Merge branch 'feat_spec_params' into work

hw/xbox/nv2a/pgraph/gl/renderer.h

@@ -122,6 +122,7 @@ typedef struct ShaderBinding {
     GLint light_infinite_direction_loc[NV2A_MAX_LIGHTS];
     GLint light_local_position_loc[NV2A_MAX_LIGHTS];
     GLint light_local_attenuation_loc[NV2A_MAX_LIGHTS];
+    int specular_power_loc;
 
     GLint clip_region_loc[8];
 

hw/xbox/nv2a/pgraph/gl/shaders.c

@@ -193,8 +193,11 @@ static void update_shader_constant_locations(ShaderBinding *binding)
     if (binding->state.fixed_function) {
         binding->material_alpha_loc =
             glGetUniformLocation(binding->gl_program, "material_alpha");
+        binding->specular_power_loc =
+            glGetUniformLocation(binding->gl_program, "specularPower");
     } else {
         binding->material_alpha_loc = -1;
+        binding->specular_power_loc = -1;
     }
 }
 
@@ -836,6 +839,10 @@ static void shader_update_constants(PGRAPHState *pg, ShaderBinding *binding,
             }
         }
 
+        if (binding->specular_power_loc != -1) {
+    	    glUniform1f(binding->specular_power_loc, pg->specular_power);
+	    }
+
         /* estimate the viewport by assuming it matches the surface ... */
         unsigned int aa_width = 1, aa_height = 1;
         pgraph_apply_anti_aliasing_factor(pg, &aa_width, &aa_height);

hw/xbox/nv2a/pgraph/glsl/vsh-ff.c

@@ -238,6 +238,8 @@ GLSL_DEFINE(materialEmissionColor, GLSL_LTCTXA(NV_IGRAPH_XF_LTCTXA_CM_COL) ".xyz
     } else {
         //FIXME: Do 2 passes if we want 2 sided-lighting?
 
+        mstring_append_fmt(uniforms, "%sfloat specularPower;\n", u);
+
         static char alpha_source_diffuse[] = "diffuse.a";
         static char alpha_source_specular[] = "specular.a";
         static char alpha_source_material[] = "material_alpha";
@@ -301,7 +303,7 @@ GLSL_DEFINE(materialEmissionColor, GLSL_LTCTXA(NV_IGRAPH_XF_LTCTXA_CM_COL) ".xyz
                     "    float nDotVP = max(0.0, dot(tNormal, VP));\n"
                     "    float nDotHV = max(0.0, dot(tNormal, halfVector));\n",
                     i, i, i, i, i,
-                    state->local_eye ? "VPeye" : "vec3(0.0, 0.0, -1.0)"
+                    state->local_eye ? "VPeye" : "vec3(0.0, 0.0, 0.0)"
                 );
             }
 
@@ -361,27 +363,15 @@ GLSL_DEFINE(materialEmissionColor, GLSL_LTCTXA(NV_IGRAPH_XF_LTCTXA_CM_COL) ".xyz
                 "    if (nDotVP == 0.0) {\n"
                 "      pf = 0.0;\n"
                 "    } else {\n"
-                "      pf = pow(nDotHV, %f);\n"
+                "      pf = pow(nDotHV, specularPower);\n"
                 "    }\n"
                 "    vec3 lightAmbient = lightAmbientColor(%d) * attenuation;\n"
                 "    vec3 lightDiffuse = lightDiffuseColor(%d) * attenuation * nDotVP;\n"
                 "    vec3 lightSpecular = lightSpecularColor(%d) * attenuation * pf;\n",
-                state->specular_power, i, i, i);
+                i, i, i);
 
-            switch (state->ambient_src) {
-            case MATERIAL_COLOR_SRC_MATERIAL:
-                mstring_append(body,
-                               "    oD0.rgb += lightAmbient;\n");
-                break;
-            case MATERIAL_COLOR_SRC_DIFFUSE:
-                mstring_append(body,
-                               "    oD0.rgb += diffuse.rgb * lightAmbient;\n");
-                break;
-            case MATERIAL_COLOR_SRC_SPECULAR:
-                mstring_append(body,
-                               "    oD0.rgb += specular.rgb * lightAmbient;\n");
-                break;
-            }
+            mstring_append(body,
+                "    oD0.xyz += lightAmbient;\n");
 
             switch (state->diffuse_src) {
             case MATERIAL_COLOR_SRC_MATERIAL:

hw/xbox/nv2a/pgraph/pgraph.c

@@ -1814,39 +1814,110 @@ DEF_METHOD_INC(NV097, SET_FOG_PLANE)
     pg->vsh_constants_dirty[NV_IGRAPH_XF_XFCTX_FOG] = true;
 }
 
