[dxvk] Refactor resolve shaders

Removes some code duplication while adding support for non-AVERAGE
resolves for plain color images.

Also nuke the AMD path since it is no longer used.

src/dxvk/dxvk_meta_resolve.cpp

@@ -8,7 +8,6 @@
 #include <dxvk_resolve_frag_d.h>
 #include <dxvk_resolve_frag_ds.h>
 #include <dxvk_resolve_frag_f.h>
-#include <dxvk_resolve_frag_f_amd.h>
 #include <dxvk_resolve_frag_u.h>
 #include <dxvk_resolve_frag_i.h>
 
@@ -52,9 +51,7 @@ namespace dxvk {
 
   DxvkMetaResolveObjects::DxvkMetaResolveObjects(const DxvkDevice* device)
   : m_vkd         (device->vkd()),
-    m_shaderFragF (device->features().amdShaderFragmentMask
+    m_shaderFragF (createShaderModule(dxvk_resolve_frag_f)),
-      ? createShaderModule(dxvk_resolve_frag_f_amd)
-      : createShaderModule(dxvk_resolve_frag_f)),
     m_shaderFragU (createShaderModule(dxvk_resolve_frag_u)),
     m_shaderFragI (createShaderModule(dxvk_resolve_frag_i)),
     m_shaderFragD (createShaderModule(dxvk_resolve_frag_d)) {

src/dxvk/meson.build

@@ -50,7 +50,6 @@ dxvk_shaders = files([
   'shaders/dxvk_resolve_frag_d.frag',
   'shaders/dxvk_resolve_frag_ds.frag',
   'shaders/dxvk_resolve_frag_f.frag',
-  'shaders/dxvk_resolve_frag_f_amd.frag',
   'shaders/dxvk_resolve_frag_i.frag',
   'shaders/dxvk_resolve_frag_u.frag',
 

src/dxvk/shaders/dxvk_resolve_common.glsl

@@ -0,0 +1,35 @@
+#define VK_RESOLVE_MODE_NONE            (0)
+#define VK_RESOLVE_MODE_SAMPLE_ZERO_BIT (1 << 0)
+#define VK_RESOLVE_MODE_AVERAGE_BIT     (1 << 1)
+#define VK_RESOLVE_MODE_MIN_BIT         (1 << 2)
+#define VK_RESOLVE_MODE_MAX_BIT         (1 << 3)
+
+#define resolve_fn(name, type, load_fn)               \
+type name(ivec3 coord, int samples, uint mode) {      \
+  if (mode == VK_RESOLVE_MODE_NONE)                   \
+    return type(0);                                   \
+  type value = load_fn(coord, 0);                     \
+                                                      \
+  switch (mode) {                                     \
+    case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT:             \
+      return value;                                   \
+                                                      \
+    case VK_RESOLVE_MODE_AVERAGE_BIT:                 \
+      for (int i = 1; i < samples; i++)               \
+        value += load_fn(coord, i);                   \
+      value /= type(c_samples);                       \
+      break;                                          \
+                                                      \
+    case VK_RESOLVE_MODE_MIN_BIT:                     \
+      for (int i = 1; i < samples; i++)               \
+        value = min(value, load_fn(coord, i));        \
+      break;                                          \
+                                                      \
+    case VK_RESOLVE_MODE_MAX_BIT:                     \
+      for (int i = 1; i < c_samples; i++)             \
+        value = min(value, load_fn(coord, i));        \
+      break;                                          \
+  }                                                   \
+                                                      \
+  return value;                                       \
+}

src/dxvk/shaders/dxvk_resolve_frag_d.frag

@@ -1,56 +1,28 @@
 #version 450
 
+#extension GL_GOOGLE_include_directive : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
-#define VK_RESOLVE_MODE_NONE            (0)
+#include "dxvk_resolve_common.glsl"
-#define VK_RESOLVE_MODE_SAMPLE_ZERO_BIT (1 << 0)
-#define VK_RESOLVE_MODE_AVERAGE_BIT     (1 << 1)
-#define VK_RESOLVE_MODE_MIN_BIT         (1 << 2)
-#define VK_RESOLVE_MODE_MAX_BIT         (1 << 3)
 
 layout(constant_id = 0) const int c_samples = 1;
-layout(constant_id = 1) const int c_mode_d  = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
+layout(constant_id = 1) const int c_mode    = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
 
 layout(binding = 0) uniform texture2DMSArray s_depth;
 
+float load_depth(ivec3 coord, int s) {
+  return texelFetch(s_depth, coord, s).r;
+}
+
+resolve_fn(resolve_depth, float, load_depth)
+
 layout(push_constant)
 uniform u_info_t {
   ivec2 offset;
 } u_info;
 
