refactor lut/motivity

11 files changed, 315 insertions, 166 deletions

diff --git a/common/accel-lookup.hpp b/common/accel-lookup.hpp
new file mode 100644
index 0000000..18c9ea5
--- /dev/null
+++ b/common/accel-lookup.hpp
@@ -0,0 +1,152 @@
+#pragma once
+
+#include "rawaccel-base.hpp"
+#include "utility.hpp"
+
+namespace rawaccel {
+
+	struct linear_range {
+		double start;
+		double stop;
+		int num;
+
+		template <typename Func>
+		void for_each(Func fn) const
+		{
+			double interval = (stop - start) / (num - 1);
+			for (int i = 0; i < num; i++) {
+				fn(i * interval + start);
+			}
+		}
+
+		int size() const
+		{
+			return num;
+		}
+	};
+
+
+	// represents the range [2^start, 2^stop], with num - 1
+	// elements linearly spaced between each exponential step
+	struct fp_rep_range {
+		int start;
+		int stop;
+		int num;
+
+		template <typename Func>
+		void for_each(Func fn) const
+		{
+			for (int e = 0; e < stop - start; e++) {
+				double exp_scale = scalbn(1, e + start) / num;
+
+				for (int i = 0; i < num; i++) {
+					fn((i + num) * exp_scale);
+				}
+			}
+
+			fn(scalbn(1, stop));
+		}
+
+		int size() const
+		{
+			return (stop - start) * num + 1;
+		}
+	};
+
+	template <typename Lookup>
+	struct lut_base {
+		enum { capacity = LUT_CAPACITY };
+		using value_t = float;
+
+		template <typename Func>
+		void fill(Func fn)
+		{
+			auto* self = static_cast<Lookup*>(this);
+
+			self->range.for_each([&, fn, i = 0](double x) mutable {
+				self->data[i++] = static_cast<value_t>(fn(x));
+			});
+		}
+
+	};
+
+	struct linear_lut : lut_base<linear_lut> {
+		linear_range range;
+		bool transfer = false;
+		value_t data[capacity] = {};
+
+		double operator()(double x) const
+		{
+			if (x > range.start) {
+				double range_dist = range.stop - range.start;
+				double idx_f = (x - range.start) * (range.num - 1) / range_dist;
+
+				unsigned idx = min(static_cast<int>(idx_f), range.size() - 2);
+
+				if (idx < capacity - 1) {
+					double y = lerp(data[idx], data[idx + 1], idx_f - idx);
+					if (transfer) y /= x;
+					return y;
+				}
+			}
+
+			double y = data[0];
+			if (transfer) y /= range.start;
+			return y;
+		}
+
+		linear_lut(const table_args& args) :
+			range({
+				args.start,
+				args.stop,
+				args.num_elements
+				}),
+			transfer(args.transfer) {}
+
+		linear_lut(const accel_args& args) :
+			linear_lut(args.lut_args) {}
+	};
+
+	struct binlog_lut : lut_base<binlog_lut> {
+		fp_rep_range range;
+		double x_start;
+		bool transfer = false;
+		value_t data[capacity] = {};
+
+		double operator()(double x) const
+		{
+			int e = min(ilogb(x), range.stop - 1);
+
+			if (e >= range.start) {
+				int idx_int_log_part = e - range.start;
+				double idx_frac_lin_part = scalbn(x, -e) - 1;
+				double idx_f = range.num * (idx_int_log_part + idx_frac_lin_part);
+
+				unsigned idx = min(static_cast<int>(idx_f), range.size() - 2);
+
+				if (idx < capacity - 1) {
+					double y = lerp(data[idx], data[idx + 1], idx_f - idx);
+					if (transfer) y /= x;
+					return y;
+				}
+			}
+
+			double y = data[0];
+			if (transfer) y /= x_start;
+			return y;
+		}
+
+		binlog_lut(const table_args& args) :
+			range({
+				static_cast<int>(args.start),
+				static_cast<int>(args.stop),
+				args.num_elements
+				}),
+			x_start(scalbn(1, range.start)),
+			transfer(args.transfer) {}
+
+		binlog_lut(const accel_args& args) :
+			binlog_lut(args.lut_args) {}
+	};
+
+}
diff --git a/common/accel-motivity.hpp b/common/accel-motivity.hpp
index a3cb027..0c15598 100644
--- a/common/accel-motivity.hpp
+++ b/common/accel-motivity.hpp
@@ -1,100 +1,56 @@
 #pragma once
 
