summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorGravatar liamwhite2023-01-19 14:58:53 -0500
committerGravatar GitHub2023-01-19 14:58:53 -0500
commit475370c8f89002e3b508eb152b981a5b89049d68 (patch)
tree79a145e24f2bf5ab5baca6824cae6cd525a8d170
parentMerge pull request #9623 from liamwhite/wp-oops (diff)
parentAddress feedback (diff)
downloadyuzu-475370c8f89002e3b508eb152b981a5b89049d68.tar.gz
yuzu-475370c8f89002e3b508eb152b981a5b89049d68.tar.xz
yuzu-475370c8f89002e3b508eb152b981a5b89049d68.zip
Merge pull request #9556 from vonchenplus/draw_texture
video_core: Implement maxwell3d draw texture method
-rw-r--r--src/video_core/CMakeLists.txt2
-rw-r--r--src/video_core/engines/draw_manager.cpp31
-rw-r--r--src/video_core/engines/draw_manager.h20
-rw-r--r--src/video_core/engines/maxwell_3d.cpp1
-rw-r--r--src/video_core/engines/maxwell_3d.h16
-rw-r--r--src/video_core/host_shaders/CMakeLists.txt2
-rw-r--r--src/video_core/host_shaders/blit_color_float.frag (renamed from src/video_core/host_shaders/vulkan_blit_color_float.frag)0
-rw-r--r--src/video_core/host_shaders/full_screen_triangle.vert13
-rw-r--r--src/video_core/rasterizer_interface.h3
-rw-r--r--src/video_core/renderer_null/null_rasterizer.cpp1
-rw-r--r--src/video_core/renderer_null/null_rasterizer.h1
-rw-r--r--src/video_core/renderer_opengl/blit_image.cpp59
-rw-r--r--src/video_core/renderer_opengl/blit_image.h38
-rw-r--r--src/video_core/renderer_opengl/gl_device.cpp1
-rw-r--r--src/video_core/renderer_opengl/gl_device.h7
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.cpp44
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.h4
-rw-r--r--src/video_core/renderer_opengl/gl_shader_manager.cpp121
-rw-r--r--src/video_core/renderer_opengl/gl_shader_manager.h129
-rw-r--r--src/video_core/renderer_vulkan/blit_image.cpp88
-rw-r--r--src/video_core/renderer_vulkan/blit_image.h6
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.cpp29
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.h1
-rw-r--r--src/video_core/texture_cache/texture_cache.h7
-rw-r--r--src/video_core/texture_cache/texture_cache_base.h3
25 files changed, 502 insertions, 125 deletions
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index f617665de..b474eb363 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -100,6 +100,8 @@ add_library(video_core STATIC
100 renderer_null/null_rasterizer.h 100 renderer_null/null_rasterizer.h
101 renderer_null/renderer_null.cpp 101 renderer_null/renderer_null.cpp
102 renderer_null/renderer_null.h 102 renderer_null/renderer_null.h
103 renderer_opengl/blit_image.cpp
104 renderer_opengl/blit_image.h
103 renderer_opengl/gl_buffer_cache.cpp 105 renderer_opengl/gl_buffer_cache.cpp
104 renderer_opengl/gl_buffer_cache.h 106 renderer_opengl/gl_buffer_cache.h
105 renderer_opengl/gl_compute_pipeline.cpp 107 renderer_opengl/gl_compute_pipeline.cpp
diff --git a/src/video_core/engines/draw_manager.cpp b/src/video_core/engines/draw_manager.cpp
index 2437121ce..1d22d25f1 100644
--- a/src/video_core/engines/draw_manager.cpp
+++ b/src/video_core/engines/draw_manager.cpp
@@ -51,6 +51,10 @@ void DrawManager::ProcessMethodCall(u32 method, u32 argument) {
51 LOG_WARNING(HW_GPU, "(STUBBED) called"); 51 LOG_WARNING(HW_GPU, "(STUBBED) called");
52 break; 52 break;
53 } 53 }
54 case MAXWELL3D_REG_INDEX(draw_texture.src_y0): {
55 DrawTexture();
56 break;
57 }
54 default: 58 default:
55 break; 59 break;
56 } 60 }
@@ -179,6 +183,33 @@ void DrawManager::DrawIndexSmall(u32 argument) {
179 ProcessDraw(true, 1); 183 ProcessDraw(true, 1);
180} 184}
181 185
186void DrawManager::DrawTexture() {
187 const auto& regs{maxwell3d->regs};
188 draw_texture_state.dst_x0 = static_cast<float>(regs.draw_texture.dst_x0) / 4096.f;
189 draw_texture_state.dst_y0 = static_cast<float>(regs.draw_texture.dst_y0) / 4096.f;
190 const auto dst_width = static_cast<float>(regs.draw_texture.dst_width) / 4096.f;
191 const auto dst_height = static_cast<float>(regs.draw_texture.dst_height) / 4096.f;
192 const bool lower_left{regs.window_origin.mode !=
193 Maxwell3D::Regs::WindowOrigin::Mode::UpperLeft};
194 if (lower_left) {
195 draw_texture_state.dst_y0 -= dst_height;
196 }
197 draw_texture_state.dst_x1 = draw_texture_state.dst_x0 + dst_width;
198 draw_texture_state.dst_y1 = draw_texture_state.dst_y0 + dst_height;
199 draw_texture_state.src_x0 = static_cast<float>(regs.draw_texture.src_x0) / 4096.f;
200 draw_texture_state.src_y0 = static_cast<float>(regs.draw_texture.src_y0) / 4096.f;
201 draw_texture_state.src_x1 =
202 (static_cast<float>(regs.draw_texture.dx_du) / 4294967296.f) * dst_width +
203 draw_texture_state.src_x0;
204 draw_texture_state.src_y1 =
205 (static_cast<float>(regs.draw_texture.dy_dv) / 4294967296.f) * dst_height +
206 draw_texture_state.src_y0;
207 draw_texture_state.src_sampler = regs.draw_texture.src_sampler;
208 draw_texture_state.src_texture = regs.draw_texture.src_texture;
209
210 maxwell3d->rasterizer->DrawTexture();
211}
212
182void DrawManager::UpdateTopology() { 213void DrawManager::UpdateTopology() {
183 const auto& regs{maxwell3d->regs}; 214 const auto& regs{maxwell3d->regs};
184 switch (regs.primitive_topology_control) { 215 switch (regs.primitive_topology_control) {
diff --git a/src/video_core/engines/draw_manager.h b/src/video_core/engines/draw_manager.h
index 58d1b2d59..7c22c49f1 100644
--- a/src/video_core/engines/draw_manager.h
+++ b/src/video_core/engines/draw_manager.h
@@ -32,6 +32,19 @@ public:
32 std::vector<u8> inline_index_draw_indexes; 32 std::vector<u8> inline_index_draw_indexes;
33 }; 33 };
34 34
35 struct DrawTextureState {
36 f32 dst_x0;
37 f32 dst_y0;
38 f32 dst_x1;
39 f32 dst_y1;
40 f32 src_x0;
41 f32 src_y0;
42 f32 src_x1;
43 f32 src_y1;
44 u32 src_sampler;
45 u32 src_texture;
46 };
47
35 struct IndirectParams { 48 struct IndirectParams {
36 bool is_indexed; 49 bool is_indexed;
37 bool include_count; 50 bool include_count;
@@ -64,6 +77,10 @@ public:
64 return draw_state; 77 return draw_state;
65 } 78 }
66 79
80 const DrawTextureState& GetDrawTextureState() const {
81 return draw_texture_state;
82 }
83
67 IndirectParams& GetIndirectParams() { 84 IndirectParams& GetIndirectParams() {
