diff options
Diffstat (limited to 'drivers')
31 files changed, 2158 insertions, 657 deletions
diff --git a/drivers/alsamidi/midi_driver_alsamidi.cpp b/drivers/alsamidi/midi_driver_alsamidi.cpp index d2a0076023..61fecccb6b 100644 --- a/drivers/alsamidi/midi_driver_alsamidi.cpp +++ b/drivers/alsamidi/midi_driver_alsamidi.cpp @@ -37,85 +37,135 @@ #include <errno.h> -static int get_message_size(uint8_t message) { - switch (message & 0xF0) { - case 0x80: // note off - case 0x90: // note on - case 0xA0: // aftertouch - case 0xB0: // continuous controller - case 0xE0: // pitch bend - case 0xF2: // song position pointer - return 3; - - case 0xC0: // patch change - case 0xD0: // channel pressure - case 0xF1: // time code quarter frame - case 0xF3: // song select +MIDIDriverALSAMidi::MessageCategory MIDIDriverALSAMidi::msg_category(uint8_t msg_part) { + if (msg_part >= 0xf8) { + return MessageCategory::RealTime; + } else if (msg_part >= 0xf0) { + // System Exclusive begin/end are specified as System Common Category messages, + // but we separate them here and give them their own categories as their + // behavior is significantly different. + if (msg_part == 0xf0) { + return MessageCategory::SysExBegin; + } else if (msg_part == 0xf7) { + return MessageCategory::SysExEnd; + } + return MessageCategory::SystemCommon; + } else if (msg_part >= 0x80) { + return MessageCategory::Voice; + } + return MessageCategory::Data; +} + +size_t MIDIDriverALSAMidi::msg_expected_data(uint8_t status_byte) { + if (msg_category(status_byte) == MessageCategory::Voice) { + // Voice messages have a channel number in the status byte, mask it out. + status_byte &= 0xf0; + } + + switch (status_byte) { + case 0x80: // Note Off + case 0x90: // Note On + case 0xA0: // Polyphonic Key Pressure (Aftertouch) + case 0xB0: // Control Change (CC) + case 0xE0: // Pitch Bend Change + case 0xF2: // Song Position Pointer return 2; - case 0xF0: // SysEx start - case 0xF4: // reserved - case 0xF5: // reserved - case 0xF6: // tune request - case 0xF7: // SysEx end - case 0xF8: // timing clock - case 0xF9: // reserved - case 0xFA: // start - case 0xFB: // continue - case 0xFC: // stop - case 0xFD: // reserved - case 0xFE: // active sensing - case 0xFF: // reset + case 0xC0: // Program Change + case 0xD0: // Channel Pressure (Aftertouch) + case 0xF1: // MIDI Time Code Quarter Frame + case 0xF3: // Song Select return 1; } - return 256; + return 0; +} + +void MIDIDriverALSAMidi::InputConnection::parse_byte(uint8_t byte, MIDIDriverALSAMidi &driver, + uint64_t timestamp) { + switch (msg_category(byte)) { + case MessageCategory::RealTime: + // Real-Time messages are single byte messages that can + // occur at any point. + // We pass them straight through. + driver.receive_input_packet(timestamp, &byte, 1); + break; + + case MessageCategory::Data: + // We don't currently forward System Exclusive messages so skip their data. + // Collect any expected data for other message types. + if (!skipping_sys_ex && expected_data > received_data) { + buffer[received_data + 1] = byte; + received_data++; + + // Forward a complete message and reset relevant state. + if (received_data == expected_data) { + driver.receive_input_packet(timestamp, buffer, received_data + 1); + received_data = 0; + + if (msg_category(buffer[0]) != MessageCategory::Voice) { + // Voice Category messages can be sent with "running status". + // This means they don't resend the status byte until it changes. + // For other categories, we reset expected data, to require a new status byte. + expected_data = 0; + } + } + } + break; + + case MessageCategory::SysExBegin: + buffer[0] = byte; + skipping_sys_ex = true; + break; + + case MessageCategory::SysExEnd: + expected_data = 0; + skipping_sys_ex = false; + break; + + case MessageCategory::Voice: + case MessageCategory::SystemCommon: + buffer[0] = byte; + received_data = 0; + expected_data = msg_expected_data(byte); + skipping_sys_ex = false; + if (expected_data == 0) { + driver.receive_input_packet(timestamp, &byte, 1); + } + break; + } +} + +int MIDIDriverALSAMidi::InputConnection::read_in(MIDIDriverALSAMidi &driver, uint64_t timestamp) { + int ret; + do { + uint8_t byte = 0; + ret = snd_rawmidi_read(rawmidi_ptr, &byte, 1); + + if (ret < 0) { + if (ret != -EAGAIN) { + ERR_PRINT("snd_rawmidi_read error: " + String(snd_strerror(ret))); + } + } else { + parse_byte(byte, driver, timestamp); + } + } while (ret > 0); + + return ret; } void MIDIDriverALSAMidi::thread_func(void *p_udata) { MIDIDriverALSAMidi *md = static_cast<MIDIDriverALSAMidi *>(p_udata); uint64_t timestamp = 0; - uint8_t buffer[256]; - int expected_size = 255; - int bytes = 0; while (!md->exit_thread.is_set()) { - int ret; - md->lock(); - for (int i = 0; i < md->connected_inputs.size(); i++) { - snd_rawmidi_t *midi_in = md->connected_inputs[i]; - do { - uint8_t byte = 0; - ret = snd_rawmidi_read(midi_in, &byte, 1); - if (ret < 0) { - if (ret != -EAGAIN) { - ERR_PRINT("snd_rawmidi_read error: " + String(snd_strerror(ret))); - } - } else { - if (byte & 0x80) { - // Flush previous packet if there is any - if (bytes) { - md->receive_input_packet(timestamp, buffer, bytes); - bytes = 0; - } - expected_size = get_message_size(byte); - // After a SysEx start, all bytes are data until a SysEx end, so - // we're going to end the command at the SES, and let the common - // driver ignore the following data bytes. - } + InputConnection *connections = md->connected_inputs.ptrw(); + size_t connection_count = md->connected_inputs.size(); - if (bytes < 256) { - buffer[bytes++] = byte; - // If we know the size of the current packet receive it if it reached the expected size - if (bytes >= expected_size) { - md->receive_input_packet(timestamp, buffer, bytes); - bytes = 0; - } - } - } - } while (ret > 0); + for (size_t i = 0; i < connection_count; i++) { + connections[i].read_in(*md, timestamp); } md->unlock(); @@ -139,7 +189,7 @@ Error MIDIDriverALSAMidi::open() { snd_rawmidi_t *midi_in; int ret = snd_rawmidi_open(&midi_in, nullptr, name, SND_RAWMIDI_NONBLOCK); if (ret >= 0) { - connected_inputs.insert(i++, midi_in); + connected_inputs.insert(i++, InputConnection(midi_in)); } } @@ -160,7 +210,7 @@ void MIDIDriverALSAMidi::close() { thread.wait_to_finish(); for (int i = 0; i < connected_inputs.size(); i++) { - snd_rawmidi_t *midi_in = connected_inputs[i]; + snd_rawmidi_t *midi_in = connected_inputs[i].rawmidi_ptr; snd_rawmidi_close(midi_in); } connected_inputs.clear(); @@ -179,7 +229,7 @@ PackedStringArray MIDIDriverALSAMidi::get_connected_inputs() { lock(); for (int i = 0; i < connected_inputs.size(); i++) { - snd_rawmidi_t *midi_in = connected_inputs[i]; + snd_rawmidi_t *midi_in = connected_inputs[i].rawmidi_ptr; snd_rawmidi_info_t *info; snd_rawmidi_info_malloc(&info); diff --git a/drivers/alsamidi/midi_driver_alsamidi.h b/drivers/alsamidi/midi_driver_alsamidi.h index ac3530b1b2..9265dede3d 100644 --- a/drivers/alsamidi/midi_driver_alsamidi.h +++ b/drivers/alsamidi/midi_driver_alsamidi.h @@ -46,12 +46,48 @@ class MIDIDriverALSAMidi : public MIDIDriver { Thread thread; Mutex mutex; - Vector<snd_rawmidi_t *> connected_inputs; + class InputConnection { + public: + InputConnection() = default; + InputConnection(snd_rawmidi_t *midi_in) : + rawmidi_ptr{ midi_in } {} + + // Read in and parse available data, forwarding any complete messages through the driver. + int read_in(MIDIDriverALSAMidi &driver, uint64_t timestamp); + + snd_rawmidi_t *rawmidi_ptr = nullptr; + + private: + static const size_t MSG_BUFFER_SIZE = 3; + uint8_t buffer[MSG_BUFFER_SIZE] = { 0 }; + size_t expected_data = 0; + size_t received_data = 0; + bool skipping_sys_ex = false; + void parse_byte(uint8_t byte, MIDIDriverALSAMidi &driver, uint64_t timestamp); + }; + + Vector<InputConnection> connected_inputs; SafeFlag exit_thread; static void thread_func(void *p_udata); + enum class MessageCategory { + Data, + Voice, + SysExBegin, + SystemCommon, // excluding System Exclusive Begin/End + SysExEnd, + RealTime, + }; + + // If the passed byte is a status byte, return the associated message category, + // else return MessageCategory::Data. + static MessageCategory msg_category(uint8_t msg_part); + + // Return the number of data bytes expected for the provided status byte. + static size_t msg_expected_data(uint8_t status_byte); + void lock() const; void unlock() const; diff --git a/drivers/gles3/effects/copy_effects.cpp b/drivers/gles3/effects/copy_effects.cpp index 92c29a4264..3acbcf6b53 100644 --- a/drivers/gles3/effects/copy_effects.cpp +++ b/drivers/gles3/effects/copy_effects.cpp @@ -117,16 +117,12 @@ CopyEffects::~CopyEffects() { void CopyEffects::copy_to_rect(const Rect2 &p_rect) { copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION); copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION); - glBindVertexArray(quad_array); - glDrawArrays(GL_TRIANGLES, 0, 6); - glBindVertexArray(0); + draw_screen_quad(); } void CopyEffects::copy_screen() { copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_DEFAULT); - glBindVertexArray(screen_triangle_array); - glDrawArrays(GL_TRIANGLES, 0, 3); - glBindVertexArray(0); + draw_screen_triangle(); } void CopyEffects::bilinear_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region) { @@ -158,8 +154,19 @@ void CopyEffects::set_color(const Color &p_color, const Rect2i &p_region) { copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_SIMPLE_COLOR); copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_region.position.x, p_region.position.y, p_region.size.x, p_region.size.y, copy.shader_version, CopyShaderGLES3::MODE_SIMPLE_COLOR); copy.shader.version_set_uniform(CopyShaderGLES3::COLOR_IN, p_color, copy.shader_version, CopyShaderGLES3::MODE_SIMPLE_COLOR); + draw_screen_quad(); +} + +void CopyEffects::draw_screen_triangle() { + glBindVertexArray(screen_triangle_array); + glDrawArrays(GL_TRIANGLES, 0, 3); + glBindVertexArray(0); +} + +void CopyEffects::draw_screen_quad() { glBindVertexArray(quad_array); glDrawArrays(GL_TRIANGLES, 0, 6); glBindVertexArray(0); } + #endif // GLES3_ENABLED diff --git a/drivers/gles3/effects/copy_effects.h b/drivers/gles3/effects/copy_effects.h index 7f16b4e0f3..a817c3f0dc 100644 --- a/drivers/gles3/effects/copy_effects.h +++ b/drivers/gles3/effects/copy_effects.h @@ -65,6 +65,8 @@ public: void copy_screen(); void bilinear_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region); void set_color(const Color &p_color, const Rect2i &p_region); + void draw_screen_triangle(); + void draw_screen_quad(); }; } //namespace GLES3 diff --git a/drivers/gles3/rasterizer_canvas_gles3.cpp b/drivers/gles3/rasterizer_canvas_gles3.cpp index 5d0edda6c5..0a64cda787 100644 --- a/drivers/gles3/rasterizer_canvas_gles3.cpp +++ b/drivers/gles3/rasterizer_canvas_gles3.cpp @@ -36,26 +36,13 @@ #include "rasterizer_scene_gles3.h" #include "core/config/project_settings.h" +#include "core/math/geometry_2d.h" #include "servers/rendering/rendering_server_default.h" #include "storage/config.h" #include "storage/material_storage.h" #include "storage/mesh_storage.h" #include "storage/texture_storage.h" -#ifndef GLES_OVER_GL -#define glClearDepth glClearDepthf -#endif - -//static const GLenum gl_primitive[] = { -// GL_POINTS, -// GL_LINES, -// GL_LINE_STRIP, -// GL_LINE_LOOP, -// GL_TRIANGLES, -// GL_TRIANGLE_STRIP, -// GL_TRIANGLE_FAN -//}; - void RasterizerCanvasGLES3::_update_transform_2d_to_mat4(const Transform2D &p_transform, float *p_mat4) { p_mat4[0] = p_transform.columns[0][0]; p_mat4[1] = p_transform.columns[0][1]; @@ -174,7 +161,7 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state.light_uniforms[index].position[0] = -canvas_light_dir.x; state.light_uniforms[index].position[1] = -canvas_light_dir.y; - //_update_transform_2d_to_mat2x4(clight->shadow.directional_xform, state.light_uniforms[index].shadow_matrix); + _update_transform_2d_to_mat2x4(clight->shadow.directional_xform, state.light_uniforms[index].shadow_matrix); state.light_uniforms[index].height = l->height; //0..1 here @@ -185,8 +172,7 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state.light_uniforms[index].color[3] = l->energy; //use alpha for energy, so base color can go separate - /* - if (state.shadow_fb.is_valid()) { + if (state.shadow_fb != 0) { state.light_uniforms[index].shadow_pixel_size = (1.0 / state.shadow_texture_size) * (1.0 + l->shadow_smooth); state.light_uniforms[index].shadow_z_far_inv = 1.0 / clight->shadow.z_far; state.light_uniforms[index].shadow_y_ofs = clight->shadow.y_offset; @@ -195,15 +181,13 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state.light_uniforms[index].shadow_z_far_inv = 1.0; state.light_uniforms[index].shadow_y_ofs = 0; } - */ state.light_uniforms[index].flags = l->blend_mode << LIGHT_FLAGS_BLEND_SHIFT; state.light_uniforms[index].flags |= l->shadow_filter << LIGHT_FLAGS_FILTER_SHIFT; - /* + if (clight->shadow.enabled) { state.light_uniforms[index].flags |= LIGHT_FLAGS_HAS_SHADOW; } - */ l->render_index_cache = index; @@ -252,24 +236,22 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state.light_uniforms[index].color[3] = l->energy; //use alpha for energy, so base color can go separate - /* - if (state.shadow_fb.is_valid()) { - state.light_uniforms[index].shadow_pixel_size = (1.0 / state.shadow_texture_size) * (1.0 + l->shadow_smooth); - state.light_uniforms[index].shadow_z_far_inv = 1.0 / clight->shadow.z_far; - state.light_uniforms[index].shadow_y_ofs = clight->shadow.y_offset; - } else { - state.light_uniforms[index].shadow_pixel_size = 1.0; - state.light_uniforms[index].shadow_z_far_inv = 1.0; - state.light_uniforms[index].shadow_y_ofs = 0; - } - */ + if (state.shadow_fb != 0) { + state.light_uniforms[index].shadow_pixel_size = (1.0 / state.shadow_texture_size) * (1.0 + l->shadow_smooth); + state.light_uniforms[index].shadow_z_far_inv = 1.0 / clight->shadow.z_far; + state.light_uniforms[index].shadow_y_ofs = clight->shadow.y_offset; + } else { + state.light_uniforms[index].shadow_pixel_size = 1.0; + state.light_uniforms[index].shadow_z_far_inv = 1.0; + state.light_uniforms[index].shadow_y_ofs = 0; + } + state.light_uniforms[index].flags = l->blend_mode << LIGHT_FLAGS_BLEND_SHIFT; state.light_uniforms[index].flags |= l->shadow_filter << LIGHT_FLAGS_FILTER_SHIFT; - /* + if (clight->shadow.enabled) { state.light_uniforms[index].flags |= LIGHT_FLAGS_HAS_SHADOW; } - */ if (clight->texture.is_valid()) { Rect2 atlas_rect = GLES3::TextureStorage::get_singleton()->texture_atlas_get_texture_rect(clight->texture); @@ -313,6 +295,13 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ } glActiveTexture(GL_TEXTURE0 + GLES3::Config::get_singleton()->max_texture_image_units - 2); glBindTexture(GL_TEXTURE_2D, texture_atlas); + GLuint shadow_tex = state.shadow_texture; + if (shadow_tex == 0) { + GLES3::Texture *tex = texture_storage->get_texture(texture_storage->texture_gl_get_default(GLES3::DEFAULT_GL_TEXTURE_WHITE)); + shadow_tex = tex->tex_id; + } + glActiveTexture(GL_TEXTURE0 + GLES3::Config::get_singleton()->max_texture_image_units - 3); + glBindTexture(GL_TEXTURE_2D, shadow_tex); } { @@ -342,8 +331,6 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state_buffer.screen_pixel_size[0] = 1.0 / render_target_size.x; state_buffer.screen_pixel_size[1] = 1.0 / render_target_size.y; - glViewport(0, 0, render_target_size.x, render_target_size.y); - state_buffer.time = state.time; state_buffer.use_pixel_snap = p_snap_2d_vertices_to_pixel; @@ -366,6 +353,7 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ state_buffer.sdf_to_tex[3] = -sdf_tex_rect.position.y / sdf_tex_rect.size.height; state_buffer.tex_to_sdf = 1.0 / ((canvas_scale.x + canvas_scale.y) * 0.5); + glBindBufferBase(GL_UNIFORM_BUFFER, BASE_UNIFORM_LOCATION, state.canvas_instance_data_buffers[state.current_buffer].state_ubo); glBufferData(GL_UNIFORM_BUFFER, sizeof(StateBuffer), &state_buffer, GL_STREAM_DRAW); @@ -375,11 +363,17 @@ void RasterizerCanvasGLES3::canvas_render_items(RID p_to_render_target, Item *p_ glBindBuffer(GL_UNIFORM_BUFFER, 0); } + glActiveTexture(GL_TEXTURE0 + GLES3::Config::get_singleton()->max_texture_image_units - 5); + glBindTexture(GL_TEXTURE_2D, texture_storage->render_target_get_sdf_texture(p_to_render_target)); + { state.default_filter = p_default_filter; state.default_repeat = p_default_repeat; } + Size2 render_target_size = texture_storage->render_target_get_size(p_to_render_target); + glViewport(0, 0, render_target_size.x, render_target_size.y); + r_sdf_used = false; int item_count = 0; bool backbuffer_cleared = false; @@ -587,6 +581,13 @@ void RasterizerCanvasGLES3::_render_items(RID p_to_render_target, int p_item_cou _record_item_commands(ci, p_canvas_transform_inverse, current_clip, blend_mode, p_lights, index, batch_broken); } + if (index == 0) { + // Nothing to render, just return. + state.current_batch_index = 0; + state.canvas_instance_batches.clear(); + return; + } + // Copy over all data needed for rendering. glBindBuffer(GL_UNIFORM_BUFFER, state.canvas_instance_data_buffers[state.current_buffer].ubo); #ifdef WEB_ENABLED @@ -621,6 +622,7 @@ void RasterizerCanvasGLES3::_render_items(RID p_to_render_target, int p_item_cou CanvasShaderGLES3::ShaderVariant variant = state.canvas_instance_batches[i].shader_variant; uint64_t specialization = 0; specialization |= uint64_t(state.canvas_instance_batches[i].lights_disabled); + specialization |= uint64_t(!GLES3::Config::get_singleton()->float_texture_supported) << 1; _bind_material(material_data, variant, specialization); GLES3::CanvasShaderData::BlendMode blend_mode = state.canvas_instance_batches[i].blend_mode; @@ -1004,13 +1006,15 @@ void RasterizerCanvasGLES3::_record_item_commands(const Item *p_item, const Tran if (primitive->point_count != state.canvas_instance_batches[state.current_batch_index].primitive_points || state.canvas_instance_batches[state.current_batch_index].command_type != Item::Command::TYPE_PRIMITIVE) { _new_batch(r_batch_broken, r_index); - state.canvas_instance_batches[state.current_batch_index].tex = RID(); + state.canvas_instance_batches[state.current_batch_index].tex = primitive->texture; state.canvas_instance_batches[state.current_batch_index].primitive_points = primitive->point_count; state.canvas_instance_batches[state.current_batch_index].command_type = Item::Command::TYPE_PRIMITIVE; state.canvas_instance_batches[state.current_batch_index].command = c; state.canvas_instance_batches[state.current_batch_index].shader_variant = CanvasShaderGLES3::MODE_PRIMITIVE; } + _prepare_canvas_texture(state.canvas_instance_batches[state.current_batch_index].tex, state.canvas_instance_batches[state.current_batch_index].filter, state.canvas_instance_batches[state.current_batch_index].repeat, r_index, texpixel_size); + for (uint32_t j = 0; j < MIN(3u, primitive->point_count); j++) { state.instance_data_array[r_index].points[j * 2 + 0] = primitive->points[j].x; state.instance_data_array[r_index].points[j * 2 + 1] = primitive->points[j].y; @@ -1323,7 +1327,7 @@ void RasterizerCanvasGLES3::_render_batch(Light *p_lights, uint32_t p_index) { void RasterizerCanvasGLES3::_add_to_batch(uint32_t &r_index, bool &r_batch_broken) { if (r_index >= data.max_instances_per_ubo - 1) { - WARN_PRINT_ONCE("Trying to draw too many items. Please increase maximum number of items in the project settings 'rendering/gl_compatibility/item_buffer_size'"); + ERR_PRINT_ONCE("Trying to draw too many items. Please increase maximum number of items in the project settings 'rendering/gl_compatibility/item_buffer_size'"); return; } @@ -1395,28 +1399,516 @@ void RasterizerCanvasGLES3::light_set_texture(RID p_rid, RID p_texture) { } void RasterizerCanvasGLES3::light_set_use_shadow(RID p_rid, bool p_enable) { + CanvasLight *cl = canvas_light_owner.get_or_null(p_rid); + ERR_FAIL_COND(!cl); + + cl->shadow.enabled = p_enable; } void RasterizerCanvasGLES3::light_update_shadow(RID p_rid, int p_shadow_index, const Transform2D &p_light_xform, int p_light_mask, float p_near, float p_far, LightOccluderInstance *p_occluders) { + GLES3::Config *config = GLES3::Config::get_singleton(); + + CanvasLight *cl = canvas_light_owner.get_or_null(p_rid); + ERR_FAIL_COND(!cl->shadow.enabled); + + _update_shadow_atlas(); + + cl->shadow.z_far = p_far; + cl->shadow.y_offset = float(p_shadow_index * 2 + 1) / float(data.max_lights_per_render * 2); + + glBindFramebuffer(GL_FRAMEBUFFER, state.shadow_fb); + glViewport(0, p_shadow_index * 2, state.shadow_texture_size, 2); + + glDepthMask(GL_TRUE); + glEnable(GL_DEPTH_TEST); + glDepthFunc(GL_LESS); + glDisable(GL_BLEND); + + glEnable(GL_SCISSOR_TEST); + glScissor(0, p_shadow_index * 2, state.shadow_texture_size, 2); + glClearColor(p_far, p_far, p_far, 1.0); + glClearDepth(1.0); + glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); + + glCullFace(GL_BACK); + glDisable(GL_CULL_FACE); + RS::CanvasOccluderPolygonCullMode cull_mode = RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED; + + CanvasOcclusionShaderGLES3::ShaderVariant variant = config->float_texture_supported ? CanvasOcclusionShaderGLES3::MODE_SHADOW : CanvasOcclusionShaderGLES3::MODE_SHADOW_RGBA; + shadow_render.shader.version_bind_shader(shadow_render.shader_version, variant); + + for (int i = 0; i < 4; i++) { + glViewport((state.shadow_texture_size / 4) * i, p_shadow_index * 2, (state.shadow_texture_size / 4), 2); + + Projection projection; + { + real_t fov = 90; + real_t nearp = p_near; + real_t farp = p_far; + real_t aspect = 1.0; + + real_t ymax = nearp * Math::tan(Math::deg_to_rad(fov * 0.5)); + real_t ymin = -ymax; + real_t xmin = ymin * aspect; + real_t xmax = ymax * aspect; + + projection.set_frustum(xmin, xmax, ymin, ymax, nearp, farp); + } + + Vector3 cam_target = Basis::from_euler(Vector3(0, 0, Math_TAU * ((i + 3) / 4.0))).xform(Vector3(0, 1, 0)); + + projection = projection * Projection(Transform3D().looking_at(cam_target, Vector3(0, 0, -1)).affine_inverse()); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::PROJECTION, projection, shadow_render.shader_version, variant); + + static const Vector2 directions[4] = { Vector2(1, 0), Vector2(0, 1), Vector2(-1, 0), Vector2(0, -1) }; + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::DIRECTION, directions[i].x, directions[i].y, shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::Z_FAR, p_far, shadow_render.shader_version, variant); + + LightOccluderInstance *instance = p_occluders; + + while (instance) { + OccluderPolygon *co = occluder_polygon_owner.get_or_null(instance->occluder); + + if (!co || co->vertex_array == 0 || !(p_light_mask & instance->light_mask)) { + instance = instance->next; + continue; + } + + Transform2D modelview = p_light_xform * instance->xform_cache; + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW1, modelview.columns[0][0], modelview.columns[1][0], 0, modelview.columns[2][0], shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW2, modelview.columns[0][1], modelview.columns[1][1], 0, modelview.columns[2][1], shadow_render.shader_version, variant); + + if (co->cull_mode != cull_mode) { + if (co->cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED) { + glDisable(GL_CULL_FACE); + } else { + if (cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED) { + // Last time was disabled, so enable and set proper face. + glEnable(GL_CULL_FACE); + } + glCullFace(co->cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_CLOCKWISE ? GL_FRONT : GL_BACK); + } + cull_mode = co->cull_mode; + } + + glBindVertexArray(co->vertex_array); + glDrawElements(GL_TRIANGLES, 3 * co->line_point_count, GL_UNSIGNED_SHORT, 0); + + instance = instance->next; + } + } + + glBindVertexArray(0); + glBindFramebuffer(GL_FRAMEBUFFER, 0); + glDepthMask(GL_FALSE); + glDisable(GL_DEPTH_TEST); + glDisable(GL_SCISSOR_TEST); } void RasterizerCanvasGLES3::light_update_directional_shadow(RID p_rid, int p_shadow_index, const Transform2D &p_light_xform, int p_light_mask, float p_cull_distance, const Rect2 &p_clip_rect, LightOccluderInstance *p_occluders) { + GLES3::Config *config = GLES3::Config::get_singleton(); + + CanvasLight *cl = canvas_light_owner.get_or_null(p_rid); + ERR_FAIL_COND(!cl->shadow.enabled); + + _update_shadow_atlas(); + + Vector2 light_dir = p_light_xform.columns[1].normalized(); + + Vector2 center = p_clip_rect.get_center(); + + float to_edge_distance = ABS(light_dir.dot(p_clip_rect.get_support(light_dir)) - light_dir.dot(center)); + + Vector2 from_pos = center - light_dir * (to_edge_distance + p_cull_distance); + float distance = to_edge_distance * 2.0 + p_cull_distance; + float half_size = p_clip_rect.size.length() * 0.5; //shadow length, must keep this no matter the angle + + cl->shadow.z_far = distance; + cl->shadow.y_offset = float(p_shadow_index * 2 + 1) / float(data.max_lights_per_render * 2); + + Transform2D to_light_xform; + + to_light_xform[2] = from_pos; + to_light_xform[1] = light_dir; + to_light_xform[0] = -light_dir.orthogonal(); + + to_light_xform.invert(); + + glBindFramebuffer(GL_FRAMEBUFFER, state.shadow_fb); + glViewport(0, p_shadow_index * 2, state.shadow_texture_size, 2); + + glDepthMask(GL_TRUE); + glEnable(GL_DEPTH_TEST); + glDepthFunc(GL_LESS); + glDisable(GL_BLEND); + + glEnable(GL_SCISSOR_TEST); + glScissor(0, p_shadow_index * 2, state.shadow_texture_size, 2); + glClearColor(1.0, 1.0, 1.0, 1.0); + glClearDepth(1.0); + glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); + + glCullFace(GL_BACK); + glDisable(GL_CULL_FACE); + RS::CanvasOccluderPolygonCullMode cull_mode = RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED; + + CanvasOcclusionShaderGLES3::ShaderVariant variant = config->float_texture_supported ? CanvasOcclusionShaderGLES3::MODE_SHADOW : CanvasOcclusionShaderGLES3::MODE_SHADOW_RGBA; + shadow_render.shader.version_bind_shader(shadow_render.shader_version, variant); + + Projection projection; + projection.set_orthogonal(-half_size, half_size, -0.5, 0.5, 0.0, distance); + projection = projection * Projection(Transform3D().looking_at(Vector3(0, 1, 0), Vector3(0, 0, -1)).affine_inverse()); + + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::PROJECTION, projection, shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::DIRECTION, 0.0, 1.0, shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::Z_FAR, distance, shadow_render.shader_version, variant); + + LightOccluderInstance *instance = p_occluders; + + while (instance) { + OccluderPolygon *co = occluder_polygon_owner.get_or_null(instance->occluder); + + if (!co || co->vertex_array == 0 || !(p_light_mask & instance->light_mask)) { + instance = instance->next; + continue; + } + + Transform2D modelview = to_light_xform * instance->xform_cache; + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW1, modelview.columns[0][0], modelview.columns[1][0], 0, modelview.columns[2][0], shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW2, modelview.columns[0][1], modelview.columns[1][1], 0, modelview.columns[2][1], shadow_render.shader_version, variant); + + if (co->cull_mode != cull_mode) { + if (co->cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED) { + glDisable(GL_CULL_FACE); + } else { + if (cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED) { + // Last time was disabled, so enable and set proper face. + glEnable(GL_CULL_FACE); + } + glCullFace(co->cull_mode == RS::CANVAS_OCCLUDER_POLYGON_CULL_CLOCKWISE ? GL_FRONT : GL_BACK); + } + cull_mode = co->cull_mode; + } + + glBindVertexArray(co->vertex_array); + glDrawElements(GL_TRIANGLES, 3 * co->line_point_count, GL_UNSIGNED_SHORT, 0); + + instance = instance->next; + } + + Transform2D to_shadow; + to_shadow.columns[0].x = 1.0 / -(half_size * 2.0); + to_shadow.columns[2].x = 0.5; + + cl->shadow.directional_xform = to_shadow * to_light_xform; + + glBindVertexArray(0); + glBindFramebuffer(GL_FRAMEBUFFER, 0); + glDepthMask(GL_FALSE); + glDisable(GL_DEPTH_TEST); + glDisable(GL_SCISSOR_TEST); + glDisable(GL_CULL_FACE); +} + +void RasterizerCanvasGLES3::_update_shadow_atlas() { + GLES3::Config *config = GLES3::Config::get_singleton(); + + if (state.shadow_fb == 0) { + glActiveTexture(GL_TEXTURE0); + + glGenFramebuffers(1, &state.shadow_fb); + glBindFramebuffer(GL_FRAMEBUFFER, state.shadow_fb); + + glGenRenderbuffers(1, &state.shadow_depth_buffer); + glBindRenderbuffer(GL_RENDERBUFFER, state.shadow_depth_buffer); + glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, state.shadow_texture_size, data.max_lights_per_render * 2); + glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, state.shadow_depth_buffer); + + glGenTextures(1, &state.shadow_texture); + glBindTexture(GL_TEXTURE_2D, state.shadow_texture); + if (config->float_texture_supported) { + glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, state.shadow_texture_size, data.max_lights_per_render * 2, 0, GL_RED, GL_FLOAT, nullptr); + } else { + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, state.shadow_texture_size, data.max_lights_per_render * 2, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr); + } + + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1); + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, state.shadow_texture, 0); + + GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER); + if (status != GL_FRAMEBUFFER_COMPLETE) { + glDeleteFramebuffers(1, &state.shadow_fb); + glDeleteTextures(1, &state.shadow_texture); + glDeleteRenderbuffers(1, &state.shadow_depth_buffer); + state.shadow_fb = 0; + state.shadow_texture = 0; + state.shadow_depth_buffer = 0; + WARN_PRINT("Could not create CanvasItem shadow atlas, status: " + GLES3::TextureStorage::get_singleton()->get_framebuffer_error(status)); + } + glBindFramebuffer(GL_FRAMEBUFFER, GLES3::TextureStorage::system_fbo); + } } void RasterizerCanvasGLES3::render_sdf(RID p_render_target, LightOccluderInstance *p_occluders) { + GLES3::TextureStorage *texture_storage = GLES3::TextureStorage::get_singleton(); + + GLuint fb = texture_storage->render_target_get_sdf_framebuffer(p_render_target); + Rect2i rect = texture_storage->render_target_get_sdf_rect(p_render_target); + + Transform2D to_sdf; + to_sdf.columns[0] *= rect.size.width; + to_sdf.columns[1] *= rect.size.height; + to_sdf.columns[2] = rect.position; + + Transform2D to_clip; + to_clip.columns[0] *= 2.0; + to_clip.columns[1] *= 2.0; + to_clip.columns[2] = -Vector2(1.0, 1.0); + + to_clip = to_clip * to_sdf.affine_inverse(); + + glBindFramebuffer(GL_FRAMEBUFFER, fb); + glViewport(0, 0, rect.size.width, rect.size.height); + + glDepthMask(GL_FALSE); + glDisable(GL_DEPTH_TEST); + glDisable(GL_BLEND); + glDisable(GL_CULL_FACE); + glDisable(GL_SCISSOR_TEST); + + glClearColor(0.0, 0.0, 0.0, 0.0); + glClear(GL_COLOR_BUFFER_BIT); + + CanvasOcclusionShaderGLES3::ShaderVariant variant = CanvasOcclusionShaderGLES3::MODE_SDF; + shadow_render.shader.version_bind_shader(shadow_render.shader_version, variant); + + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::PROJECTION, Projection(), shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::DIRECTION, 