-#include "rawaccel-base.hpp"
+#include "accel-lookup.hpp"
 
 #include <math.h>
 
-#define RA_LOOKUP
-
 namespace rawaccel {
 
-	constexpr size_t LUT_SIZE = 601;
-
-	struct si_pair { 
-		double slope = 0;
-		double intercept = 0; 
-	};
-
-	/// <summary> Struct to hold sigmoid (s-shaped) gain implementation. </summary>
-	struct motivity {
-		double rate;
-		double limit;
+	struct sigmoid {
+		double accel;
+		double motivity;
 		double midpoint;
-		double subtractive_constant;
-
-		motivity(const accel_args& args) :
-			rate(pow(10,args.accel_motivity)), 
-			limit(2*log10(args.limit)), 
-			midpoint(log10(args.midpoint))
-		{
-			subtractive_constant = limit / 2;
-		}
+		double constant;
 
-		double operator()(double speed) const 
-		{
-			double log_speed = log10(speed);
-			return pow(10, limit / (exp(-rate * (log_speed - midpoint)) + 1) - subtractive_constant);
+		sigmoid(const accel_args& args) :
+			accel(exp(args.accel_motivity)),
+			motivity(2 * log(args.motivity)),
+			midpoint(log(args.midpoint)),
+			constant(-motivity / 2) {}
 
-		}
-
-		double apply(si_pair* lookup, double speed) const
+		double operator()(double x) const
 		{
-			si_pair pair = lookup[map(speed)];
-			return pair.slope + pair.intercept / speed;
+			double denom = exp(accel * (midpoint - log(x))) + 1;
+			return exp(motivity / denom + constant);
 		}
+	};
 
-		int map(double speed) const
-		{
-			int index = speed > 0 ? (int)(100 * log10(speed) + 201) : 0;
-
-			if (index < 0) return 0;
-			if (index >= LUT_SIZE) return LUT_SIZE - 1;
+	/// <summary> Struct to hold sigmoid (s-shaped) gain implementation. </summary>
+	struct motivity : binlog_lut {
 
-			return index;
-		}
+		using binlog_lut::operator();
 
-		void fill(si_pair* lookup) const
+		motivity(const accel_args& args) :
+			binlog_lut(args)
 		{
-			double lookup_speed = 0;
-			double integral_interval = 0;
-			double gain_integral_speed = 0;
-			double gain_integral_speed_prev = 0;
-			double gain = 0;
-			double intercept = 0;
-			double output = 0;
-			double output_prev = 0;
-			double x = -2;
-
-			double logarithm_interval = 0.01;
-			double integral_intervals_per_speed = 10;
-			double integral_interval_factor = pow(10, logarithm_interval) / integral_intervals_per_speed;
-
-			lookup[0] = {};
-
-			for (size_t i = 1; i < LUT_SIZE; i++)
-			{
-				x += logarithm_interval;
-
-				// Each lookup speed will be 10^0.01 = 2.33% higher than the previous
-				// To get 10 integral intervals per speed, set interval to 0.233%
-				lookup_speed = pow(10, x);
-				integral_interval = lookup_speed * integral_interval_factor;
-
-				while (gain_integral_speed < lookup_speed)
-				{
-					output_prev = output;
-					gain_integral_speed_prev = gain_integral_speed;
-					gain_integral_speed += integral_interval;
-					gain = operator()(gain_integral_speed);
-					output += gain * integral_interval;
+			double sum = 0;
+			double a = 0;
+			auto sigmoid_sum = [&, sig = sigmoid(args)](double b) mutable {
+				double interval = (b - a) / args.lut_args.partitions;
+				for (int i = 1; i <= args.lut_args.partitions; i++) {
+					sum += sig(a + i * interval) * interval;
 				}
-
-				intercept = output_prev - gain_integral_speed_prev * gain;
-
-				lookup[i] = { gain, intercept };
-			}
-
+				a = b;
+				return sum;
+			};
+
+			fill([&](double x) {
+				double y = sigmoid_sum(x);
+				if (!this->transfer) y /= x;
+				return y;
+			});
 		}
+
 	};
 