68 return indirect_state; 85 return indirect_state;
69 } 86 }
@@ -81,6 +98,8 @@ private:
81 98
82 void DrawIndexSmall(u32 argument); 99 void DrawIndexSmall(u32 argument);
83 100
101 void DrawTexture();
102
84 void UpdateTopology(); 103 void UpdateTopology();
85 104
86 void ProcessDraw(bool draw_indexed, u32 instance_count); 105 void ProcessDraw(bool draw_indexed, u32 instance_count);
@@ -89,6 +108,7 @@ private:
89 108
90 Maxwell3D* maxwell3d{}; 109 Maxwell3D* maxwell3d{};
91 State draw_state{}; 110 State draw_state{};
111 DrawTextureState draw_texture_state{};
92 IndirectParams indirect_state{}; 112 IndirectParams indirect_state{};
93}; 113};
94} // namespace Tegra::Engines 114} // namespace Tegra::Engines
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 68ceda519..ae9da6290 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -149,6 +149,7 @@ bool Maxwell3D::IsMethodExecutable(u32 method) {
149 case MAXWELL3D_REG_INDEX(inline_index_4x8.index0): 149 case MAXWELL3D_REG_INDEX(inline_index_4x8.index0):
150 case MAXWELL3D_REG_INDEX(vertex_array_instance_first): 150 case MAXWELL3D_REG_INDEX(vertex_array_instance_first):
151 case MAXWELL3D_REG_INDEX(vertex_array_instance_subsequent): 151 case MAXWELL3D_REG_INDEX(vertex_array_instance_subsequent):
152 case MAXWELL3D_REG_INDEX(draw_texture.src_y0):
152 case MAXWELL3D_REG_INDEX(wait_for_idle): 153 case MAXWELL3D_REG_INDEX(wait_for_idle):
153 case MAXWELL3D_REG_INDEX(shadow_ram_control): 154 case MAXWELL3D_REG_INDEX(shadow_ram_control):
154 case MAXWELL3D_REG_INDEX(load_mme.instruction_ptr): 155 case MAXWELL3D_REG_INDEX(load_mme.instruction_ptr):
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 0b2fd2928..c89969bb4 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -1599,6 +1599,20 @@ public:
1599 }; 1599 };
1600 static_assert(sizeof(TIRModulationCoeff) == 0x4); 1600 static_assert(sizeof(TIRModulationCoeff) == 0x4);
1601 1601
1602 struct DrawTexture {
1603 s32 dst_x0;
1604 s32 dst_y0;
1605 s32 dst_width;
1606 s32 dst_height;
1607 s64 dx_du;
1608 s64 dy_dv;
1609 u32 src_sampler;
1610 u32 src_texture;
1611 s32 src_x0;
1612 s32 src_y0;
1613 };
1614 static_assert(sizeof(DrawTexture) == 0x30);
1615
1602 struct ReduceColorThreshold { 1616 struct ReduceColorThreshold {
1603 union { 1617 union {
1604 BitField<0, 8, u32> all_hit_once; 1618 BitField<0, 8, u32> all_hit_once;
@@ -2751,7 +2765,7 @@ public:
2751 u32 reserved_sw_method2; ///< 0x102C 2765 u32 reserved_sw_method2; ///< 0x102C
2752 std::array<TIRModulationCoeff, 5> tir_modulation_coeff; ///< 0x1030 2766 std::array<TIRModulationCoeff, 5> tir_modulation_coeff; ///< 0x1030
2753 std::array<u32, 15> spare_nop; ///< 0x1044 2767 std::array<u32, 15> spare_nop; ///< 0x1044
2754 INSERT_PADDING_BYTES_NOINIT(0x30); 2768 DrawTexture draw_texture; ///< 0x1080
2755 std::array<u32, 7> reserved_sw_method3_to_7; ///< 0x10B0 2769 std::array<u32, 7> reserved_sw_method3_to_7; ///< 0x10B0
2756 ReduceColorThreshold reduce_color_thresholds_unorm8; ///< 0x10CC 2770 ReduceColorThreshold reduce_color_thresholds_unorm8; ///< 0x10CC
2757 std::array<u32, 4> reserved_sw_method10_to_13; ///< 0x10D0 2771 std::array<u32, 4> reserved_sw_method10_to_13; ///< 0x10D0
diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index f275b2aa9..e968ae220 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -11,6 +11,7 @@ set(GLSL_INCLUDES
11 11
12set(SHADER_FILES 12set(SHADER_FILES
13 astc_decoder.comp 13 astc_decoder.comp
14 blit_color_float.frag
14 block_linear_unswizzle_2d.comp 15 block_linear_unswizzle_2d.comp
15 block_linear_unswizzle_3d.comp 16 block_linear_unswizzle_3d.comp
16 convert_abgr8_to_d24s8.frag 17 convert_abgr8_to_d24s8.frag
@@ -36,7 +37,6 @@ set(SHADER_FILES
36 smaa_blending_weight_calculation.frag 37 smaa_blending_weight_calculation.frag
37 smaa_neighborhood_blending.vert 38 smaa_neighborhood_blending.vert
38 smaa_neighborhood_blending.frag 39 smaa_neighborhood_blending.frag
39 vulkan_blit_color_float.frag
40 vulkan_blit_depth_stencil.frag 40 vulkan_blit_depth_stencil.frag
41 vulkan_fidelityfx_fsr_easu_fp16.comp 41 vulkan_fidelityfx_fsr_easu_fp16.comp
42 vulkan_fidelityfx_fsr_easu_fp32.comp 42 vulkan_fidelityfx_fsr_easu_fp32.comp
diff --git a/src/video_core/host_shaders/vulkan_blit_color_float.frag b/src/video_core/host_shaders/blit_color_float.frag
index c0c832296..c0c832296 100644
--- a/src/video_core/host_shaders/vulkan_blit_color_float.frag
+++ b/src/video_core/host_shaders/blit_color_float.frag
diff --git a/src/video_core/host_shaders/full_screen_triangle.vert b/src/video_core/host_shaders/full_screen_triangle.vert
index 2c976b19f..d16d98995 100644
--- a/src/video_core/host_shaders/full_screen_triangle.vert
+++ b/src/video_core/host_shaders/full_screen_triangle.vert
@@ -4,13 +4,20 @@
4#version 450 4#version 450
5 5
6#ifdef VULKAN 6#ifdef VULKAN
7#define VERTEX_ID gl_VertexIndex
7#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants { 8#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
8#define END_PUSH_CONSTANTS }; 9#define END_PUSH_CONSTANTS };
9#define UNIFORM(n) 10#define UNIFORM(n)
11#define FLIPY 1
10#else // ^^^ Vulkan ^^^ // vvv OpenGL vvv 12#else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
13#define VERTEX_ID gl_VertexID
11#define BEGIN_PUSH_CONSTANTS 14#define BEGIN_PUSH_CONSTANTS
12#define END_PUSH_CONSTANTS 15#define END_PUSH_CONSTANTS
16#define FLIPY -1
13#define UNIFORM(n) layout (location = n) uniform 17#define UNIFORM(n) layout (location = n) uniform
18out gl_PerVertex {
19 vec4 gl_Position;
20};
14#endif 21#endif
15 22
16BEGIN_PUSH_CONSTANTS 23BEGIN_PUSH_CONSTANTS
@@ -21,8 +28,8 @@ END_PUSH_CONSTANTS
21layout(location = 0) out vec2 texcoord; 28layout(location = 0) out vec2 texcoord;
22 29
23void main() { 30void main() {
24 float x = float((gl_VertexIndex & 1) << 2); 31 float x = float((VERTEX_ID & 1) << 2);
25 float y = float((gl_VertexIndex & 2) << 1); 32 float y = float((VERTEX_ID & 2) << 1);
26 gl_Position = vec4(x - 1.0, y - 1.0, 0.0, 1.0); 33 gl_Position = vec4(x - 1.0, FLIPY * (y - 1.0), 0.0, 1.0);
27 texcoord = fma(vec2(x, y) / 2.0, tex_scale, tex_offset); 34 texcoord = fma(vec2(x, y) / 2.0, tex_scale, tex_offset);
