nv2a: add new NV_PGRAPH parameters, fix W_param tests, ZMinMaxControl tests, see through and missing shadows

hw/xbox/nv2a/nv2a_regs.h

@@ -461,6 +461,7 @@
 #       define NV_PGRAPH_CONTROL_2_STENCIL_OP_V_INCR                7
 #       define NV_PGRAPH_CONTROL_2_STENCIL_OP_V_DECR                8
 #define NV_PGRAPH_CONTROL_3                              0x00001958
+#   define NV_PGRAPH_CONTROL_3_TEXTURE_PERSPECTIVE_ENABLE       (1 << 6)
 #   define NV_PGRAPH_CONTROL_3_SHADEMODE                        (1 << 7)
 #       define NV_PGRAPH_CONTROL_3_SHADEMODE_FLAT                   0
 #       define NV_PGRAPH_CONTROL_3_SHADEMODE_SMOOTH                 1
@@ -619,6 +620,7 @@
 #define NV_PGRAPH_WINDOWCLIPY6                           0x00001A7C
 #define NV_PGRAPH_WINDOWCLIPY7                           0x00001A80
 #define NV_PGRAPH_ZCOMPRESSOCCLUDE                       0x00001A84
+#   define NV_PGRAPH_ZCOMPRESSOCCLUDE_CULL_NEAR_FAR_EN          (1 << 0)
 #   define NV_PGRAPH_ZCOMPRESSOCCLUDE_ZCLAMP_EN                 (1 << 4)
 #       define NV_PGRAPH_ZCOMPRESSOCCLUDE_ZCLAMP_EN_CULL            0
 #       define NV_PGRAPH_ZCOMPRESSOCCLUDE_ZCLAMP_EN_CLAMP           1
@@ -879,6 +881,7 @@
 #       define NV097_SET_CONTROL0_STENCIL_WRITE_ENABLE            (1 << 0)
 #       define NV097_SET_CONTROL0_Z_FORMAT                        (1 << 12)
 #       define NV097_SET_CONTROL0_Z_PERSPECTIVE_ENABLE            (1 << 16)
+#       define NV097_SET_CONTROL0_TEXTURE_PERSPECTIVE_ENABLE      (1 << 20)
 #   define NV097_SET_COLOR_MATERIAL                           0x00000298
 #   define NV097_SET_FOG_MODE                                 0x0000029C
 #       define NV097_SET_FOG_MODE_V_LINEAR                        0x2601
@@ -1233,6 +1236,7 @@
 #       define NV097_SET_ZMIN_MAX_CONTROL_ZCLAMP_EN               0x000000F0
 #       define NV097_SET_ZMIN_MAX_CONTROL_ZCLAMP_EN_CULL              0
 #       define NV097_SET_ZMIN_MAX_CONTROL_ZCLAMP_EN_CLAMP             1
+#       define NV097_SET_ZMIN_MAX_CONTROL_CULL_NEAR_FAR_EN_TRUE   (1 << 0)
 #   define NV097_SET_ANTI_ALIASING_CONTROL                    0x00001D7C
 #       define NV097_SET_ANTI_ALIASING_CONTROL_ENABLE             (1 << 0)
 #   define NV097_SET_ZSTENCIL_CLEAR_VALUE                     0x00001D8C

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

@@ -219,7 +219,9 @@ static void generate_shaders(ShaderBinding *binding)
                                  state->polygon_back_mode,
                                  state->primitive_mode,
                                  state->smooth_shading,
-                                 false);
+                                 false,
+                                 state->z_perspective || state->texture_perspective
+            );
     if (geometry_shader_code) {
         const char* geometry_shader_code_str =
              mstring_get_str(geometry_shader_code);
@@ -891,6 +893,17 @@ static void shader_update_constants(PGRAPHState *pg, ShaderBinding *binding,
         float zclip_max = *(float *)&v[1];
         glUniform4f(binding->clip_range_loc, 0, zmax, zclip_min, zclip_max);
     }
+    if (binding->zbias_loc != -1) {
+        float zbias = 0.0f;
+        if (pgraph_reg_r(pg, NV_PGRAPH_SETUPRASTER) &
+            (NV_PGRAPH_SETUPRASTER_POFFSETFILLENABLE |
+             NV_PGRAPH_SETUPRASTER_POFFSETLINEENABLE |
+             NV_PGRAPH_SETUPRASTER_POFFSETPOINTENABLE)) {
+            uint32_t zbias_u32 = pgraph_reg_r(pg, NV_PGRAPH_ZOFFSETBIAS);
+            zbias = *(float *)&zbias_u32;
+        }
+        glUniform1f(binding->zbias_loc, zbias);
+    }
 