 }
diff --git a/common/accel-power.hpp b/common/accel-power.hpp
index c8faabb..a21f926 100644
--- a/common/accel-power.hpp
+++ b/common/accel-power.hpp
@@ -24,4 +24,5 @@ namespace rawaccel {
 		}
 	};
 
+	using power_legacy = power;
 }
diff --git a/common/accel-union.hpp b/common/accel-union.hpp
index 97496e1..c63a9cc 100644
--- a/common/accel-union.hpp
+++ b/common/accel-union.hpp
@@ -16,53 +16,59 @@ namespace rawaccel {
         jump_gain,
         natural_lgcy,
         natural_gain,
-        power,
-        motivity,
+        power_lgcy,
+        power_gain,
+        motivity_lgcy,
+        motivity_gain,
+        lut_log,
+        lut_lin,
         noaccel
     };
 
-    constexpr internal_mode make_mode(accel_mode m, bool legacy) 
+    constexpr internal_mode make_mode(accel_mode mode, table_mode lut_mode, bool legacy) 
     {
-        switch (m) {
-        case accel_mode::classic:
-            return legacy ? internal_mode::classic_lgcy : internal_mode::classic_gain;
-        case accel_mode::jump:
-            return legacy ? internal_mode::jump_lgcy : internal_mode::jump_gain;
-        case accel_mode::natural:
-            return legacy ? internal_mode::natural_lgcy : internal_mode::natural_gain;
-        case accel_mode::power:
-            return internal_mode::power;
-        case accel_mode::motivity:
-            return internal_mode::motivity;
-        default:
+        if (lut_mode != table_mode::off) {
+            switch (lut_mode) {
+            case table_mode::binlog: return internal_mode::lut_log;
+            case table_mode::linear: return internal_mode::lut_lin;
+            default: return internal_mode::noaccel;
+            }
+        }
+        else if (mode < accel_mode{} || mode >= accel_mode::noaccel) {
             return internal_mode::noaccel;
         }
+        else {
+            int im = static_cast<int>(mode) * 2 + (legacy ? 0 : 1);
+            return static_cast<internal_mode>(im);
+        }
     }
 
     constexpr internal_mode make_mode(const accel_args& args) 
     {
-        return make_mode(args.mode, args.legacy);
+        return make_mode(args.mode, args.lut_args.mode, args.legacy);
     }
 
     template <typename Visitor, typename Variant>
     inline auto visit_accel(Visitor vis, Variant&& var) 
     {
         switch (var.tag) {
-        case internal_mode::classic_lgcy: return vis(var.u.classic_l);
-        case internal_mode::classic_gain: return vis(var.u.classic_g);
-        case internal_mode::jump_lgcy: return vis(var.u.jump_l);
-        case internal_mode::jump_gain: return vis(var.u.jump_g);
-        case internal_mode::natural_lgcy: return vis(var.u.natural_l);
-        case internal_mode::natural_gain: return vis(var.u.natural_g);
-        case internal_mode::power:        return vis(var.u.power);
-        case internal_mode::motivity:     return vis(var.u.motivity);
-        default:                          return vis(var.u.noaccel);
+        case internal_mode::classic_lgcy:  return vis(var.u.classic_l);
+        case internal_mode::classic_gain:  return vis(var.u.classic_g);
+        case internal_mode::jump_lgcy:     return vis(var.u.jump_l);
+        case internal_mode::jump_gain:     return vis(var.u.jump_g);
+        case internal_mode::natural_lgcy:  return vis(var.u.natural_l);
+        case internal_mode::natural_gain:  return vis(var.u.natural_g);
+        case internal_mode::power_lgcy:    return vis(var.u.power_l);
+        case internal_mode::power_gain:    return vis(var.u.power_g);
+        case internal_mode::motivity_lgcy: return vis(var.u.motivity_l);
+        case internal_mode::motivity_gain: return vis(var.u.motivity_g);
+        case internal_mode::lut_log:       return vis(var.u.log_lut);
+        case internal_mode::lut_lin:       return vis(var.u.lin_lut);
+        default:                           return vis(var.u.noaccel);
         }
     }
 
