rtl/gfx: implement shuffle, swizzle

4 files changed, 45 insertions, 1 deletions

diff --git a/gfx_hw.tcl b/gfx_hw.tcl
index fd78163..9941f43 100644
--- a/gfx_hw.tcl
+++ b/gfx_hw.tcl
@@ -85,6 +85,8 @@ add_fileset_file gfx_sp.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_sp.sv
 add_fileset_file gfx_sp_batch.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_sp_batch.sv
 add_fileset_file gfx_sp_fetch.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_sp_fetch.sv
 add_fileset_file gfx_sp_widener.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_sp_widener.sv
+add_fileset_file gfx_shuffle.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_shuffle.sv
+add_fileset_file gfx_swizzle.sv SYSTEM_VERILOG PATH rtl/gfx/gfx_swizzle.sv
 
 
 # 
diff --git a/rtl/gfx/gfx_defs.sv b/rtl/gfx/gfx_defs.sv
index fa297e6..01fe2e3 100644
--- a/rtl/gfx/gfx_defs.sv
+++ b/rtl/gfx/gfx_defs.sv
@@ -212,6 +212,9 @@ typedef struct packed
 `define GFX_BATCH_FIFO_DEPTH 4
 `define GFX_SP_LANES         `VECS_PER_MAT
 
-typedef logic[`GFX_SP_LANES - 1:0] lane_mask;
+typedef logic[`GFX_SP_LANES - 1:0]   lane_mask;
+typedef logic[`FLOATS_PER_VEC - 1:0] vec_mask;
+
+typedef logic[`FLOATS_PER_VEC - 1:0][$clog2(`FLOATS_PER_VEC) - 1:0] swizzle_lanes;
 
 `endif
diff --git a/rtl/gfx/gfx_shuffle.sv b/rtl/gfx/gfx_shuffle.sv
new file mode 100644
index 0000000..c52ef25
--- /dev/null
+++ b/rtl/gfx/gfx_shuffle.sv
@@ -0,0 +1,20 @@
+`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_swizzle.sv b/rtl/gfx/gfx_swizzle.sv
new file mode 100644
index 0000000..1a57c11
--- /dev/null
+++ b/rtl/gfx/gfx_swizzle.sv
@@ -0,0 +1,19 @@
+`include "gfx/gfx_defs.sv"
+
+module gfx_swizzle
+(
+	input  logic         clk,
+
+	input  vec4          in,
+	input  swizzle_lanes select,
+	input  logic         stall,
+
+	output vec4          out
+);
+
+	always_ff @(posedge clk)
+		if (!stall)
+			for (integer i = 0; i < `FLOATS_PER_VEC; ++i)
+				out[i] <= in[select[i]];
+
+endmodule