     if (binding->depth_offset_loc != -1) {
         float zbias = 0.0f;
@@ -1029,7 +1042,6 @@ void pgraph_gl_bind_shaders(PGRAPHState *pg)
     bool binding_changed = false;
     if (r->shader_binding && !test_shaders_dirty(pg) && !pg->program_data_dirty) {
         nv2a_profile_inc_counter(NV2A_PROF_SHADER_BIND_NOTDIRTY);
-        goto update_constants;
     }
 
     ShaderBinding *old_binding = r->shader_binding;
@@ -1069,7 +1081,6 @@ void pgraph_gl_bind_shaders(PGRAPHState *pg)
 
     NV2A_GL_DGROUP_END();
 
-update_constants:
     assert(r->shader_binding);
     assert(r->shader_binding->initialized);
     shader_update_constants(pg, r->shader_binding, binding_changed);

hw/xbox/nv2a/pgraph/glsl/common.c

@@ -52,5 +52,9 @@ MString *pgraph_get_glsl_vtx_header(MString *out, bool location, bool smooth, bo
                            suffix_s);
     }
 
+    if (location) {
+        mstring_append(out, " layout(location = 9) ");
+    }
+    mstring_append_fmt(out, "%s float %sdepthBuf%s;\n", in_out_s, prefix_s, suffix_s);
     return out;
 }

hw/xbox/nv2a/pgraph/glsl/common.h

@@ -33,6 +33,6 @@
 
 #define GLSL_DEFINE(a, b) "#define " stringify(a) " " b "\n"
 
-MString *pgraph_get_glsl_vtx_header(MString *out, bool location, bool smooth, bool in, bool prefix, bool array);
+MString *pgraph_get_glsl_vtx_header(MString *out, bool location, bool smooth, bool in, bool prefix, bool array, bool z_perspective);
 
 #endif

hw/xbox/nv2a/pgraph/glsl/geom.c

@@ -27,7 +27,8 @@ MString *pgraph_gen_geom_glsl(enum ShaderPolygonMode polygon_front_mode,
                               enum ShaderPolygonMode polygon_back_mode,
                               enum ShaderPrimitiveMode primitive_mode,
                               bool smooth_shading,
-                              bool vulkan)
+                              bool vulkan,
+                              bool z_perspective)
 {
     /* FIXME: Missing support for 2-sided-poly mode */
     assert(polygon_front_mode == polygon_back_mode);
@@ -174,14 +175,15 @@ MString *pgraph_gen_geom_glsl(enum ShaderPolygonMode polygon_front_mode,
     mstring_append(s, layout_in);
     mstring_append(s, layout_out);
     mstring_append(s, "\n");
-    pgraph_get_glsl_vtx_header(s, vulkan, smooth_shading, true, true, true);
-    pgraph_get_glsl_vtx_header(s, vulkan, smooth_shading, false, false, false);
+    pgraph_get_glsl_vtx_header(s, vulkan, smooth_shading, true, true, true, z_perspective);
+    pgraph_get_glsl_vtx_header(s, vulkan, smooth_shading, false, false, false, z_perspective);
 
     if (smooth_shading) {
         mstring_append(s,
                        "void emit_vertex(int index, int _unused) {\n"
                        "  gl_Position = gl_in[index].gl_Position;\n"
                        "  gl_PointSize = gl_in[index].gl_PointSize;\n"
+                       "  depthBuf = v_depthBuf[index];\n"
                        "  vtxD0 = v_vtxD0[index];\n"
                        "  vtxD1 = v_vtxD1[index];\n"
                        "  vtxB0 = v_vtxB0[index];\n"
@@ -198,6 +200,7 @@ MString *pgraph_gen_geom_glsl(enum ShaderPolygonMode polygon_front_mode,
                        "void emit_vertex(int index, int provoking_index) {\n"
                        "  gl_Position = gl_in[index].gl_Position;\n"
                        "  gl_PointSize = gl_in[index].gl_PointSize;\n"
+                       "  depthBuf = v_depthBuf[index];\n"
                        "  vtxD0 = v_vtxD0[provoking_index];\n"
                        "  vtxD1 = v_vtxD1[provoking_index];\n"
                        "  vtxB0 = v_vtxB0[provoking_index];\n"