     struct accel_variant {
-        si_pair* lookup;
-
         internal_mode tag = internal_mode::noaccel;
 
         union union_t {
@@ -72,30 +78,25 @@ namespace rawaccel {
             jump_legacy jump_l;
             natural natural_g;
             natural_legacy natural_l;
-            power power;
-            motivity motivity;
+            power power_g;
+            power_legacy power_l;
+            sigmoid motivity_l;
+            motivity motivity_g;
+            linear_lut lin_lut;
+            binlog_lut log_lut;
             accel_noaccel noaccel = {};
         } u = {};
 
-        accel_variant(const accel_args& args, si_pair* lut = nullptr) :
-            tag(make_mode(args)), lookup(lut)
+        accel_variant(const accel_args& args) :
+            tag(make_mode(args))
         {
             visit_accel([&](auto& impl) {
                 impl = { args };
             }, *this);
-
-            if (lookup && tag == internal_mode::motivity) {
-                u.motivity.fill(lookup);
-            }
-
         }
 
         double apply(double speed) const 
         {
-            if (lookup && tag == internal_mode::motivity) {
-                return u.motivity.apply(lookup, speed);
-            }
-
             return visit_accel([=](auto&& impl) {
                 return impl(speed);
             }, *this);
diff --git a/common/common.vcxitems b/common/common.vcxitems
index 6ee47ed..4cbe2b7 100644
--- a/common/common.vcxitems
+++ b/common/common.vcxitems
@@ -16,6 +16,7 @@
   <ItemGroup>
     <ClInclude Include="$(MSBuildThisFileDirectory)accel-classic.hpp" />
     <ClInclude Include="$(MSBuildThisFileDirectory)accel-jump.hpp" />
+    <ClInclude Include="$(MSBuildThisFileDirectory)accel-lookup.hpp" />
     <ClInclude Include="$(MSBuildThisFileDirectory)accel-motivity.hpp" />
     <ClInclude Include="$(MSBuildThisFileDirectory)accel-natural.hpp" />
     <ClInclude Include="$(MSBuildThisFileDirectory)accel-noaccel.hpp" />
diff --git a/common/rawaccel-base.hpp b/common/rawaccel-base.hpp
index ac60ff0..6996164 100644
--- a/common/rawaccel-base.hpp
+++ b/common/rawaccel-base.hpp
@@ -13,19 +13,38 @@ namespace rawaccel {
 
     inline constexpr size_t MAX_DEV_ID_LEN = 200;
 
+    inline constexpr size_t LUT_CAPACITY = 1025;
+
     enum class accel_mode {
         classic,
         jump,
         natural,
-        motivity,
         power,
+        motivity,
         noaccel
     };
 
+    enum class table_mode {
+        off,
+        binlog,
+        linear
+    };
+
+    struct table_args {
+        table_mode mode = table_mode::off;
+        bool transfer = true;
+        unsigned char partitions = 2;
+        short num_elements = 8;
+        double start = 0;
+        double stop = 8;
+    };
+
     struct accel_args {
         accel_mode mode = accel_mode::noaccel;
         bool legacy = false;
 
+        table_args lut_args = {};
+
         double offset = 0;
         double cap = 1.5;
         double accel_classic = 0.005;
diff --git a/common/rawaccel-io.hpp b/common/rawaccel-io.hpp
index 0d3ddee..da496fa 100644
--- a/common/rawaccel-io.hpp
+++ b/common/rawaccel-io.hpp
@@ -1,14 +1,15 @@
 #pragma once
 
