platform/wavelet3d: implement floating-point min/max

5 files changed, 99 insertions, 15 deletions

diff --git a/platform/wavelet3d/gfx_fpint.sv b/platform/wavelet3d/gfx_fpint.sv
index e5c457c..b3108a4 100644
--- a/platform/wavelet3d/gfx_fpint.sv
+++ b/platform/wavelet3d/gfx_fpint.sv
@@ -13,6 +13,9 @@ module gfx_fpint
 	                 mnorm_put_mul,
 	                 mnorm_zero_b,
 	                 mnorm_zero_flags,
+	                 minmax_abs,
+	                 minmax_swap,
+	                 minmax_zero_min,
 	                 minmax_copy_flags,
 	                 shiftr_int_signed,
 	                 addsub_copy_flags,
@@ -42,6 +45,9 @@ module gfx_fpint
 	assign op.mnorm_put_mul = mnorm_put_mul;
 	assign op.mnorm_zero_b = mnorm_zero_b;
 	assign op.mnorm_zero_flags = mnorm_zero_flags;
+	assign op.minmax_abs = minmax_abs;
+	assign op.minmax_swap = minmax_swap;
+	assign op.minmax_zero_min = minmax_zero_min;
 	assign op.minmax_copy_flags = minmax_copy_flags;
 	assign op.shiftr_int_signed = shiftr_int_signed;
 	assign op.addsub_copy_flags = addsub_copy_flags;
@@ -68,6 +74,9 @@ module gfx_fpint
 				.put_mul_2(stage_op[2].mnorm_put_mul),
 				.zero_b_2(stage_op[2].mnorm_zero_b),
 				.zero_flags_2(stage_op[2].mnorm_zero_flags),
+				.abs_3(stage_op[3].minmax_abs),
+				.swap_3(stage_op[3].minmax_swap),
+				.zero_min_3(stage_op[3].minmax_zero_min),
 				.copy_flags_3(stage_op[3].minmax_copy_flags),
 				.int_signed_5(stage_op[5].shiftr_int_signed),
 				.copy_flags_6(stage_op[6].addsub_copy_flags),
diff --git a/platform/wavelet3d/gfx_fpint_lane.sv b/platform/wavelet3d/gfx_fpint_lane.sv
index 8cb77a8..b52a393 100644
--- a/platform/wavelet3d/gfx_fpint_lane.sv
+++ b/platform/wavelet3d/gfx_fpint_lane.sv
@@ -25,6 +25,9 @@ module gfx_fpint_lane
 	                     put_mul_2,
 	                     zero_b_2,
 	                     zero_flags_2,
+	                     abs_3,
+	                     swap_3,
+	                     zero_min_3,
 	                     copy_flags_3,
 	                     int_signed_5,
 	                     copy_flags_6,
@@ -128,6 +131,9 @@ module gfx_fpint_lane
 		.clk(clk),
 		.in(mnorm_minmax),
 		.out(minmax_expdiff),
+		.abs(abs_3),
+		.swap(swap_3),
+		.zero_min(zero_min_3),
 		.copy_flags(copy_flags_3)
 	);
 
@@ -351,25 +357,44 @@ module gfx_fpint_lane_minmax
 	input  logic                     clk,
 
 	input  gfx::fpint_mnorm_minmax   in,
-	input  logic                     copy_flags,
+	input  logic                     abs,
+	                                 swap,
+	                                 zero_min,
+	                                 copy_flags,
 
 	output gfx::fpint_minmax_expdiff out
 );
 
 	import gfx::*;
 
+	logic abs_b_gt_abs_a, b_gt_a;
+
+	/* Wiki dice:
+	 *
+	 * A property of the single- and double-precision formats is that
+	 * their encoding allows one to easily sort them without using
+	 * floating-point hardware, as if the bits represented sign-magnitude
+	 * integers, although it is unclear whether this was a design
+	 * consideration (it seems noteworthy that the earlier IBM hexadecimal
+	 * floating-point representation also had this property for normalized
+	 * numbers).
+	 */
+	assign abs_b_gt_abs_a = {in.b.exp, in.b.mant} > {in.a.exp, in.a.mant};
+
+	always_comb begin
+		unique case ({in.b.sign, in.a.sign})
+			2'b00: b_gt_a = abs_b_gt_abs_a;
+			2'b01: b_gt_a = 1;
+			2'b10: b_gt_a = 0;
+			2'b11: b_gt_a = abs_b_gt_abs_a;
+		endcase
+
+		if (abs)
+			b_gt_a = abs_b_gt_abs_a;
+	end
+
 	always_ff @(posedge clk) begin
-		/* Wiki dice:
-		 *
-		 * A property of the single- and double-precision formats is that
-		 * their encoding allows one to easily sort them without using
-		 * floating-point hardware, as if the bits represented sign-magnitude
-		 * integers, although it is unclear whether this was a design
-		 * consideration (it seems noteworthy that the earlier IBM hexadecimal
-		 * floating-point representation also had this property for normalized
-		 * numbers).
-		 */
-		if ({in.b.exp, in.b.mant} > {in.a.exp, in.a.mant}) begin
+		if (b_gt_a ^ swap) begin
 			out.max <= in.b;
 			out.min <= in.a;
 			out.max_class <= in.b_class;
@@ -381,6 +406,11 @@ module gfx_fpint_lane_minmax
 			out.min_class <= in.b_class;
 		end
 