hw/xbox/nv2a/pgraph/glsl/geom.h

@@ -29,6 +29,7 @@ MString *pgraph_gen_geom_glsl(enum ShaderPolygonMode polygon_front_mode,
                               enum ShaderPolygonMode polygon_back_mode,
                               enum ShaderPrimitiveMode primitive_mode,
                               bool smooth_shading,
-                              bool vulkan);
+                              bool vulkan,
+                              bool z_perspective);
 
 #endif

hw/xbox/nv2a/pgraph/glsl/psh.c

@@ -733,11 +733,13 @@ static void apply_convolution_filter(const struct PixelShader *ps, MString *vars
 
 static MString* psh_convert(struct PixelShader *ps)
 {
+    bool z_perspective = ps->state.z_perspective;
+    bool tex = ps->state.texture_perspective;
     const char *u = ps->state.vulkan ? "" : "uniform "; // FIXME: Remove
 
     MString *preflight = mstring_new();
     pgraph_get_glsl_vtx_header(preflight, ps->state.vulkan,
-                             ps->state.smooth_shading, true, false, false);
+                             ps->state.smooth_shading, true, false, false, tex || z_perspective);
 
     if (ps->state.vulkan) {
         mstring_append_fmt(preflight,
@@ -869,6 +871,43 @@ static MString* psh_convert(struct PixelShader *ps)
     }
 
     /* Depth clipping */
+    /* OGL/VK zbias is applied to z coord, so we need to apply it 
+     * to w coord manually when w-buffering is enabled */
+    mstring_append(clip, "float w = depthBuf + zbias;\n"); 
+    if (ps->state.vulkan) {
+        mstring_append(clip, "float z = gl_FragCoord.z;\n");
+    } else {
+        /* Changing clipping range to [-1, 1] here 
+         * prevents floating point precission loss in OGL */
+        mstring_append(clip, "float z = gl_FragCoord.z * 2.0 - 1.0;\n");
+    }
+
+    const char *z = ps->state.z_perspective ? "w" : "z * clipRange.y";
+    if (ps->state.clipping) {
+        mstring_append_fmt(clip,
+            "if (%s < clipRange.z || clipRange.w < %s) {\n"
+            "  discard;\n"
+            "}\n", z, z
+        );
+    }
+    if (ps->state.near_far) {
+        /* FIXME: set correct custom clip planes 
+         * (see pgraph ZMinMaxControl tests)*/
+        mstring_append(clip,  
+            // near plane
+            "if (z * clipRange.y < clipRange.z*2.0/3.0) {\n"
+            "  discard;\n"
+            "}\n"
+            // z=n->f far plane
+            //"if (z * clipRange.y > clipRange.w*1.2) {\n"   
+            // z=inc far plane
+            "if (z * clipRange.y > clipRange.w + 2.0*(clipRange.w + clipRange.z)/(clipRange.w - clipRange.z) - 1) {\n"  
+            "  discard;\n"
+            "}\n"           
+        );        
+    }
+
+    /* calculate perspective-correct inputs
     if (ps->state.depth_clipping) {
         if (ps->state.z_perspective) {
             mstring_append(
@@ -877,7 +916,7 @@ static MString* psh_convert(struct PixelShader *ps)
                       "  discard;\n"
                       "}\n");
         } else {
-            /* Take care of floating point precision problems. MS dashboard
+             Take care of floating point precision problems. MS dashboard
              * outputs exactly 0.0 z-coordinates and then our fixed function
              * vertex shader outputs -w as the z-coordinate when OpenGL is
              * used. Since -w/w = -1, this should give us exactly 0.0 as
@@ -893,7 +932,7 @@ static MString* psh_convert(struct PixelShader *ps)
              * depth buffer.) Radeon OpenGL problem could also be fixed by using
              * glClipControl(), but it requires OpenGL 4.5.
              * Above is based on experiments with Linux and Mesa.
-             */
+
             if (ps->state.vulkan) {
                 mstring_append(
                     clip, "if (gl_FragCoord.z*clipRange.y < clipRange.z ||\n"
@@ -908,7 +947,7 @@ static MString* psh_convert(struct PixelShader *ps)
                           "}\n");
             }
         }
-    }
+    } */
 