-#include <system_error>
+#include "rawaccel-io-def.h"
+#include "rawaccel.hpp"
+#include "rawaccel-version.h"
+#include "rawaccel-error.hpp"
 
 #define NOMINMAX
+#define WIN32_LEAN_AND_MEAN
 #include <Windows.h>
 
-#include "rawaccel-io-def.h"
-#include "rawaccel-base.hpp"
-#include "rawaccel-version.h"
-#include "rawaccel-error.hpp"
+#include <system_error>
 
 #pragma warning(push)
 #pragma warning(disable:4245) // int -> DWORD conversion while passing CTL_CODE
@@ -45,18 +46,14 @@ namespace rawaccel {
 		}
 	}
 
-	inline settings read() 
+	inline void read(io_t& args)
 	{
-		settings args;
-		io_control(RA_READ, NULL, 0, &args, sizeof(settings));
-		return args;
+		io_control(RA_READ, NULL, 0, &args, sizeof(io_t));
 	}
 
-
-	inline void write(const settings& args) 
+	inline void write(const io_t& args) 
 	{
-		auto in_ptr = const_cast<settings*>(&args);
-		io_control(RA_WRITE, in_ptr, sizeof(settings), NULL, 0);
+		io_control(RA_WRITE, const_cast<io_t*>(&args), sizeof(io_t), NULL, 0);
 	}
 
 	inline version_t get_version() 
diff --git a/common/rawaccel.hpp b/common/rawaccel.hpp
index f6bc0fd..5d3ee15 100644
--- a/common/rawaccel.hpp
+++ b/common/rawaccel.hpp
@@ -145,7 +145,7 @@ namespace rawaccel {
         vec2d sensitivity = { 1, 1 };
         vec2d directional_multipliers = {};
 
-        mouse_modifier(const settings& args, vec2<si_pair*> luts = {}) :
+        mouse_modifier(const settings& args) :
             by_component(!args.combine_mags),
             dpi_factor(1000 / args.dpi),
             speed_cap(args.speed_cap)
@@ -172,8 +172,8 @@ namespace rawaccel {
                 return;
             }
 
-            accels.x = accel_variant(args.argsv.x, luts.x);
-            accels.y = accel_variant(args.argsv.y, luts.y);
+            accels.x = accel_variant(args.argsv.x);
+            accels.y = accel_variant(args.argsv.y);
 
             distance = weighted_distance(args.dom_args);
             directional = direction_weight(args.range_weights);
@@ -243,4 +243,9 @@ namespace rawaccel {
         mouse_modifier() = default;
     };
 
+    struct io_t {
+        settings args;
+        mouse_modifier mod;
+    };
+
 } // rawaccel
diff --git a/common/utility.hpp b/common/utility.hpp
index ae14b48..de90d44 100644
--- a/common/utility.hpp
+++ b/common/utility.hpp
@@ -34,6 +34,33 @@ namespace rawaccel {
         return minsd(maxsd(v, lo), hi);
     }
 
+    template <typename T>
+    constexpr const T& min(const T& a, const T& b)
+    {
+        return (b < a) ? b : a;
+    }
+
+    template <typename T>
+    constexpr const T& max(const T& a, const T& b)
+    {
+        return (b < a) ? a : b;
+    }
+
+    template <typename T>
+    constexpr const T& clamp(const T& v, const T& lo, const T& hi)
+    {
+        return (v < lo) ? lo : (hi < v) ? hi : v;
+    }
+
+    constexpr double lerp(double a, double b, double t)
+    {
+        double x = a + t * (b - a);
+        if ((t > 1) == (a < b)) {
+            return maxsd(x, b);
+        }
+        return minsd(x, b);
+    }
+
     // returns the unbiased exponent of x if x is normal 
     inline int ilogb(double x)
     {
@@ -41,6 +68,14 @@ namespace rawaccel {
 	    return static_cast<int>((u.i >> 52) & 0x7ff) - 0x3ff;
     }
 