+		if (zero_min) begin
+			out.min <= 0;
+			out.min_class <= classify_float(0);
+		end
+
 		out.guard <= in.guard;
 		out.round <= in.round;
 		out.sticky <= in.sticky;
diff --git a/platform/wavelet3d/gfx_pkg.sv b/platform/wavelet3d/gfx_pkg.sv
index bd71eff..72c8f80 100644
--- a/platform/wavelet3d/gfx_pkg.sv
+++ b/platform/wavelet3d/gfx_pkg.sv
@@ -84,6 +84,9 @@ package gfx;
 		      mnorm_put_mul,
 		      mnorm_zero_b,
 		      mnorm_zero_flags,
+		      minmax_abs,
+		      minmax_swap,
+		      minmax_zero_min,
 		      minmax_copy_flags,
 		      shiftr_int_signed,
 		      addsub_copy_flags,
diff --git a/platform/wavelet3d/gfx_top.sv b/platform/wavelet3d/gfx_top.sv
index 79ffc2d..77126b9 100644
--- a/platform/wavelet3d/gfx_top.sv
+++ b/platform/wavelet3d/gfx_top.sv
@@ -13,6 +13,9 @@ module gfx_top
 	                 mnorm_put_mul,
 	                 mnorm_zero_b,
 	                 mnorm_zero_flags,
+	                 minmax_abs,
+	                 minmax_swap,
+	                 minmax_zero_min,
 	                 minmax_copy_flags,
 	                 shiftr_int_signed,
 	                 addsub_copy_flags,
diff --git a/platform/wavelet3d/main.cpp b/platform/wavelet3d/main.cpp
index 2555d1d..74435fd 100644
--- a/platform/wavelet3d/main.cpp
+++ b/platform/wavelet3d/main.cpp
@@ -41,7 +41,7 @@ int main(int argc, char **argv)
 	{
 		std::printf
 		(
-			"[%03d] %s %s %s\n",
+			"[%03d] <= %s %s %s\n",
 			time, std::to_string(a).c_str(), op, std::to_string(b).c_str()
 		);
 
@@ -64,6 +64,9 @@ int main(int argc, char **argv)
 		top.mnorm_put_mul = 0;
 		top.mnorm_zero_flags = 1;
 		top.mnorm_zero_b = 1;
+		top.minmax_abs = 1;
+		top.minmax_swap = 0;
+		top.minmax_zero_min = 0;
 		top.minmax_copy_flags = 0;
 		top.shiftr_int_signed = 1;
 		top.addsub_int_operand = 1;
@@ -86,6 +89,9 @@ int main(int argc, char **argv)
 		top.mnorm_put_mul = 0;
 		top.mnorm_zero_flags = 1;
 		top.mnorm_zero_b = 1;
+		top.minmax_abs = 1;
+		top.minmax_swap = 0;
+		top.minmax_zero_min = 0;
 		top.minmax_copy_flags = 1;
 		top.shiftr_int_signed = 0;
 		top.addsub_int_operand = 0;
@@ -109,6 +115,9 @@ int main(int argc, char **argv)
 		top.mnorm_put_mul = 1;
 		top.mnorm_zero_flags = 0;
 		top.mnorm_zero_b = 1;
+		top.minmax_abs = 1;
+		top.minmax_swap = 0;
+		top.minmax_zero_min = 0;
 		top.minmax_copy_flags = 1;
 		top.shiftr_int_signed = 0;
 		top.addsub_int_operand = 0;
@@ -131,6 +140,9 @@ int main(int argc, char **argv)
 		top.mnorm_put_mul = 0;
 		top.mnorm_zero_flags = 0;
 		top.mnorm_zero_b = 0;
+		top.minmax_abs = 1;
+		top.minmax_swap = 0;
+		top.minmax_zero_min = 0;
 		top.minmax_copy_flags = 0;
 		top.shiftr_int_signed = 0;
 		top.addsub_int_operand = 0;
@@ -144,6 +156,31 @@ int main(int argc, char **argv)
 		send_op(a, "+", b);
 	};
 
+	auto min_max_fp = [&](float a, float b, bool min = false)
+	{
+		top.setup_mul_float = 0;
+		top.setup_unit_b = 1;
+		top.mnorm_put_hi = 0;
+		top.mnorm_put_lo = 1;
+		top.mnorm_put_mul = 0;
+		top.mnorm_zero_flags = 0;
+		top.mnorm_zero_b = 0;
+		top.minmax_abs = 0;
+		top.minmax_swap = min;
+		top.minmax_zero_min = 1;
+		top.minmax_copy_flags = 1;
+		top.shiftr_int_signed = 0;
+		top.addsub_int_operand = 0;
+		top.addsub_copy_flags = 1;
+		top.clz_force_nop = 1;
+		top.shiftl_copy_flags = 1;
+		top.round_copy_flags = 1;
+		top.round_enable = 0;
+		top.encode_enable = 0;
+
+		send_op(a, min ? "min" : "max", b);
+	};
+
 	top.rst_n = 0;
 	top.in_valid = 0;
 	cycle();
@@ -167,8 +204,10 @@ int main(int argc, char **argv)
 	mul_int(a, b);
 	mul_fp(a_flt, b_flt);
 	add_fp(a_flt, b_flt);
+	min_max_fp(a_flt, b_flt);
+	min_max_fp(a_flt, b_flt, true);
 
-	while (time < 20) {
+	while (time < 50) {
 		cycle();
 
 		if (!top.out_valid)
@@ -180,7 +219,7 @@ int main(int argc, char **argv)
 
 		std::printf
 		(
-			"[%03d] q=0x%08x, q_flt=%g, q_int=%d, q_uint=%u\n",
+			"[%03d] => q=0x%08x, q_flt=%g, q_int=%d, q_uint=%u\n",
 			time, q_bits, q_flt, q_int, q_bits
 		);
 	}