     MString *vars = mstring_new();
     mstring_append(vars, "vec4 pD0 = vtxD0;\n");
@@ -919,6 +958,7 @@ static MString* psh_convert(struct PixelShader *ps)
     mstring_append(vars, "vec4 pT0 = vtxT0;\n");
     mstring_append(vars, "vec4 pT1 = vtxT1;\n");
     mstring_append(vars, "vec4 pT2 = vtxT2;\n");
+
     if (ps->state.point_sprite) {
         assert(!ps->state.rect_tex[3]);
         mstring_append(vars, "vec4 pT3 = vec4(gl_PointCoord, 1.0, 1.0);\n");
@@ -1245,6 +1285,17 @@ static MString* psh_convert(struct PixelShader *ps)
         }
     }
 
+    /* NV097_SET_CONTROL0_Z_PERSPECTIVE_ENABLE enables w-buffering
+     * not only gl_Position gets divided by the homogeneous coordinate, 
+     * but also all other interpolated variables, which requires 
+     * the division to be after the rasterization */
+    if (z_perspective) {   
+        mstring_append(ps->code, "gl_FragDepth = w / clipRange.y;\n");
+    }
+    else if (!ps->state.vulkan) {
+        mstring_append(ps->code, "gl_FragDepth = z;\n");
+    }
+
     for (int i = 0; i < ps->num_var_refs; i++) {
         mstring_append_fmt(vars, "vec4 %s = vec4(0);\n", ps->var_refs[i]);
         if (strcmp(ps->var_refs[i], "r0") == 0) {

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

@@ -423,9 +423,19 @@ GLSL_DEFINE(materialEmissionColor, GLSL_LTCTXA(NV_IGRAPH_XF_LTCTXA_CM_COL) ".xyz
     mstring_append(body,
     "   oPos = tPosition * compositeMat;\n"
     "   oPos.w = clampAwayZeroInf(oPos.w);\n"
+    "   depthBuf = (invViewport * vec4(oPos.xyz/oPos.w, oPos.w)).w;\n"
     "   oPos = invViewport * oPos;\n"
     );
 
+    if (!state->texture_perspective) {
+        mstring_append(body,
+            "    if (oPos.w >= 0.0) {\n"
+            "      oPos.xyz /= oPos.w;\n"
+            "      oPos.w = 1.0;\n"
+            "    }\n"  
+        );  
+    }
+    mstring_append(body,"   oPos = invViewport * oPos;\n");
     if (state->vulkan) {
         mstring_append(body, "   oPos.y *= -1;\n");
     }

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

@@ -33,6 +33,7 @@
 #include <assert.h>
 
 #include "hw/xbox/nv2a/pgraph/vsh.h"
+#include "hw/xbox/nv2a/pgraph/shaders.h"
 #include "common.h"
 #include "vsh-prog.h"
 
@@ -790,11 +791,13 @@ static const char* vsh_header =
     "}\n";
 
 void pgraph_gen_vsh_prog_glsl(uint16_t version,
-                   const uint32_t *tokens,
-                   unsigned int length,
-                   bool vulkan,
+                   const ShaderState *state,
                    MString *header, MString *body)
 {
+    const uint32_t *tokens = (uint32_t *)state->program_data;
+    unsigned int length = state->program_length;
+    bool texture = state->texture_perspective;
+    bool vulkan = state->vulkan;
 
     mstring_append(header, vsh_header);
 
@@ -837,6 +840,7 @@ void pgraph_gen_vsh_prog_glsl(uint16_t version,
     }
 
     mstring_append(body,
+        "  depthBuf = oPos.w;\n"
         "  oPos.z = oPos.z / clipRange.y;\n"
         "  oPos.w = clampAwayZeroInf(oPos.w);\n"
 

hw/xbox/nv2a/pgraph/glsl/vsh-prog.h

@@ -28,8 +28,7 @@
 #ifndef HW_XBOX_NV2A_PGRAPH_GLSL_VSH_PROG_H
 #define HW_XBOX_NV2A_PGRAPH_GLSL_VSH_PROG_H
 
-void pgraph_gen_vsh_prog_glsl(uint16_t version, const uint32_t *tokens,
-                              unsigned int length,
-                              bool vulkan, MString *header, MString *body);
+void pgraph_gen_vsh_prog_glsl(uint16_t version, const ShaderState *state,
+                              MString *header, MString *body);
 