28} 35}
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 1735b6164..33e2610bc 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -47,6 +47,9 @@ public:
47 /// Dispatches an indirect draw invocation 47 /// Dispatches an indirect draw invocation
48 virtual void DrawIndirect() {} 48 virtual void DrawIndirect() {}
49 49
50 /// Dispatches an draw texture invocation
51 virtual void DrawTexture() = 0;
52
50 /// Clear the current framebuffer 53 /// Clear the current framebuffer
51 virtual void Clear(u32 layer_count) = 0; 54 virtual void Clear(u32 layer_count) = 0;
52 55
diff --git a/src/video_core/renderer_null/null_rasterizer.cpp b/src/video_core/renderer_null/null_rasterizer.cpp
index 2c11345d7..2b5c7defa 100644
--- a/src/video_core/renderer_null/null_rasterizer.cpp
+++ b/src/video_core/renderer_null/null_rasterizer.cpp
@@ -21,6 +21,7 @@ RasterizerNull::RasterizerNull(Core::Memory::Memory& cpu_memory_, Tegra::GPU& gp
21RasterizerNull::~RasterizerNull() = default; 21RasterizerNull::~RasterizerNull() = default;
22 22
23void RasterizerNull::Draw(bool is_indexed, u32 instance_count) {} 23void RasterizerNull::Draw(bool is_indexed, u32 instance_count) {}
24void RasterizerNull::DrawTexture() {}
24void RasterizerNull::Clear(u32 layer_count) {} 25void RasterizerNull::Clear(u32 layer_count) {}
25void RasterizerNull::DispatchCompute() {} 26void RasterizerNull::DispatchCompute() {}
26void RasterizerNull::ResetCounter(VideoCore::QueryType type) {} 27void RasterizerNull::ResetCounter(VideoCore::QueryType type) {}
diff --git a/src/video_core/renderer_null/null_rasterizer.h b/src/video_core/renderer_null/null_rasterizer.h
index 2112aa70e..51f896e43 100644
--- a/src/video_core/renderer_null/null_rasterizer.h
+++ b/src/video_core/renderer_null/null_rasterizer.h
@@ -31,6 +31,7 @@ public:
31 ~RasterizerNull() override; 31 ~RasterizerNull() override;
32 32
33 void Draw(bool is_indexed, u32 instance_count) override; 33 void Draw(bool is_indexed, u32 instance_count) override;
34 void DrawTexture() override;
34 void Clear(u32 layer_count) override; 35 void Clear(u32 layer_count) override;
35 void DispatchCompute() override; 36 void DispatchCompute() override;
36 void ResetCounter(VideoCore::QueryType type) override; 37 void ResetCounter(VideoCore::QueryType type) override;
diff --git a/src/video_core/renderer_opengl/blit_image.cpp b/src/video_core/renderer_opengl/blit_image.cpp
new file mode 100644
index 000000000..9a560a73b
--- /dev/null
+++ b/src/video_core/renderer_opengl/blit_image.cpp
@@ -0,0 +1,59 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#include <algorithm>
5
6#include "video_core/host_shaders/blit_color_float_frag.h"
7#include "video_core/host_shaders/full_screen_triangle_vert.h"
8#include "video_core/renderer_opengl/blit_image.h"
9#include "video_core/renderer_opengl/gl_shader_manager.h"
10#include "video_core/renderer_opengl/gl_shader_util.h"
11
12namespace OpenGL {
13
14BlitImageHelper::BlitImageHelper(ProgramManager& program_manager_)
15 : program_manager(program_manager_),
16 full_screen_vert(CreateProgram(HostShaders::FULL_SCREEN_TRIANGLE_VERT, GL_VERTEX_SHADER)),
17 blit_color_to_color_frag(
18 CreateProgram(HostShaders::BLIT_COLOR_FLOAT_FRAG, GL_FRAGMENT_SHADER)) {}
19
20BlitImageHelper::~BlitImageHelper() = default;
21
22void BlitImageHelper::BlitColor(GLuint dst_framebuffer, GLuint src_image_view, GLuint src_sampler,
23 const Region2D& dst_region, const Region2D& src_region,
24 const Extent3D& src_size) {
25 glEnable(GL_CULL_FACE);
26 glDisable(GL_COLOR_LOGIC_OP);
27 glDisable(GL_DEPTH_TEST);
28 glDisable(GL_STENCIL_TEST);
29 glDisable(GL_POLYGON_OFFSET_FILL);
30 glDisable(GL_RASTERIZER_DISCARD);
31 glDisable(GL_ALPHA_TEST);
32 glDisablei(GL_BLEND, 0);
33 glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
34 glCullFace(GL_BACK);
35 glFrontFace(GL_CW);
36 glColorMaski(0, GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
37 glDepthRangeIndexed(0, 0.0, 0.0);
38
39 program_manager.BindPresentPrograms(full_screen_vert.handle, blit_color_to_color_frag.handle);
40 glProgramUniform2f(full_screen_vert.handle, 0,
41 static_cast<float>(src_region.end.x - src_region.start.x) /
42 static_cast<float>(src_size.width),
43 static_cast<float>(src_region.end.y - src_region.start.y) /
44 static_cast<float>(src_size.height));
45 glProgramUniform2f(full_screen_vert.handle, 1,
46 static_cast<float>(src_region.start.x) / static_cast<float>(src_size.width),
47 static_cast<float>(src_region.start.y) /
48 static_cast<float>(src_size.height));
49 glViewport(std::min(dst_region.start.x, dst_region.end.x),
50 std::min(dst_region.start.y, dst_region.end.y),
51 std::abs(dst_region.end.x - dst_region.start.x),
52 std::abs(dst_region.end.y - dst_region.start.y));
53 glBindFramebuffer(GL_DRAW_FRAMEBUFFER, dst_framebuffer);
54 glBindSampler(0, src_sampler);
55 glBindTextureUnit(0, src_image_view);
56 glClear(GL_COLOR_BUFFER_BIT);
57 glDrawArrays(GL_TRIANGLES, 0, 3);
58}
59} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/blit_image.h b/src/video_core/renderer_opengl/blit_image.h
new file mode 100644
index 000000000..5a2b12d16
--- /dev/null
+++ b/src/video_core/renderer_opengl/blit_image.h
@@ -0,0 +1,38 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#pragma once
5
6#include <glad/glad.h>
7
8#include "video_core/engines/fermi_2d.h"
9#include "video_core/renderer_opengl/gl_resource_manager.h"
10#include "video_core/texture_cache/types.h"
11
12namespace OpenGL {
13
14using VideoCommon::Extent3D;
15using VideoCommon::Offset2D;
16using VideoCommon::Region2D;
17
18class ProgramManager;
19class Framebuffer;
20class ImageView;
21
22class BlitImageHelper {
23public:
24 explicit BlitImageHelper(ProgramManager& program_manager);
25 ~BlitImageHelper();
26
27 void BlitColor(GLuint dst_framebuffer, GLuint src_image_view, GLuint src_sampler,
28 const Region2D& dst_region, const Region2D& src_region,
29 const Extent3D& src_size);
30
31private:
32 ProgramManager& program_manager;
33
34 OGLProgram full_screen_vert;
35 OGLProgram blit_color_to_color_frag;
36};
37
38} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp
index cee5c3247..22ed16ebf 100644
--- a/src/video_core/renderer_opengl/gl_device.cpp
+++ b/src/video_core/renderer_opengl/gl_device.cpp
@@ -166,6 +166,7 @@ Device::Device(Core::Frontend::EmuWindow& emu_window) {
166 has_shader_int64 = HasExtension(extensions, "GL_ARB_gpu_shader_int64"); 166 has_shader_int64 = HasExtension(extensions, "GL_ARB_gpu_shader_int64");
167 has_amd_shader_half_float = GLAD_GL_AMD_gpu_shader_half_float; 167 has_amd_shader_half_float = GLAD_GL_AMD_gpu_shader_half_float;
168 has_sparse_texture_2 = GLAD_GL_ARB_sparse_texture2; 168 has_sparse_texture_2 = GLAD_GL_ARB_sparse_texture2;
169 has_draw_texture = GLAD_GL_NV_draw_texture;
169 warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel; 170 warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
170 need_fastmath_off = is_nvidia; 171 need_fastmath_off = is_nvidia;
171 can_report_memory = GLAD_GL_NVX_gpu_memory_info; 172 can_report_memory = GLAD_GL_NVX_gpu_memory_info;
diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h
index 2a72d84be..3ff8cad83 100644
--- a/src/video_core/renderer_opengl/gl_device.h
+++ b/src/video_core/renderer_opengl/gl_device.h
@@ -4,6 +4,8 @@
4#pragma once 4#pragma once
5 5
6#include <cstddef> 6#include <cstddef>
7#include <string>
8
7#include "common/common_types.h" 9#include "common/common_types.h"
8#include "core/frontend/emu_window.h" 10#include "core/frontend/emu_window.h"
9#include "shader_recompiler/stage.h" 11#include "shader_recompiler/stage.h"
@@ -146,6 +148,10 @@ public:
146 return has_sparse_texture_2; 148 return has_sparse_texture_2;
147 } 149 }
148 150
151 bool HasDrawTexture() const {
152 return has_draw_texture;
153 }
154
149 bool IsWarpSizePotentiallyLargerThanGuest() const { 155 bool IsWarpSizePotentiallyLargerThanGuest() const {
150 return warp_size_potentially_larger_than_guest; 156 return warp_size_potentially_larger_than_guest;
151 } 157 }
@@ -216,6 +222,7 @@ private:
216 bool has_shader_int64{}; 222 bool has_shader_int64{};
217 bool has_amd_shader_half_float{}; 223 bool has_amd_shader_half_float{};
218 bool has_sparse_texture_2{}; 224 bool has_sparse_texture_2{};
225 bool has_draw_texture{};
219 bool warp_size_potentially_larger_than_guest{}; 226 bool warp_size_potentially_larger_than_guest{};
220 bool need_fastmath_off{}; 227 bool need_fastmath_off{};
221 bool has_cbuf_ftou_bug{}; 228 bool has_cbuf_ftou_bug{};
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 181857d9c..7bced675c 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -64,7 +64,8 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra
64 shader_cache(*this, emu_window_, device, texture_cache, buffer_cache, program_manager, 64 shader_cache(*this, emu_window_, device, texture_cache, buffer_cache, program_manager,
65 state_tracker, gpu.ShaderNotify()), 65 state_tracker, gpu.ShaderNotify()),
66 query_cache(*this), accelerate_dma(buffer_cache), 66 query_cache(*this), accelerate_dma(buffer_cache),
67 fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache) {} 67 fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache),
68 blit_image(program_manager_) {}
68 69
69RasterizerOpenGL::~RasterizerOpenGL() = default; 70RasterizerOpenGL::~RasterizerOpenGL() = default;
70 71
@@ -320,6 +321,47 @@ void RasterizerOpenGL::DrawIndirect() {
320 buffer_cache.SetDrawIndirect(nullptr); 321 buffer_cache.SetDrawIndirect(nullptr);
321} 322}
322 323
324void RasterizerOpenGL::DrawTexture() {
325 MICROPROFILE_SCOPE(OpenGL_Drawing);
326
327 SCOPE_EXIT({ gpu.TickWork(); });
328 query_cache.UpdateCounters();
329
330 texture_cache.SynchronizeGraphicsDescriptors();
331 texture_cache.UpdateRenderTargets(false);
332
333 SyncState();
334
335 const auto& draw_texture_state = maxwell3d->draw_manager->GetDrawTextureState();
336 const auto& sampler = texture_cache.GetGraphicsSampler(draw_texture_state.src_sampler);
337 const auto& texture = texture_cache.GetImageView(draw_texture_state.src_texture);
338
339 if (device.HasDrawTexture()) {
340 state_tracker.BindFramebuffer(texture_cache.GetFramebuffer()->Handle());
341
342 glDrawTextureNV(texture.DefaultHandle(), sampler->Handle(), draw_texture_state.dst_x0,
343 draw_texture_state.dst_y0, draw_texture_state.dst_x1,
344 draw_texture_state.dst_y1, 0,
345 draw_texture_state.src_x0 / static_cast<float>(texture.size.width),
346 draw_texture_state.src_y0 / static_cast<float>(texture.size.height),
347 draw_texture_state.src_x1 / static_cast<float>(texture.size.width),
348 draw_texture_state.src_y1 / static_cast<float>(texture.size.height));
349 } else {
350 Region2D dst_region = {Offset2D{.x = static_cast<s32>(draw_texture_state.dst_x0),
351 .y = static_cast<s32>(draw_texture_state.dst_y0)},
352 Offset2D{.x = static_cast<s32>(draw_texture_state.dst_x1),
353 .y = static_cast<s32>(draw_texture_state.dst_y1)}};
354 Region2D src_region = {Offset2D{.x = static_cast<s32>(draw_texture_state.src_x0),
355 .y = static_cast<s32>(draw_texture_state.src_y0)},
356 Offset2D{.x = static_cast<s32>(draw_texture_state.src_x1),
357 .y = static_cast<s32>(draw_texture_state.src_y1)}};
358 blit_image.BlitColor(texture_cache.GetFramebuffer()->Handle(), texture.DefaultHandle(),
359 sampler->Handle(), dst_region, src_region, texture.size);
360 }
361
362 ++num_queued_commands;
363}
364
323void RasterizerOpenGL::DispatchCompute() { 365void RasterizerOpenGL::DispatchCompute() {
324 gpu_memory->FlushCaching(); 366 gpu_memory->FlushCaching();
325 ComputePipeline* const pipeline{shader_cache.CurrentComputePipeline()}; 367 ComputePipeline* const pipeline{shader_cache.CurrentComputePipeline()};
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index be4f76c18..0c45832ae 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -16,6 +16,7 @@
16#include "video_core/engines/maxwell_dma.h" 16#include "video_core/engines/maxwell_dma.h"
17#include "video_core/rasterizer_accelerated.h" 17#include "video_core/rasterizer_accelerated.h"
18#include "video_core/rasterizer_interface.h" 18#include "video_core/rasterizer_interface.h"
19#include "video_core/renderer_opengl/blit_image.h"
19#include "video_core/renderer_opengl/gl_buffer_cache.h" 20#include "video_core/renderer_opengl/gl_buffer_cache.h"