-// Based on curve fitting to observed values from DirectX uses.
-//    x = -log2(30.80722523) / power
-//    c3_param = -1.01441946 * pow(exp2( x ), 2) + 0.01451685
-#define SPECULAR_POWER_NUMERATOR_CONSTANT    30.80722523f
-#define SPECULAR_POWER_COEFFICIENT_A         -1.0141946f
-#define SPECULAR_POWER_CONSTANT_COEFFICIENT  0.01451685f
-//#define SPECULAR_POWER_LOG_CONSTANT  (-2.f * log2(SPECULAR_POWER_NUMERATOR_CONSTANT))
-#define SPECULAR_POWER_LOG_CONSTANT -9.8903942108154297f
-static float reconstruct_specular_power_from_c3(uint32_t c3_parameter)
-{
-    float c3 = *(float*)&c3_parameter;
+struct CurveCoefficients {
+  float a;
+  float b;
+  float c;
+};
 
-    // FIXME: This handling is not correct, but is visually distinct without causing a crash.
-    // It does not appear possible for a DirectX-generated value to be positive, so while this differs from hardware
-    // behavior, it may be irrelevant in practice.
-    float invert = 1.f;
-    if (c3 > 0.0f) {
-        invert = -1.f;
-        c3 *= invert;
+static const struct CurveCoefficients curve_coefficients[] = {
+  {1.000108475163, -9.838607076280, 54.829089549713},
+  {1.199164441703, -3.292603784852, 7.799987995214},
+  {8.653441252033, 29.189473787191, 43.586027561823},
+  {-531.307758450301, 117.398468683934, 113.155490738338},
+  {-4.662713151292, 1.221108944572, 1.217360986939},
+  {-124.435242105211, 35.401219563514, 35.408114377045},
+  {10672560.259502287954, 21565843.555823743343, 10894794.336297152564},
+  {-51973801.463933646679, -104199997.554352939129, -52225454.356278456748},
+  {972270.324080004124, 2025882.096547174733, 1054898.052467488218},
+};
+
+static const float kCoefficient0StepPoints[] = {
+  -0.022553957999, // power = 1.25
+  -0.421539008617, // power = 4.00
+  -0.678715527058, // power = 9.00
+  -0.838916420937, // power = 20.00
+  -0.961754500866, // power = 90.00
+  -0.990773200989, // power = 375.00
+  -0.994858562946, // power = 650.00
+  -0.996561050415, // power = 1000.00
+  -0.999547004700, // power = 1250.00
+};
+
+static float reconstruct_quadratic(float c0, const struct CurveCoefficients *coefficients) {
+  return coefficients->a + coefficients->b * c0 + coefficients->c * c0 * c0;
+}
+
+static float reconstruct_saturation_growth_rate(float c0, const struct CurveCoefficients *coefficients) {
+  return (coefficients->a * c0) / (coefficients->b + coefficients->c * c0);
+}
+
+static float (* const reconstruct_func_map[])(float, const struct CurveCoefficients *) = {
+  reconstruct_quadratic, // 1.0..1.25 max error 0.01 %
+  reconstruct_quadratic, // 1.25..4.0 max error 2.2 %
+  reconstruct_quadratic, // 4.0..9.0 max error 2.3 %
+  reconstruct_saturation_growth_rate, // 9.0..20.0 max error 1.4 %
+  reconstruct_saturation_growth_rate, // 20.0..90.0 max error 2.1 %
+  reconstruct_saturation_growth_rate, // 90.0..375.0 max error 2.8%
+  reconstruct_quadratic, // 375..650 max error 1.0 %
+  reconstruct_quadratic, // 650..1000 max error 1.7%
+  reconstruct_quadratic, // 1000..1250 max error 1.0%
+};
+
+static float reconstruct_specular_power(const float *params) {
+  // See https://github.com/dracc/xgu/blob/db3172d8c983629f0dc971092981846da22438ae/xgux.h#L279
+
+  // Values < 1.0 will result in a positive c1 and (c2 - c0 * 2) will be very
+  // close to the original value.
+  if (params[1] > 0.0f && params[2] < 1.0f) {
+    return params[2] - (params[0] * 2.0f);
+  }
+
+  float c0 = params[0];
+  float c3 = params[3];
+  // FIXME: This handling is not correct, but is distinct without crashing.
+  // It does not appear possible for a DirectX-generated value to be positive,
+  // so while this differs from hardware behavior, it may be irrelevant in
+  // practice.
+  if (c0 > 0.0f || c3 > 0.0f) {
+    return 0.0001f;
+  }
+
+  float reconstructed_power = 0.f;
+  for (uint32_t i = 0; i < sizeof(kCoefficient0StepPoints) / sizeof(kCoefficient0StepPoints[0]); ++i) {
+    if (c0 > kCoefficient0StepPoints[i]) {
+      reconstructed_power = reconstruct_func_map[i](c0, &curve_coefficients[i]);
+      break;
     }
+  }
 