-float resolve_depth(ivec3 coord) {
-  float depth = 0.0f;
-
-  switch (c_mode_d) {
-    case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      break;
-    
-    case VK_RESOLVE_MODE_AVERAGE_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth += texelFetch(s_depth, coord, i).r;
-      depth /= float(c_samples);
-      break;
-    
-    case VK_RESOLVE_MODE_MIN_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth = min(depth, texelFetch(s_depth, coord, i).r);
-      break;
-    
-    case VK_RESOLVE_MODE_MAX_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth = max(depth, texelFetch(s_depth, coord, i).r);
-      break;
-  }
-
-  return depth;
-}
-
 void main() {
   ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
 
-  gl_FragDepth = resolve_depth(coord);
+  gl_FragDepth = resolve_depth(coord, c_samples, c_mode);
 }

src/dxvk/shaders/dxvk_resolve_frag_ds.frag

@@ -1,13 +1,10 @@
 #version 450
 
+#extension GL_GOOGLE_include_directive : enable
 #extension GL_ARB_shader_stencil_export : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
-#define VK_RESOLVE_MODE_NONE            (0)
+#include "dxvk_resolve_common.glsl"
-#define VK_RESOLVE_MODE_SAMPLE_ZERO_BIT (1 << 0)
-#define VK_RESOLVE_MODE_AVERAGE_BIT     (1 << 1)
-#define VK_RESOLVE_MODE_MIN_BIT         (1 << 2)
-#define VK_RESOLVE_MODE_MAX_BIT         (1 << 3)
 
 layout(constant_id = 0) const int c_samples = 1;
 layout(constant_id = 1) const int c_mode_d  = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
@@ -16,69 +13,25 @@ layout(constant_id = 2) const int c_mode_s  = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
 layout(binding = 0) uniform  texture2DMSArray s_depth;
 layout(binding = 1) uniform utexture2DMSArray s_stencil;
 
+float load_depth(ivec3 coord, int s) {
+  return texelFetch(s_depth, coord, s).r;
+}
+
+uint load_stencil(ivec3 coord, int s) {
+  return texelFetch(s_stencil, coord, s).r;
+}
+
+resolve_fn(resolve_depth, float, load_depth)
+resolve_fn(resolve_stencil, uint, load_stencil)
+
 layout(push_constant)
 uniform u_info_t {
   ivec2 offset;
 } u_info;
 
-float resolve_depth(ivec3 coord) {
-  float depth = 0.0f;
-
-  switch (c_mode_d) {
-    case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      break;
-    
-    case VK_RESOLVE_MODE_AVERAGE_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth += texelFetch(s_depth, coord, i).r;
-      depth /= float(c_samples);
-      break;
-    
-    case VK_RESOLVE_MODE_MIN_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth = min(depth, texelFetch(s_depth, coord, i).r);
-      break;
-    
-    case VK_RESOLVE_MODE_MAX_BIT:
-      depth = texelFetch(s_depth, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        depth = max(depth, texelFetch(s_depth, coord, i).r);
-      break;
-  }
-
-  return depth;
-}
-
-int resolve_stencil(ivec3 coord) {
-  uint stencil = 0u;
-
-  switch (c_mode_s) {
-    case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT:
-      stencil = texelFetch(s_stencil, coord, 0).r;
-      break;
-    
-    case VK_RESOLVE_MODE_MIN_BIT:
-      stencil = texelFetch(s_stencil, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        stencil = min(stencil, texelFetch(s_stencil, coord, i).r);
-      break;
-    
-    case VK_RESOLVE_MODE_MAX_BIT:
-      stencil = texelFetch(s_stencil, coord, 0).r;
-      for (int i = 1; i < c_samples; i++)
-        stencil = max(stencil, texelFetch(s_stencil, coord, i).r);
-      break;
-  }
-
-  return int(stencil);
-}
-
 void main() {
   ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
 
-  gl_FragDepth         = resolve_depth(coord);
+  gl_FragDepth = resolve_depth(coord, c_samples, c_mode_d);
-  gl_FragStencilRefARB = resolve_stencil(coord);
+  gl_FragStencilRefARB = int(resolve_stencil(coord, c_samples, c_mode_s));
 }

src/dxvk/shaders/dxvk_resolve_frag_f.frag

@@ -1,13 +1,23 @@
 #version 450
 
+#extension GL_GOOGLE_include_directive : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
+#include "dxvk_resolve_common.glsl"
+
 layout(constant_id = 0) const int c_samples = 1;
+layout(constant_id = 1) const int c_mode    = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
 
 layout(binding = 0) uniform texture2DMSArray s_image;
 
 layout(location = 0) out vec4 o_color;
 