20#include "video_core/renderer_opengl/gl_device.h" 21#include "video_core/renderer_opengl/gl_device.h"
21#include "video_core/renderer_opengl/gl_fence_manager.h" 22#include "video_core/renderer_opengl/gl_fence_manager.h"
@@ -70,6 +71,7 @@ public:
70 71
71 void Draw(bool is_indexed, u32 instance_count) override; 72 void Draw(bool is_indexed, u32 instance_count) override;
72 void DrawIndirect() override; 73 void DrawIndirect() override;
74 void DrawTexture() override;
73 void Clear(u32 layer_count) override; 75 void Clear(u32 layer_count) override;
74 void DispatchCompute() override; 76 void DispatchCompute() override;
75 void ResetCounter(VideoCore::QueryType type) override; 77 void ResetCounter(VideoCore::QueryType type) override;
@@ -224,6 +226,8 @@ private:
224 AccelerateDMA accelerate_dma; 226 AccelerateDMA accelerate_dma;
225 FenceManagerOpenGL fence_manager; 227 FenceManagerOpenGL fence_manager;
226 228
229 BlitImageHelper blit_image;
230
227 boost::container::static_vector<u32, MAX_IMAGE_VIEWS> image_view_indices; 231 boost::container::static_vector<u32, MAX_IMAGE_VIEWS> image_view_indices;
228 std::array<ImageViewId, MAX_IMAGE_VIEWS> image_view_ids; 232 std::array<ImageViewId, MAX_IMAGE_VIEWS> image_view_ids;
229 boost::container::static_vector<GLuint, MAX_TEXTURES> sampler_handles; 233 boost::container::static_vector<GLuint, MAX_TEXTURES> sampler_handles;
diff --git a/src/video_core/renderer_opengl/gl_shader_manager.cpp b/src/video_core/renderer_opengl/gl_shader_manager.cpp
index d9c29d8b7..98841ae65 100644
--- a/src/video_core/renderer_opengl/gl_shader_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_manager.cpp
@@ -1,2 +1,123 @@
1// SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project 1// SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later 2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#include <glad/glad.h>
5
6#include "video_core/renderer_opengl/gl_shader_manager.h"
7
8namespace OpenGL {
9
10static constexpr std::array ASSEMBLY_PROGRAM_ENUMS{
11 GL_VERTEX_PROGRAM_NV, GL_TESS_CONTROL_PROGRAM_NV, GL_TESS_EVALUATION_PROGRAM_NV,
12 GL_GEOMETRY_PROGRAM_NV, GL_FRAGMENT_PROGRAM_NV,
13};
14
15ProgramManager::ProgramManager(const Device& device) {
16 glCreateProgramPipelines(1, &pipeline.handle);
17 if (device.UseAssemblyShaders()) {
18 glEnable(GL_COMPUTE_PROGRAM_NV);
19 }
20}
21
22void ProgramManager::BindComputeProgram(GLuint program) {
23 glUseProgram(program);
24 is_compute_bound = true;
25}
26
27void ProgramManager::BindComputeAssemblyProgram(GLuint program) {
28 if (current_assembly_compute_program != program) {
29 current_assembly_compute_program = program;
30 glBindProgramARB(GL_COMPUTE_PROGRAM_NV, program);
31 }
32 UnbindPipeline();
33}
34
35void ProgramManager::BindSourcePrograms(std::span<const OGLProgram, NUM_STAGES> programs) {
36 static constexpr std::array<GLenum, 5> stage_enums{
37 GL_VERTEX_SHADER_BIT, GL_TESS_CONTROL_SHADER_BIT, GL_TESS_EVALUATION_SHADER_BIT,
38 GL_GEOMETRY_SHADER_BIT, GL_FRAGMENT_SHADER_BIT,
39 };
40 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
41 if (current_programs[stage] != programs[stage].handle) {
42 current_programs[stage] = programs[stage].handle;
43 glUseProgramStages(pipeline.handle, stage_enums[stage], programs[stage].handle);
44 }
45 }
46 BindPipeline();
47}
48
49void ProgramManager::BindPresentPrograms(GLuint vertex, GLuint fragment) {
50 if (current_programs[0] != vertex) {
51 current_programs[0] = vertex;
52 glUseProgramStages(pipeline.handle, GL_VERTEX_SHADER_BIT, vertex);
53 }
54 if (current_programs[4] != fragment) {
55 current_programs[4] = fragment;
56 glUseProgramStages(pipeline.handle, GL_FRAGMENT_SHADER_BIT, fragment);
57 }
58 glUseProgramStages(
59 pipeline.handle,
60 GL_TESS_CONTROL_SHADER_BIT | GL_TESS_EVALUATION_SHADER_BIT | GL_GEOMETRY_SHADER_BIT, 0);
61 current_programs[1] = 0;
62 current_programs[2] = 0;
63 current_programs[3] = 0;
64
65 if (current_stage_mask != 0) {
66 current_stage_mask = 0;
67 for (const GLenum program_type : ASSEMBLY_PROGRAM_ENUMS) {
68 glDisable(program_type);
69 }
70 }
71 BindPipeline();
72}
73
74void ProgramManager::BindAssemblyPrograms(std::span<const OGLAssemblyProgram, NUM_STAGES> programs,
75 u32 stage_mask) {
76 const u32 changed_mask = current_stage_mask ^ stage_mask;
77 current_stage_mask = stage_mask;
78
79 if (changed_mask != 0) {
80 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
81 if (((changed_mask >> stage) & 1) != 0) {
82 if (((stage_mask >> stage) & 1) != 0) {
83 glEnable(ASSEMBLY_PROGRAM_ENUMS[stage]);
84 } else {
85 glDisable(ASSEMBLY_PROGRAM_ENUMS[stage]);
86 }
87 }
88 }
89 }
90 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
91 if (current_programs[stage] != programs[stage].handle) {
92 current_programs[stage] = programs[stage].handle;
93 glBindProgramARB(ASSEMBLY_PROGRAM_ENUMS[stage], programs[stage].handle);
94 }
95 }
96 UnbindPipeline();
97}
98
99void ProgramManager::RestoreGuestCompute() {}
100
101void ProgramManager::BindPipeline() {
102 if (!is_pipeline_bound) {
103 is_pipeline_bound = true;
104 glBindProgramPipeline(pipeline.handle);
105 }
106 UnbindCompute();
107}
108
109void ProgramManager::UnbindPipeline() {
110 if (is_pipeline_bound) {
111 is_pipeline_bound = false;
112 glBindProgramPipeline(0);
113 }
114 UnbindCompute();
115}
116
117void ProgramManager::UnbindCompute() {
118 if (is_compute_bound) {
119 is_compute_bound = false;
120 glUseProgram(0);
121 }
122}
123} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_shader_manager.h b/src/video_core/renderer_opengl/gl_shader_manager.h
index a84f5aeb3..07ffab77f 100644
--- a/src/video_core/renderer_opengl/gl_shader_manager.h
+++ b/src/video_core/renderer_opengl/gl_shader_manager.h
@@ -6,8 +6,6 @@
6#include <array> 6#include <array>
7#include <span> 7#include <span>
8 8
9#include <glad/glad.h>
10
11#include "video_core/renderer_opengl/gl_device.h" 9#include "video_core/renderer_opengl/gl_device.h"
12#include "video_core/renderer_opengl/gl_resource_manager.h" 10#include "video_core/renderer_opengl/gl_resource_manager.h"
13 11