-    c3 -= SPECULAR_POWER_CONSTANT_COEFFICIENT;
-    float ret = SPECULAR_POWER_LOG_CONSTANT / log2(c3 / SPECULAR_POWER_COEFFICIENT_A);
-    assert(!isnan(ret) && "Failed to reconstruct specular power factor");
-    return ret * invert;
+  float reconstructed_half_power = 0.f;
+  for (uint32_t i = 0; i < sizeof(kCoefficient0StepPoints) / sizeof(kCoefficient0StepPoints[0]); ++i) {
+    if (c3 > kCoefficient0StepPoints[i]) {
+      reconstructed_half_power = reconstruct_func_map[i](c3, &curve_coefficients[i]);
+      break;
+    }
+  }
+
+  // The range can be extended beyond 1250 by using the half power params. This
+  // will only work for DirectX generated values, arbitrary params could
+  // erroneously trigger this.
+  //
+  // There are some very low power (~1) values that have inverted powers, but
+  // they are easily identified by comparatively high c0 parameters.
+  if (reconstructed_power == 0.f || (reconstructed_half_power > reconstructed_power && c0 < -0.1f)) {
+    return reconstructed_half_power * 2.f;
+  }
+
+  return reconstructed_power;
 }
 
 DEF_METHOD_INC(NV097, SET_SPECULAR_PARAMS)
 {
     int slot = (method - NV097_SET_SPECULAR_PARAMS) / 4;
-    NV2A_DPRINTF("NV097_SET_SPECULAR_PARAMS[%d] 0x%X\n", slot, parameter);
-    if (slot == 3) {
-        pg->specular_power = reconstruct_specular_power_from_c3(parameter);
+    pg->specular_params[slot] = *(float *)&parameter;
+    if (slot == 5) {
+        pg->specular_power = reconstruct_specular_power(pg->specular_params);
     }
 }
 
@@ -2831,9 +2902,9 @@ DEF_METHOD_INC(NV097, SET_SPECULAR_FOG_FACTOR)
 DEF_METHOD_INC(NV097, SET_SPECULAR_PARAMS_BACK)
 {
     int slot = (method - NV097_SET_SPECULAR_PARAMS_BACK) / 4;
-    NV2A_DPRINTF("SET_SPECULAR_PARAMS_BACK[%d] 0x%X\n", slot, parameter);
-    if (slot == 3) {
-        pg->specular_power_back = reconstruct_specular_power_from_c3(parameter);
+    pg->specular_params_back[slot] = *(float *)&parameter;
+    if (slot == 5) {
+        pg->specular_power_back = reconstruct_specular_power(pg->specular_params_back);
     }
 }
 

hw/xbox/nv2a/pgraph/pgraph.h

@@ -197,7 +197,9 @@ typedef struct PGRAPHState {
     float light_local_position[NV2A_MAX_LIGHTS][3];
     float light_local_attenuation[NV2A_MAX_LIGHTS][3];
 
+    float specular_params[6];
     float specular_power;
+    float specular_params_back[6];
     float specular_power_back;
 
     float point_params[8];

hw/xbox/nv2a/pgraph/vk/renderer.h

@@ -190,6 +190,7 @@ typedef struct ShaderBinding {
     int light_infinite_direction_loc[NV2A_MAX_LIGHTS];
     int light_local_position_loc[NV2A_MAX_LIGHTS];
     int light_local_attenuation_loc[NV2A_MAX_LIGHTS];
+    int specular_power_loc;
 
     int clip_region_loc;