
+		.out(wb),
+		.*
+	);
+
+	genvar gen_i;
+	generate
+		for (gen_i = 0; gen_i < `VECS_PER_MAT; ++gen_i) begin: lanes
+			gfx_sp_select select
+			(
+				.a(a[gen_i]),
+				.b(b[gen_i]),
+				.out(select_out[gen_i]),
+				.deco(deco.shuffler),
+				.*
+			);
+
+			gfx_sp_swizzle swizzle
+			(
+				.in(a[gen_i]),
+				.out(swizzle_out[gen_i]),
+				.deco(deco.shuffler),
+				.*
+			);
+		end
+	endgenerate
+
+	always_ff @(posedge clk)
+		if (!stall) begin
+			hold_dst <= deco.dst;
+			is_swizzle <= deco.shuffler.is_swizzle;
+
+			wb_out.dst <= hold_dst;
+			for (integer i = 0; i < `VECS_PER_MAT; ++i)
+				wb_out.data[i] <= is_swizzle ? swizzle_out[i] : select_out[i];
+		end
+
+endmodule
diff --git a/rtl/gfx/gfx_swizzle.sv b/rtl/gfx/gfx_sp_swizzle.sv
index 1a57c11..d07d934 100644
--- a/rtl/gfx/gfx_swizzle.sv
+++ b/rtl/gfx/gfx_sp_swizzle.sv
@@ -1,11 +1,11 @@
 `include "gfx/gfx_defs.sv"
 
-module gfx_swizzle
+module gfx_sp_swizzle
 (
 	input  logic         clk,
 
 	input  vec4          in,
-	input  swizzle_lanes select,
+	input  shuffler_deco deco,
 	input  logic         stall,
 
 	output vec4          out
@@ -14,6 +14,6 @@ module gfx_swizzle
 	always_ff @(posedge clk)
 		if (!stall)
 			for (integer i = 0; i < `FLOATS_PER_VEC; ++i)
-				out[i] <= in[select[i]];
+				out[i] <= in[deco.swizzle_op[i]];
 
 endmodule