@@ -16,121 +14,28 @@ namespace OpenGL {
16class ProgramManager { 14class ProgramManager {
17 static constexpr size_t NUM_STAGES = 5; 15 static constexpr size_t NUM_STAGES = 5;
18 16
19 static constexpr std::array ASSEMBLY_PROGRAM_ENUMS{
20 GL_VERTEX_PROGRAM_NV, GL_TESS_CONTROL_PROGRAM_NV, GL_TESS_EVALUATION_PROGRAM_NV,
21 GL_GEOMETRY_PROGRAM_NV, GL_FRAGMENT_PROGRAM_NV,
22 };
23
24public: 17public:
25 explicit ProgramManager(const Device& device) { 18 explicit ProgramManager(const Device& device);
26 glCreateProgramPipelines(1, &pipeline.handle); 19
27 if (device.UseAssemblyShaders()) { 20 void BindComputeProgram(GLuint program);
28 glEnable(GL_COMPUTE_PROGRAM_NV); 21
29 } 22 void BindComputeAssemblyProgram(GLuint program);
30 } 23
31 24 void BindSourcePrograms(std::span<const OGLProgram, NUM_STAGES> programs);
32 void BindComputeProgram(GLuint program) { 25
33 glUseProgram(program); 26 void BindPresentPrograms(GLuint vertex, GLuint fragment);
34 is_compute_bound = true;
35 }
36
37 void BindComputeAssemblyProgram(GLuint program) {
38 if (current_assembly_compute_program != program) {
39 current_assembly_compute_program = program;
40 glBindProgramARB(GL_COMPUTE_PROGRAM_NV, program);
41 }
42 UnbindPipeline();
43 }
44
45 void BindSourcePrograms(std::span<const OGLProgram, NUM_STAGES> programs) {
46 static constexpr std::array<GLenum, 5> stage_enums{
47 GL_VERTEX_SHADER_BIT, GL_TESS_CONTROL_SHADER_BIT, GL_TESS_EVALUATION_SHADER_BIT,
48 GL_GEOMETRY_SHADER_BIT, GL_FRAGMENT_SHADER_BIT,
49 };
50 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
51 if (current_programs[stage] != programs[stage].handle) {
52 current_programs[stage] = programs[stage].handle;
53 glUseProgramStages(pipeline.handle, stage_enums[stage], programs[stage].handle);
54 }
55 }
56 BindPipeline();
57 }
58
59 void BindPresentPrograms(GLuint vertex, GLuint fragment) {
60 if (current_programs[0] != vertex) {
61 current_programs[0] = vertex;
62 glUseProgramStages(pipeline.handle, GL_VERTEX_SHADER_BIT, vertex);
63 }
64 if (current_programs[4] != fragment) {
65 current_programs[4] = fragment;
66 glUseProgramStages(pipeline.handle, GL_FRAGMENT_SHADER_BIT, fragment);
67 }
68 glUseProgramStages(
69 pipeline.handle,
70 GL_TESS_CONTROL_SHADER_BIT | GL_TESS_EVALUATION_SHADER_BIT | GL_GEOMETRY_SHADER_BIT, 0);
71 current_programs[1] = 0;
72 current_programs[2] = 0;
73 current_programs[3] = 0;
74
75 if (current_stage_mask != 0) {
76 current_stage_mask = 0;
77 for (const GLenum program_type : ASSEMBLY_PROGRAM_ENUMS) {
78 glDisable(program_type);
79 }
80 }
81 BindPipeline();
82 }
83 27
84 void BindAssemblyPrograms(std::span<const OGLAssemblyProgram, NUM_STAGES> programs, 28 void BindAssemblyPrograms(std::span<const OGLAssemblyProgram, NUM_STAGES> programs,
85 u32 stage_mask) { 29 u32 stage_mask);
86 const u32 changed_mask = current_stage_mask ^ stage_mask; 30
87 current_stage_mask = stage_mask; 31 void RestoreGuestCompute();
88
89 if (changed_mask != 0) {
90 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
91 if (((changed_mask >> stage) & 1) != 0) {
92 if (((stage_mask >> stage) & 1) != 0) {
93 glEnable(ASSEMBLY_PROGRAM_ENUMS[stage]);
94 } else {
95 glDisable(ASSEMBLY_PROGRAM_ENUMS[stage]);
96 }
97 }
98 }
99 }
100 for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
101 if (current_programs[stage] != programs[stage].handle) {
102 current_programs[stage] = programs[stage].handle;
103 glBindProgramARB(ASSEMBLY_PROGRAM_ENUMS[stage], programs[stage].handle);
104 }
105 }
106 UnbindPipeline();
107 }
108
109 void RestoreGuestCompute() {}
110 32
111private: 33private:
112 void BindPipeline() { 34 void BindPipeline();
113 if (!is_pipeline_bound) { 35
114 is_pipeline_bound = true; 36 void UnbindPipeline();
115 glBindProgramPipeline(pipeline.handle); 37
116 } 38 void UnbindCompute();
117 UnbindCompute();
118 }
119
120 void UnbindPipeline() {
121 if (is_pipeline_bound) {
122 is_pipeline_bound = false;
123 glBindProgramPipeline(0);
124 }
125 UnbindCompute();
126 }
127
128 void UnbindCompute() {
129 if (is_compute_bound) {
130 is_compute_bound = false;
131 glUseProgram(0);
132 }
133 }
134 39
135 OGLPipeline pipeline; 40 OGLPipeline pipeline;
136 bool is_pipeline_bound{}; 41 bool is_pipeline_bound{};
diff --git a/src/video_core/renderer_vulkan/blit_image.cpp b/src/video_core/renderer_vulkan/blit_image.cpp
index 3f2b139e0..dd00d3edf 100644
--- a/src/video_core/renderer_vulkan/blit_image.cpp
+++ b/src/video_core/renderer_vulkan/blit_image.cpp
@@ -4,13 +4,13 @@
4#include <algorithm> 4#include <algorithm>
5 5
6#include "common/settings.h" 6#include "common/settings.h"
7#include "video_core/host_shaders/blit_color_float_frag_spv.h"
7#include "video_core/host_shaders/convert_abgr8_to_d24s8_frag_spv.h" 8#include "video_core/host_shaders/convert_abgr8_to_d24s8_frag_spv.h"
8#include "video_core/host_shaders/convert_d24s8_to_abgr8_frag_spv.h" 9#include "video_core/host_shaders/convert_d24s8_to_abgr8_frag_spv.h"
9#include "video_core/host_shaders/convert_depth_to_float_frag_spv.h" 10#include "video_core/host_shaders/convert_depth_to_float_frag_spv.h"
10#include "video_core/host_shaders/convert_float_to_depth_frag_spv.h" 11#include "video_core/host_shaders/convert_float_to_depth_frag_spv.h"
11#include "video_core/host_shaders/convert_s8d24_to_abgr8_frag_spv.h" 12#include "video_core/host_shaders/convert_s8d24_to_abgr8_frag_spv.h"
12#include "video_core/host_shaders/full_screen_triangle_vert_spv.h" 13#include "video_core/host_shaders/full_screen_triangle_vert_spv.h"
13#include "video_core/host_shaders/vulkan_blit_color_float_frag_spv.h"
14#include "video_core/host_shaders/vulkan_blit_depth_stencil_frag_spv.h" 14#include "video_core/host_shaders/vulkan_blit_depth_stencil_frag_spv.h"
15#include "video_core/renderer_vulkan/blit_image.h" 15#include "video_core/renderer_vulkan/blit_image.h"
16#include "video_core/renderer_vulkan/maxwell_to_vk.h" 16#include "video_core/renderer_vulkan/maxwell_to_vk.h"
@@ -303,7 +303,7 @@ void UpdateTwoTexturesDescriptorSet(const Device& device, VkDescriptorSet descri
303} 303}
304 304