+vec4 load_color(ivec3 coord, int s) {
+  return texelFetch(s_image, coord, s);
+}
+
+resolve_fn(resolve_color, vec4, load_color)
+
 layout(push_constant)
 uniform u_info_t {
   ivec2 offset;
@@ -15,8 +25,5 @@ uniform u_info_t {
 
 void main() {
   ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
-  vec4 color = vec4(0.0f);
+  o_color = resolve_color(coord, c_samples, c_mode);
-  for (int i = 0; i < c_samples; i++)
-    color += texelFetch(s_image, coord, i);
-  o_color = color / float(c_samples);
 }

src/dxvk/shaders/dxvk_resolve_frag_f_amd.frag

@@ -1,56 +0,0 @@
-#version 450
-
-#extension GL_EXT_samplerless_texture_functions : enable
-#extension GL_EXT_spirv_intrinsics : enable
-
-// GL_AMD_shader_fragment_mask was never updated to support
-// sampler-less functions, so we have to define these manually
-spirv_instruction(extensions = ["SPV_AMD_shader_fragment_mask"], capabilities = [5010], id = 5011)
-uint fragment_mask_fetch(texture2DMSArray tex, ivec3 coord);
-
-spirv_instruction(extensions = ["SPV_AMD_shader_fragment_mask"], capabilities = [5010], id = 5012)
-vec4 fragment_fetch(texture2DMSArray tex, ivec3 coord, uint index);
-
-layout(constant_id = 0) const int c_samples = 1;
-
-layout(set = 0, binding = 0)
-uniform texture2DMSArray s_image;
-
-layout(location = 0) out vec4 o_color;
-
-layout(push_constant)
-uniform u_info_t {
-  ivec2 offset;
-} u_info;
-
-void main() {
-  ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
-
-  // get a four-bit fragment index for each sample
-  uint fragMask = fragment_mask_fetch(s_image, coord);
-
-  // count number of occurences of each fragment
-  // index in one four-bit counter for each sample
-  uint fragCount = 0u;
-
-  for (int i = 0; i < 4 * c_samples; i += 4) {
-    uint fragIndex = bitfieldExtract(fragMask, i, 4);
-    fragCount += 1u << (fragIndex << 2);
-  }
-
-  // perform necessary texture lookups to compute
-  // final fragment color
-  o_color = vec4(0.0f);
-  
-  while (fragCount != 0) {
-    int fragIndex = findLSB(fragCount) >> 2;
-    int fragShift = fragIndex << 2;
-
-    o_color += fragment_fetch(s_image, coord, fragIndex)
-      * float(bitfieldExtract(fragCount, fragShift, 4));
-
-    fragCount = bitfieldInsert(fragCount, 0, fragShift, 4);
-  }
-
-  o_color /= float(c_samples);
-}

src/dxvk/shaders/dxvk_resolve_frag_i.frag

@@ -1,11 +1,23 @@
 #version 450
 
+#extension GL_GOOGLE_include_directive : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
+#include "dxvk_resolve_common.glsl"
+
+layout(constant_id = 0) const int c_samples = 1;
+layout(constant_id = 1) const int c_mode    = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
+
 layout(binding = 0) uniform itexture2DMSArray s_image;
 
 layout(location = 0) out ivec4 o_color;
 
+ivec4 load_color(ivec3 coord, int s) {
+  return texelFetch(s_image, coord, s);
+}
+
+resolve_fn(resolve_color, ivec4, load_color)
+
 layout(push_constant)
 uniform u_info_t {
   ivec2 offset;
@@ -13,5 +25,5 @@ uniform u_info_t {
 
 void main() {
   ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
-  o_color = texelFetch(s_image, coord, 0);
+  o_color = resolve_color(coord, c_samples, c_mode);
 }

src/dxvk/shaders/dxvk_resolve_frag_u.frag

@@ -1,10 +1,22 @@
 #version 450
 
+#extension GL_GOOGLE_include_directive : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
+#include "dxvk_resolve_common.glsl"
+
+layout(constant_id = 0) const int c_samples = 1;
+layout(constant_id = 1) const int c_mode    = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
+
 layout(binding = 0) uniform utexture2DMSArray s_image;
 
-layout(location = 0) out uvec4 o_color;
+layout(location = 0) out vec4 o_color;
+
+uvec4 load_color(ivec3 coord, int s) {
+  return texelFetch(s_image, coord, s);
+}
+
+resolve_fn(resolve_color, uvec4, load_color)
 
 layout(push_constant)
 uniform u_info_t {
@@ -13,5 +25,5 @@ uniform u_info_t {
 
 void main() {
   ivec3 coord = ivec3(gl_FragCoord.xy + u_info.offset, gl_Layer);
-  o_color = texelFetch(s_image, coord, 0);
+  o_color = resolve_color(coord, c_samples, c_mode);
 }