 #endif

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

@@ -85,10 +85,11 @@ MString *pgraph_gen_vsh_glsl(const ShaderState *state, bool prefix_outputs)
         "}\n");
 
     pgraph_get_glsl_vtx_header(header, state->vulkan, state->smooth_shading,
-                             false, prefix_outputs, false);
+                             false, prefix_outputs, false, state->texture_perspective || state->z_perspective);
 
     if (prefix_outputs) {
         mstring_append(header,
+                       "#define depthBuf v_depthBuf\n"
                        "#define vtxD0 v_vtxD0\n"
                        "#define vtxD1 v_vtxD1\n"
                        "#define vtxB0 v_vtxB0\n"
@@ -148,10 +149,8 @@ MString *pgraph_gen_vsh_glsl(const ShaderState *state, bool prefix_outputs)
     if (state->fixed_function) {
         pgraph_gen_vsh_ff_glsl(state, header, body, uniforms);
     } else if (state->vertex_program) {
-        pgraph_gen_vsh_prog_glsl(VSH_VERSION_XVS,
-                                 (uint32_t *)state->program_data,
-                                 state->program_length,
-                                 state->vulkan, header, body);
+        pgraph_gen_vsh_prog_glsl(VSH_VERSION_XVS, state, header, body);
+
     } else {
         assert(false);
     }
@@ -241,6 +240,7 @@ MString *pgraph_gen_vsh_glsl(const ShaderState *state, bool prefix_outputs)
     }
 
     /* Set outputs */
+
     mstring_append(body, "\n"
                    "  vtxD0 = clamp(oD0, 0.0, 1.0);\n"
                    "  vtxD1 = clamp(oD1, 0.0, 1.0);\n"
@@ -252,19 +252,9 @@ MString *pgraph_gen_vsh_glsl(const ShaderState *state, bool prefix_outputs)
                    "  vtxT2 = oT2;\n"
                    "  vtxT3 = oT3;\n"
                    "  gl_PointSize = oPts.x;\n"
-    );
-
-    if (state->vulkan) {
-        mstring_append(body,
                    "  gl_Position = oPos;\n"
-        );
-    } else {
-        mstring_append(body,
-                   "  gl_Position = vec4(oPos.x, oPos.y, 2.0*oPos.z - oPos.w, oPos.w);\n"
-        );
-    }
-
-    mstring_append(body, "}\n");
+                   "}\n"
+    );
 
     /* Return combined header + source */
     if (state->vulkan) {

hw/xbox/nv2a/pgraph/pgraph.c

@@ -1073,6 +1073,12 @@ DEF_METHOD(NV097, SET_CONTROL0)
     PG_SET_MASK(NV_PGRAPH_CONTROL_0,
              NV_PGRAPH_CONTROL_0_Z_PERSPECTIVE_ENABLE,
              z_perspective);
+
+    bool texture_perspective =
+        parameter & NV097_SET_CONTROL0_TEXTURE_PERSPECTIVE_ENABLE;
+    PG_SET_MASK(NV_PGRAPH_CONTROL_3,
+             NV_PGRAPH_CONTROL_3_TEXTURE_PERSPECTIVE_ENABLE,
+             texture_perspective);
 }
 
 DEF_METHOD(NV097, SET_COLOR_MATERIAL)
@@ -2683,6 +2689,8 @@ DEF_METHOD(NV097, SET_ZMIN_MAX_CONTROL)
         assert(!"Invalid zclamp value");
         break;
     }
+    PG_SET_MASK(NV_PGRAPH_ZCOMPRESSOCCLUDE, NV_PGRAPH_ZCOMPRESSOCCLUDE_CULL_NEAR_FAR_EN,
+            GET_MASK(parameter, NV097_SET_ZMIN_MAX_CONTROL_CULL_NEAR_FAR_EN_TRUE));
 }
 
 DEF_METHOD(NV097, SET_ANTI_ALIASING_CONTROL)