305void BindBlitState(vk::CommandBuffer cmdbuf, VkPipelineLayout layout, const Region2D& dst_region, 305void BindBlitState(vk::CommandBuffer cmdbuf, VkPipelineLayout layout, const Region2D& dst_region,
306 const Region2D& src_region) { 306 const Region2D& src_region, const Extent3D& src_size = {1, 1, 1}) {
307 const VkOffset2D offset{ 307 const VkOffset2D offset{
308 .x = std::min(dst_region.start.x, dst_region.end.x), 308 .x = std::min(dst_region.start.x, dst_region.end.x),
309 .y = std::min(dst_region.start.y, dst_region.end.y), 309 .y = std::min(dst_region.start.y, dst_region.end.y),
@@ -325,12 +325,15 @@ void BindBlitState(vk::CommandBuffer cmdbuf, VkPipelineLayout layout, const Regi
325 .offset = offset, 325 .offset = offset,
326 .extent = extent, 326 .extent = extent,
327 }; 327 };
328 const float scale_x = static_cast<float>(src_region.end.x - src_region.start.x); 328 const float scale_x = static_cast<float>(src_region.end.x - src_region.start.x) /
329 const float scale_y = static_cast<float>(src_region.end.y - src_region.start.y); 329 static_cast<float>(src_size.width);
330 const float scale_y = static_cast<float>(src_region.end.y - src_region.start.y) /
331 static_cast<float>(src_size.height);
330 const PushConstants push_constants{ 332 const PushConstants push_constants{
331 .tex_scale = {scale_x, scale_y}, 333 .tex_scale = {scale_x, scale_y},
332 .tex_offset = {static_cast<float>(src_region.start.x), 334 .tex_offset = {static_cast<float>(src_region.start.x) / static_cast<float>(src_size.width),
333 static_cast<float>(src_region.start.y)}, 335 static_cast<float>(src_region.start.y) /
336 static_cast<float>(src_size.height)},
334 }; 337 };
335 cmdbuf.SetViewport(0, viewport); 338 cmdbuf.SetViewport(0, viewport);
336 cmdbuf.SetScissor(0, scissor); 339 cmdbuf.SetScissor(0, scissor);
@@ -347,6 +350,51 @@ VkExtent2D GetConversionExtent(const ImageView& src_image_view) {
347 .height = is_rescaled ? resolution.ScaleUp(height) : height, 350 .height = is_rescaled ? resolution.ScaleUp(height) : height,
348 }; 351 };
349} 352}
353
354void TransitionImageLayout(vk::CommandBuffer& cmdbuf, VkImage image, VkImageLayout target_layout,
355 VkImageLayout source_layout = VK_IMAGE_LAYOUT_GENERAL) {
356 constexpr VkFlags flags{VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
357 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT};
358 const VkImageMemoryBarrier barrier{
359 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
360 .pNext = nullptr,
361 .srcAccessMask = flags,
362 .dstAccessMask = flags,
363 .oldLayout = source_layout,
364 .newLayout = target_layout,
365 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
366 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
367 .image = image,
368 .subresourceRange{
369 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
370 .baseMipLevel = 0,
371 .levelCount = 1,
372 .baseArrayLayer = 0,
373 .layerCount = 1,
374 },
375 };
376 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
377 0, barrier);
378}
379
380void BeginRenderPass(vk::CommandBuffer& cmdbuf, const Framebuffer* framebuffer) {
381 const VkRenderPass render_pass = framebuffer->RenderPass();
382 const VkFramebuffer framebuffer_handle = framebuffer->Handle();
383 const VkExtent2D render_area = framebuffer->RenderArea();
384 const VkRenderPassBeginInfo renderpass_bi{
385 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
386 .pNext = nullptr,
387 .renderPass = render_pass,
388 .framebuffer = framebuffer_handle,
389 .renderArea{
390 .offset{},
391 .extent = render_area,
392 },
393 .clearValueCount = 0,
394 .pClearValues = nullptr,
395 };
396 cmdbuf.BeginRenderPass(renderpass_bi, VK_SUBPASS_CONTENTS_INLINE);
397}
350} // Anonymous namespace 398} // Anonymous namespace
351 399
352BlitImageHelper::BlitImageHelper(const Device& device_, Scheduler& scheduler_, 400BlitImageHelper::BlitImageHelper(const Device& device_, Scheduler& scheduler_,
@@ -365,7 +413,7 @@ BlitImageHelper::BlitImageHelper(const Device& device_, Scheduler& scheduler_,
365 two_textures_pipeline_layout(device.GetLogical().CreatePipelineLayout( 413 two_textures_pipeline_layout(device.GetLogical().CreatePipelineLayout(
366 PipelineLayoutCreateInfo(two_textures_set_layout.address()))), 414 PipelineLayoutCreateInfo(two_textures_set_layout.address()))),
367 full_screen_vert(BuildShader(device, FULL_SCREEN_TRIANGLE_VERT_SPV)), 415 full_screen_vert(BuildShader(device, FULL_SCREEN_TRIANGLE_VERT_SPV)),
368 blit_color_to_color_frag(BuildShader(device, VULKAN_BLIT_COLOR_FLOAT_FRAG_SPV)), 416 blit_color_to_color_frag(BuildShader(device, BLIT_COLOR_FLOAT_FRAG_SPV)),
369 blit_depth_stencil_frag(BuildShader(device, VULKAN_BLIT_DEPTH_STENCIL_FRAG_SPV)), 417 blit_depth_stencil_frag(BuildShader(device, VULKAN_BLIT_DEPTH_STENCIL_FRAG_SPV)),
370 convert_depth_to_float_frag(BuildShader(device, CONVERT_DEPTH_TO_FLOAT_FRAG_SPV)), 418 convert_depth_to_float_frag(BuildShader(device, CONVERT_DEPTH_TO_FLOAT_FRAG_SPV)),
371 convert_float_to_depth_frag(BuildShader(device, CONVERT_FLOAT_TO_DEPTH_FRAG_SPV)), 419 convert_float_to_depth_frag(BuildShader(device, CONVERT_FLOAT_TO_DEPTH_FRAG_SPV)),
@@ -404,6 +452,32 @@ void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, VkImageView
404 scheduler.InvalidateState(); 452 scheduler.InvalidateState();
405} 453}
406 454
455void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, VkImageView src_image_view,
456 VkImage src_image, VkSampler src_sampler,
457 const Region2D& dst_region, const Region2D& src_region,
458 const Extent3D& src_size) {
459 const BlitImagePipelineKey key{
460 .renderpass = dst_framebuffer->RenderPass(),
461 .operation = Tegra::Engines::Fermi2D::Operation::SrcCopy,
462 };
463 const VkPipelineLayout layout = *one_texture_pipeline_layout;
464 const VkPipeline pipeline = FindOrEmplaceColorPipeline(key);
465 scheduler.RequestOutsideRenderPassOperationContext();
466 scheduler.Record([this, dst_framebuffer, src_image_view, src_image, src_sampler, dst_region,
467 src_region, src_size, pipeline, layout](vk::CommandBuffer cmdbuf) {
468 TransitionImageLayout(cmdbuf, src_image, VK_IMAGE_LAYOUT_READ_ONLY_OPTIMAL);
469 BeginRenderPass(cmdbuf, dst_framebuffer);