+    // returns x * 2^n if n is in [-1022, 1023] 
+    inline double scalbn(double x, int n)
+    {
+        union { double f; unsigned long long i; } u;
+        u.i = static_cast<unsigned long long>(0x3ff + n) << 52;
+        return x * u.f;
+    }
+
     inline bool infnan(double x)
     {
 	    return ilogb(x) == 0x400;
diff --git a/driver/driver.cpp b/driver/driver.cpp
index 49f375d..89f09dc 100644
--- a/driver/driver.cpp
+++ b/driver/driver.cpp
@@ -13,13 +13,11 @@
 #endif
 
 using milliseconds = double;
-using lut_value_t = ra::si_pair;
 
 struct {
     ra::settings args;
     milliseconds tick_interval = 0; // set in DriverEntry
     ra::mouse_modifier modifier;
-    vec2<lut_value_t*> lookups = {};
 } global = {};
 
 VOID
@@ -164,7 +162,7 @@ Return Value:
     case RA_READ:
         status = WdfRequestRetrieveOutputBuffer(
             Request,
-            sizeof(ra::settings),
+            sizeof(ra::io_t),
             &buffer,
             NULL
         );
@@ -172,14 +170,18 @@ Return Value:
             DebugPrint(("RetrieveOutputBuffer failed: 0x%x\n", status));
         }
         else {
-            *reinterpret_cast<ra::settings*>(buffer) = global.args;
-            bytes_out = sizeof(ra::settings);
+            ra::io_t& output = *reinterpret_cast<ra::io_t*>(buffer);
+
+            output.args = global.args;
+            output.mod = global.modifier;
+
+            bytes_out = sizeof(ra::io_t);
         }
         break;
     case RA_WRITE:
         status = WdfRequestRetrieveInputBuffer(
             Request,
-            sizeof(ra::settings),
+            sizeof(ra::io_t),
             &buffer,
             NULL
         );
@@ -191,14 +193,10 @@ Return Value:
             interval.QuadPart = static_cast<LONGLONG>(ra::WRITE_DELAY) * -10000;
             KeDelayExecutionThread(KernelMode, FALSE, &interval);
 
-            ra::settings new_settings = *reinterpret_cast<ra::settings*>(buffer);
-
-            if (new_settings.time_min <= 0 || _isnanf(static_cast<float>(new_settings.time_min))) {
-                new_settings.time_min = ra::settings{}.time_min;
-            }
+            ra::io_t& input = *reinterpret_cast<ra::io_t*>(buffer);
 
-            global.args = new_settings;
-            global.modifier = { global.args, global.lookups };
+            global.args = input.args;
+            global.modifier = input.mod;
         }
         break;
     case RA_GET_VERSION:
@@ -276,23 +274,6 @@ Routine Description:
         KeQueryPerformanceCounter(&freq);
         global.tick_interval = 1e3 / freq.QuadPart;
 
-        auto make_lut = [] {
-            const size_t POOL_SIZE = sizeof(lut_value_t) * ra::LUT_SIZE;
-
-            auto pool = ExAllocatePoolWithTag(NonPagedPool, POOL_SIZE, 'AR');
-
-            if (!pool) {
-                DebugPrint(("RA - failed to allocate LUT\n"));
-            }
-            else {
-                RtlZeroMemory(pool, POOL_SIZE);
-            }
-
-            return reinterpret_cast<lut_value_t*>(pool);
-        };
-
-        global.lookups = { make_lut(), make_lut() };
-
         CreateControlDevice(driver);
     }
     else {
diff --git a/driver/driver.h b/driver/driver.h
index bb1ae41..56742f6 100644
--- a/driver/driver.h
+++ b/driver/driver.h
@@ -2,6 +2,7 @@
 
 #include "rawaccel-base.hpp"
 
+#define NOMINMAX
 #include <ntddk.h>
 #include <kbdmou.h>
 #include <wdf.h>