470 const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
471 UpdateOneTextureDescriptorSet(device, descriptor_set, src_sampler, src_image_view);
472 cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
473 cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, layout, 0, descriptor_set,
474 nullptr);
475 BindBlitState(cmdbuf, layout, dst_region, src_region, src_size);
476 cmdbuf.Draw(3, 1, 0, 0);
477 cmdbuf.EndRenderPass();
478 });
479}
480
407void BlitImageHelper::BlitDepthStencil(const Framebuffer* dst_framebuffer, 481void BlitImageHelper::BlitDepthStencil(const Framebuffer* dst_framebuffer,
408 VkImageView src_depth_view, VkImageView src_stencil_view, 482 VkImageView src_depth_view, VkImageView src_stencil_view,
409 const Region2D& dst_region, const Region2D& src_region, 483 const Region2D& dst_region, const Region2D& src_region,
diff --git a/src/video_core/renderer_vulkan/blit_image.h b/src/video_core/renderer_vulkan/blit_image.h
index 5df679fb4..be8a9a2f6 100644
--- a/src/video_core/renderer_vulkan/blit_image.h
+++ b/src/video_core/renderer_vulkan/blit_image.h
@@ -10,6 +10,8 @@
10 10
11namespace Vulkan { 11namespace Vulkan {
12 12
13using VideoCommon::Extent3D;
14using VideoCommon::Offset2D;
13using VideoCommon::Region2D; 15using VideoCommon::Region2D;
14 16
15class Device; 17class Device;
@@ -36,6 +38,10 @@ public:
36 Tegra::Engines::Fermi2D::Filter filter, 38 Tegra::Engines::Fermi2D::Filter filter,
37 Tegra::Engines::Fermi2D::Operation operation); 39 Tegra::Engines::Fermi2D::Operation operation);
38 40
41 void BlitColor(const Framebuffer* dst_framebuffer, VkImageView src_image_view,
42 VkImage src_image, VkSampler src_sampler, const Region2D& dst_region,
43 const Region2D& src_region, const Extent3D& src_size);
44
39 void BlitDepthStencil(const Framebuffer* dst_framebuffer, VkImageView src_depth_view, 45 void BlitDepthStencil(const Framebuffer* dst_framebuffer, VkImageView src_depth_view,
40 VkImageView src_stencil_view, const Region2D& dst_region, 46 VkImageView src_stencil_view, const Region2D& dst_region,
41 const Region2D& src_region, Tegra::Engines::Fermi2D::Filter filter, 47 const Region2D& src_region, Tegra::Engines::Fermi2D::Filter filter,
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index b75b8eec6..86ef0daeb 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -266,6 +266,35 @@ void RasterizerVulkan::DrawIndirect() {
266 buffer_cache.SetDrawIndirect(nullptr); 266 buffer_cache.SetDrawIndirect(nullptr);
267} 267}
268 268
269void RasterizerVulkan::DrawTexture() {
270 MICROPROFILE_SCOPE(Vulkan_Drawing);
271
272 SCOPE_EXIT({ gpu.TickWork(); });
273 FlushWork();
274
275 query_cache.UpdateCounters();
276
277 texture_cache.SynchronizeGraphicsDescriptors();
278 texture_cache.UpdateRenderTargets(false);
279
280 UpdateDynamicStates();
281
282 const auto& draw_texture_state = maxwell3d->draw_manager->GetDrawTextureState();
283 const auto& sampler = texture_cache.GetGraphicsSampler(draw_texture_state.src_sampler);
284 const auto& texture = texture_cache.GetImageView(draw_texture_state.src_texture);
285 Region2D dst_region = {Offset2D{.x = static_cast<s32>(draw_texture_state.dst_x0),
286 .y = static_cast<s32>(draw_texture_state.dst_y0)},
287 Offset2D{.x = static_cast<s32>(draw_texture_state.dst_x1),
288 .y = static_cast<s32>(draw_texture_state.dst_y1)}};
289 Region2D src_region = {Offset2D{.x = static_cast<s32>(draw_texture_state.src_x0),
290 .y = static_cast<s32>(draw_texture_state.src_y0)},
291 Offset2D{.x = static_cast<s32>(draw_texture_state.src_x1),
292 .y = static_cast<s32>(draw_texture_state.src_y1)}};
293 blit_image.BlitColor(texture_cache.GetFramebuffer(), texture.RenderTarget(),
294 texture.ImageHandle(), sampler->Handle(), dst_region, src_region,
295 texture.size);
296}
297
269void RasterizerVulkan::Clear(u32 layer_count) { 298void RasterizerVulkan::Clear(u32 layer_count) {
270 MICROPROFILE_SCOPE(Vulkan_Clearing); 299 MICROPROFILE_SCOPE(Vulkan_Clearing);
271 300
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 472cc64d9..a0508b57c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -66,6 +66,7 @@ public:
66 66
67 void Draw(bool is_indexed, u32 instance_count) override; 67 void Draw(bool is_indexed, u32 instance_count) override;
68 void DrawIndirect() override; 68 void DrawIndirect() override;
69 void DrawTexture() override;
69 void Clear(u32 layer_count) override; 70 void Clear(u32 layer_count) override;
70 void DispatchCompute() override; 71 void DispatchCompute() override;
71 void ResetCounter(VideoCore::QueryType type) override; 72 void ResetCounter(VideoCore::QueryType type) override;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 87152c8e9..1b01990a4 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -149,6 +149,13 @@ typename P::ImageView& TextureCache<P>::GetImageView(ImageViewId id) noexcept {
149} 149}
150 150
151template <class P> 151template <class P>
152typename P::ImageView& TextureCache<P>::GetImageView(u32 index) noexcept {
153 const auto image_view_id = VisitImageView(channel_state->graphics_image_table,
154 channel_state->graphics_image_view_ids, index);
155 return slot_image_views[image_view_id];
156}
157
158template <class P>
152void TextureCache<P>::MarkModification(ImageId id) noexcept { 159void TextureCache<P>::MarkModification(ImageId id) noexcept {
153 MarkModification(slot_images[id]); 160 MarkModification(slot_images[id]);
154} 161}
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 4eea1f609..485eaabaa 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -129,6 +129,9 @@ public:
129 /// Return a reference to the given image view id 129 /// Return a reference to the given image view id
130 [[nodiscard]] ImageView& GetImageView(ImageViewId id) noexcept; 130 [[nodiscard]] ImageView& GetImageView(ImageViewId id) noexcept;
131 131
132 /// Get the imageview from the graphics descriptor table in the specified index
133 [[nodiscard]] ImageView& GetImageView(u32 index) noexcept;
134
132 /// Mark an image as modified from the GPU 135 /// Mark an image as modified from the GPU
133 void MarkModification(ImageId id) noexcept; 136 void MarkModification(ImageId id) noexcept;
134 137