0.0, 0.0, shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::Z_FAR, 0.0, shadow_render.shader_version, variant); + + LightOccluderInstance *instance = p_occluders; + + while (instance) { + OccluderPolygon *oc = occluder_polygon_owner.get_or_null(instance->occluder); + + if (!oc || oc->sdf_vertex_array == 0) { + instance = instance->next; + continue; + } + + Transform2D modelview = to_clip * instance->xform_cache; + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW1, modelview.columns[0][0], modelview.columns[1][0], 0, modelview.columns[2][0], shadow_render.shader_version, variant); + shadow_render.shader.version_set_uniform(CanvasOcclusionShaderGLES3::MODELVIEW2, modelview.columns[0][1], modelview.columns[1][1], 0, modelview.columns[2][1], shadow_render.shader_version, variant); + + glBindVertexArray(oc->sdf_vertex_array); + glDrawElements(oc->sdf_is_lines ? GL_LINES : GL_TRIANGLES, oc->sdf_index_count, GL_UNSIGNED_INT, 0); + + instance = instance->next; + } + + texture_storage->render_target_sdf_process(p_render_target); //done rendering, process it + glBindVertexArray(0); + glBindFramebuffer(GL_FRAMEBUFFER, 0); } RID RasterizerCanvasGLES3::occluder_polygon_create() { - return RID(); + OccluderPolygon occluder; + + return occluder_polygon_owner.make_rid(occluder); } void RasterizerCanvasGLES3::occluder_polygon_set_shape(RID p_occluder, const Vector<Vector2> &p_points, bool p_closed) { + OccluderPolygon *oc = occluder_polygon_owner.get_or_null(p_occluder); + ERR_FAIL_COND(!oc); + + Vector<Vector2> lines; + + if (p_points.size()) { + int lc = p_points.size() * 2; + + lines.resize(lc - (p_closed ? 0 : 2)); + { + Vector2 *w = lines.ptrw(); + const Vector2 *r = p_points.ptr(); + + int max = lc / 2; + if (!p_closed) { + max--; + } + for (int i = 0; i < max; i++) { + Vector2 a = r[i]; + Vector2 b = r[(i + 1) % (lc / 2)]; + w[i * 2 + 0] = a; + w[i * 2 + 1] = b; + } + } + } + + if (oc->line_point_count != lines.size() && oc->vertex_array != 0) { + glDeleteVertexArrays(1, &oc->vertex_array); + glDeleteBuffers(1, &oc->vertex_buffer); + glDeleteBuffers(1, &oc->index_buffer); + + oc->vertex_array = 0; + oc->vertex_buffer = 0; + oc->index_buffer = 0; + } + + if (lines.size()) { + Vector<uint8_t> geometry; + Vector<uint8_t> indices; + int lc = lines.size(); + + geometry.resize(lc * 6 * sizeof(float)); + indices.resize(lc * 3 * sizeof(uint16_t)); + + { + uint8_t *vw = geometry.ptrw(); + float *vwptr = reinterpret_cast<float *>(vw); + uint8_t *iw = indices.ptrw(); + uint16_t *iwptr = (uint16_t *)iw; + + const Vector2 *lr = lines.ptr(); + + const int POLY_HEIGHT = 16384; + + for (int i = 0; i < lc / 2; i++) { + vwptr[i * 12 + 0] = lr[i * 2 + 0].x; + vwptr[i * 12 + 1] = lr[i * 2 + 0].y; + vwptr[i * 12 + 2] = POLY_HEIGHT; + + vwptr[i * 12 + 3] = lr[i * 2 + 1].x; + vwptr[i * 12 + 4] = lr[i * 2 + 1].y; + vwptr[i * 12 + 5] = POLY_HEIGHT; + + vwptr[i * 12 + 6] = lr[i * 2 + 1].x; + vwptr[i * 12 + 7] = lr[i * 2 + 1].y; + vwptr[i * 12 + 8] = -POLY_HEIGHT; + + vwptr[i * 12 + 9] = lr[i * 2 + 0].x; + vwptr[i * 12 + 10] = lr[i * 2 + 0].y; + vwptr[i * 12 + 11] = -POLY_HEIGHT; + + iwptr[i * 6 + 0] = i * 4 + 0; + iwptr[i * 6 + 1] = i * 4 + 1; + iwptr[i * 6 + 2] = i * 4 + 2; + + iwptr[i * 6 + 3] = i * 4 + 2; + iwptr[i * 6 + 4] = i * 4 + 3; + iwptr[i * 6 + 5] = i * 4 + 0; + } + } + + if (oc->vertex_array == 0) { + oc->line_point_count = lc; + glGenVertexArrays(1, &oc->vertex_array); + glBindVertexArray(oc->vertex_array); + glGenBuffers(1, &oc->vertex_buffer); + glBindBuffer(GL_ARRAY_BUFFER, oc->vertex_buffer); + + glBufferData(GL_ARRAY_BUFFER, lc * 6 * sizeof(float), geometry.ptr(), GL_STATIC_DRAW); + glEnableVertexAttribArray(RS::ARRAY_VERTEX); + glVertexAttribPointer(RS::ARRAY_VERTEX, 3, GL_FLOAT, GL_FALSE, 3 * sizeof(float), nullptr); + + glGenBuffers(1, &oc->index_buffer); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, oc->index_buffer); + glBufferData(GL_ELEMENT_ARRAY_BUFFER, 3 * lc * sizeof(uint16_t), indices.ptr(), GL_STATIC_DRAW); + glBindVertexArray(0); + } else { + glBindVertexArray(oc->vertex_array); + glBindBuffer(GL_ARRAY_BUFFER, oc->vertex_buffer); + glBufferData(GL_ARRAY_BUFFER, lc * 6 * sizeof(float), geometry.ptr(), GL_STATIC_DRAW); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, oc->index_buffer); + glBufferData(GL_ELEMENT_ARRAY_BUFFER, 3 * lc * sizeof(uint16_t), indices.ptr(), GL_STATIC_DRAW); + } + } + + // sdf + + Vector<int> sdf_indices; + + if (p_points.size()) { + if (p_closed) { + sdf_indices = Geometry2D::triangulate_polygon(p_points); + oc->sdf_is_lines = false; + } else { + int max = p_points.size(); + sdf_indices.resize(max * 2); + + int *iw = sdf_indices.ptrw(); + for (int i = 0; i < max; i++) { + iw[i * 2 + 0] = i; + iw[i * 2 + 1] = (i + 1) % max; + } + oc->sdf_is_lines = true; + } + } + + if (oc->sdf_index_count != sdf_indices.size() && oc->sdf_point_count != p_points.size() && oc->sdf_vertex_array != 0) { + glDeleteVertexArrays(1, &oc->sdf_vertex_array); + glDeleteBuffers(1, &oc->sdf_vertex_buffer); + glDeleteBuffers(1, &oc->sdf_index_buffer); + + oc->sdf_vertex_array = 0; + oc->sdf_vertex_buffer = 0; + oc->sdf_index_buffer = 0; + + oc->sdf_index_count = sdf_indices.size(); + oc->sdf_point_count = p_points.size(); + } + + if (sdf_indices.size()) { + if (oc->sdf_vertex_array == 0) { + oc->sdf_index_count = sdf_indices.size(); + oc->sdf_point_count = p_points.size(); + glGenVertexArrays(1, &oc->sdf_vertex_array); + glBindVertexArray(oc->sdf_vertex_array); + glGenBuffers(1, &oc->sdf_vertex_buffer); + glBindBuffer(GL_ARRAY_BUFFER, oc->sdf_vertex_buffer); + + glBufferData(GL_ARRAY_BUFFER, p_points.size() * 2 * sizeof(float), p_points.to_byte_array().ptr(), GL_STATIC_DRAW); + glEnableVertexAttribArray(RS::ARRAY_VERTEX); + glVertexAttribPointer(RS::ARRAY_VERTEX, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr); + + glGenBuffers(1, &oc->sdf_index_buffer); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, oc->sdf_index_buffer); + glBufferData(GL_ELEMENT_ARRAY_BUFFER, sdf_indices.size() * sizeof(uint32_t), sdf_indices.to_byte_array().ptr(), GL_STATIC_DRAW); + glBindVertexArray(0); + } else { + glBindBuffer(GL_ARRAY_BUFFER, oc->sdf_vertex_buffer); + glBufferData(GL_ARRAY_BUFFER, p_points.size() * 2 * sizeof(float), p_points.to_byte_array().ptr(), GL_STATIC_DRAW); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, oc->sdf_index_buffer); + glBufferData(GL_ELEMENT_ARRAY_BUFFER, sdf_indices.size() * sizeof(uint32_t), sdf_indices.to_byte_array().ptr(), GL_STATIC_DRAW); + glBindBuffer(GL_ARRAY_BUFFER, 0); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0); + } + } } void RasterizerCanvasGLES3::occluder_polygon_set_cull_mode(RID p_occluder, RS::CanvasOccluderPolygonCullMode p_mode) { + OccluderPolygon *oc = occluder_polygon_owner.get_or_null(p_occluder); + ERR_FAIL_COND(!oc); + oc->cull_mode = p_mode; } void RasterizerCanvasGLES3::set_shadow_texture_size(int p_size) { + GLES3::Config *config = GLES3::Config::get_singleton(); + p_size = nearest_power_of_2_templated(p_size); + if (p_size == state.shadow_texture_size) { + return; + } + + if (p_size > config->max_texture_size) { + p_size = config->max_texture_size; + WARN_PRINT("Attempting to set CanvasItem shadow atlas size to " + itos(p_size) + " which is beyond limit of " + itos(config->max_texture_size) + "supported by hardware."); + } + + state.shadow_texture_size = p_size; } bool RasterizerCanvasGLES3::free(RID p_rid) { @@ -1424,6 +1916,9 @@ bool RasterizerCanvasGLES3::free(RID p_rid) { CanvasLight *cl = canvas_light_owner.get_or_null(p_rid); ERR_FAIL_COND_V(!cl, false); canvas_light_owner.free(p_rid); + } else if (occluder_polygon_owner.owns(p_rid)) { + occluder_polygon_set_shape(p_rid, Vector<Vector2>(), false); + occluder_polygon_owner.free(p_rid); } else { return false; } @@ -1623,16 +2118,16 @@ void RasterizerCanvasGLES3::reset_canvas() { glEnable(GL_BLEND); glBlendFuncSeparate(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ZERO, GL_ONE); + glActiveTexture(GL_TEXTURE0 + GLES3::Config::get_singleton()->max_texture_image_units - 2); + glBindTexture(GL_TEXTURE_2D, 0); + glActiveTexture(GL_TEXTURE0 + GLES3::Config::get_singleton()->max_texture_image_units - 3); + glBindTexture(GL_TEXTURE_2D, 0); + glActiveTexture(GL_TEXTURE0); + glBindBuffer(GL_ARRAY_BUFFER, 0); glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0); } -void RasterizerCanvasGLES3::canvas_debug_viewport_shadows(Light *p_lights_with_shadow) { -} - -void RasterizerCanvasGLES3::canvas_light_shadow_buffer_update(RID p_buffer, const Transform2D &p_light_xform, int p_light_mask, float p_near, float p_far, LightOccluderInstance *p_occluders, Projection *p_xform_cache) { -} - void RasterizerCanvasGLES3::draw_lens_distortion_rect(const Rect2 &p_rect, float p_k1, float p_k2, const Vector2 &p_eye_center, float p_oversample) { } @@ -1663,7 +2158,6 @@ RendererCanvasRender::PolygonID RasterizerCanvasGLES3::request_polygon(const Vec polygon_buffer.resize(buffer_size * sizeof(float)); { glBindBuffer(GL_ARRAY_BUFFER, pb.vertex_buffer); - glBufferData(GL_ARRAY_BUFFER, stride * vertex_count * sizeof(float), nullptr, GL_STATIC_DRAW); // TODO may not be necessary uint8_t *r = polygon_buffer.ptrw(); float *fptr = reinterpret_cast<float *>(r); uint32_t *uptr = (uint32_t *)r; @@ -1772,7 +2266,6 @@ RendererCanvasRender::PolygonID RasterizerCanvasGLES3::request_polygon(const Vec } glGenBuffers(1, &pb.index_buffer); glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, pb.index_buffer); - glBufferData(GL_ELEMENT_ARRAY_BUFFER, p_indices.size() * 4, nullptr, GL_STATIC_DRAW); // TODO may not be necessary glBufferData(GL_ELEMENT_ARRAY_BUFFER, p_indices.size() * 4, index_buffer.ptr(), GL_STATIC_DRAW); pb.count = p_indices.size(); } @@ -2064,6 +2557,9 @@ RasterizerCanvasGLES3::RasterizerCanvasGLES3() { data.canvas_shader_default_version = GLES3::MaterialStorage::get_singleton()->shaders.canvas_shader.version_create(); GLES3::MaterialStorage::get_singleton()->shaders.canvas_shader.version_bind_shader(data.canvas_shader_default_version, CanvasShaderGLES3::MODE_QUAD); + shadow_render.shader.initialize(); + shadow_render.shader_version = shadow_render.shader.version_create(); + { default_canvas_group_shader = material_storage->shader_allocate(); material_storage->shader_initialize(default_canvas_group_shader); @@ -2116,9 +2612,11 @@ void fragment() { } RasterizerCanvasGLES3::~RasterizerCanvasGLES3() { - GLES3::MaterialStorage *material_storage = GLES3::MaterialStorage::get_singleton(); + singleton = nullptr; + GLES3::MaterialStorage *material_storage = GLES3::MaterialStorage::get_singleton(); material_storage->shaders.canvas_shader.version_free(data.canvas_shader_default_version); + shadow_render.shader.version_free(shadow_render.shader_version); material_storage->material_free(default_canvas_group_material); material_storage->shader_free(default_canvas_group_shader); material_storage->material_free(default_clip_children_material); @@ -2134,6 +2632,15 @@ RasterizerCanvasGLES3::~RasterizerCanvasGLES3() { GLES3::TextureStorage::get_singleton()->canvas_texture_free(default_canvas_texture); memdelete_arr(state.instance_data_array); memdelete_arr(state.light_uniforms); + + if (state.shadow_fb != 0) { + glDeleteFramebuffers(1, &state.shadow_fb); + glDeleteTextures(1, &state.shadow_texture); + glDeleteRenderbuffers(1, &state.shadow_depth_buffer); + state.shadow_fb = 0; + state.shadow_texture = 0; + state.shadow_depth_buffer = 0; + } } #endif // GLES3_ENABLED diff --git a/drivers/gles3/rasterizer_canvas_gles3.h b/drivers/gles3/rasterizer_canvas_gles3.h index aee2782b62..d672d05e14 100644 --- a/drivers/gles3/rasterizer_canvas_gles3.h +++ b/drivers/gles3/rasterizer_canvas_gles3.h @@ -40,6 +40,7 @@ #include "storage/texture_storage.h" #include "shaders/canvas.glsl.gen.h" +#include "shaders/canvas_occlusion.glsl.gen.h" class RasterizerSceneGLES3; @@ -102,10 +103,40 @@ class RasterizerCanvasGLES3 : public RendererCanvasRender { struct CanvasLight { RID texture; + struct { + bool enabled = false; + float z_far; + float y_offset; + Transform2D directional_xform; + } shadow; }; RID_Owner<CanvasLight> canvas_light_owner; + struct OccluderPolygon { + RS::CanvasOccluderPolygonCullMode cull_mode = RS::CANVAS_OCCLUDER_POLYGON_CULL_DISABLED; + int line_point_count = 0; + GLuint vertex_buffer = 0; + GLuint vertex_array = 0; + GLuint index_buffer = 0; + + int sdf_point_count = 0; + int sdf_index_count = 0; + GLuint sdf_vertex_buffer = 0; + GLuint sdf_vertex_array = 0; + GLuint sdf_index_buffer = 0; + bool sdf_is_lines = false; + }; + + RID_Owner<OccluderPolygon> occluder_polygon_owner; + + void _update_shadow_atlas(); + + struct { + CanvasOcclusionShaderGLES3 shader; + RID shader_version; + } shadow_render; + struct LightUniform { float matrix[8]; //light to texture coordinate matrix float shadow_matrix[8]; //light to shadow coordinate matrix @@ -153,9 +184,9 @@ public: }; struct PolygonBuffers { - GLuint vertex_buffer; - GLuint vertex_array; - GLuint index_buffer; + GLuint vertex_buffer = 0; + GLuint vertex_array = 0; + GLuint index_buffer = 0; int count = 0; bool color_disabled = false; Color color; @@ -265,6 +296,11 @@ public: LightUniform *light_uniforms = nullptr; + GLuint shadow_texture = 0; + GLuint shadow_depth_buffer = 0; + GLuint shadow_fb = 0; + int shadow_texture_size = 2048; + bool using_directional_lights = false; RID current_tex = RID(); @@ -295,9 +331,6 @@ public: void draw_lens_distortion_rect(const Rect2 &p_rect, float p_k1, float p_k2, const Vector2 &p_eye_center, float p_oversample); void reset_canvas(); - void canvas_light_shadow_buffer_update(RID p_buffer, const Transform2D &p_light_xform, int p_light_mask, float p_near, float p_far, LightOccluderInstance *p_occluders, Projection *p_xform_cache); - - virtual void canvas_debug_viewport_shadows(Light *p_lights_with_shadow) override; RID light_create() override; void light_set_texture(RID p_rid, RID p_texture) override; diff --git a/drivers/gles3/rasterizer_gles3.cpp b/drivers/gles3/rasterizer_gles3.cpp index ea97dff522..b2d01b02fb 100644 --- a/drivers/gles3/rasterizer_gles3.cpp +++ b/drivers/gles3/rasterizer_gles3.cpp @@ -101,8 +101,7 @@ void RasterizerGLES3::begin_frame(double frame_step) { scene->set_time(time_total, frame_step); GLES3::Utilities *utils = GLES3::Utilities::get_singleton(); - utils->info.render_final = utils->info.render; - utils->info.render.reset(); + utils->_capture_timestamps_begin(); //scene->iteration(); } @@ -272,6 +271,13 @@ RasterizerGLES3::~RasterizerGLES3() { } void RasterizerGLES3::prepare_for_blitting_render_targets() { + // This is a hack, but this function is called one time after all viewports have been updated. + // So it marks the end of the frame for all viewports + // In the OpenGL renderer we have to call end_frame for each viewport so we can swap the + // buffers for each window before proceeding to the next. + // This allows us to only increment the frame after all viewports are done. + GLES3::Utilities *utils = GLES3::Utilities::get_singleton(); + utils->capture_timestamps_end(); } void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, DisplayServer::WindowID p_screen, const Rect2 &p_screen_rect, uint32_t p_layer) { diff --git a/drivers/gles3/rasterizer_scene_gles3.cpp b/drivers/gles3/rasterizer_scene_gles3.cpp index 6e9d4e5963..734911ccdb 100644 --- a/drivers/gles3/rasterizer_scene_gles3.cpp +++ b/drivers/gles3/rasterizer_scene_gles3.cpp @@ -595,6 +595,7 @@ void RasterizerSceneGLES3::_setup_sky(const RenderDataGLES3 *p_render_data, cons sky->reflection_dirty = true; } + glBindBufferBase(GL_UNIFORM_BUFFER, SKY_DIRECTIONAL_LIGHT_UNIFORM_LOCATION, sky_globals.directional_light_buffer); if (shader_data->uses_light) { sky_globals.directional_light_count = 0; for (int i = 0; i < (int)p_lights.size(); i++) { @@ -678,7 +679,6 @@ void RasterizerSceneGLES3::_setup_sky(const RenderDataGLES3 *p_render_data, cons } if (light_data_dirty) { - glBindBufferBase(GL_UNIFORM_BUFFER, SKY_DIRECTIONAL_LIGHT_UNIFORM_LOCATION, sky_globals.directional_light_buffer); glBufferData(GL_UNIFORM_BUFFER, sizeof(DirectionalLightData) * sky_globals.max_directional_lights, sky_globals.directional_lights, GL_STREAM_DRAW); glBindBuffer(GL_UNIFORM_BUFFER, 0); @@ -847,7 +847,7 @@ void RasterizerSceneGLES3::_update_sky_radiance(RID p_env, const Projection &p_p Projection cm; cm.set_perspective(90, 1, 0.01, 10.0); Projection correction; - correction.set_depth_correction(true); + correction.columns[1][1] = -1.0; cm = correction * cm; GLES3::MaterialStorage::get_singleton()->shaders.sky_shader.version_bind_shader(shader_data->version, SkyShaderGLES3::MODE_CUBEMAP); @@ -1265,7 +1265,7 @@ void RasterizerSceneGLES3::_fill_render_list(RenderListType p_render_list, const // Needs to be called after _setup_lights so that directional_light_count is accurate. void RasterizerSceneGLES3::_setup_environment(const RenderDataGLES3 *p_render_data, bool p_no_fog, const Size2i &p_screen_size, bool p_flip_y, const Color &p_default_bg_color, bool p_pancake_shadows) { Projection correction; - correction.set_depth_correction(p_flip_y); + correction.columns[1][1] = p_flip_y ? -1.0 : 1.0; Projection projection = correction * p_render_data->cam_projection; //store camera into ubo GLES3::MaterialStorage::store_camera(projection, scene_state.ubo.projection_matrix); @@ -1642,6 +1642,9 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_ ERR_FAIL_COND(rb.is_null()); } + GLES3::RenderTarget *rt = texture_storage->get_render_target(rb->render_target); + ERR_FAIL_COND(!rt); + // Assign render data // Use the format from rendererRD RenderDataGLES3 render_data; @@ -1729,8 +1732,20 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_ scene_state.ubo.emissive_exposure_normalization = -1.0; // Use default exposure normalization. + bool flip_y = !render_data.reflection_probe.is_valid(); + + if (rt->overridden.color.is_valid()) { + // If we've overridden the render target's color texture, then don't render upside down. + // We're probably rendering directly to an XR device. + flip_y = false; + } + if (!flip_y) { + // If we're rendering right-side up, then we need to change the winding order. + glFrontFace(GL_CW); + } + _setup_lights(&render_data, false, render_data.directional_light_count, render_data.omni_light_count, render_data.spot_light_count); - _setup_environment(&render_data, render_data.reflection_probe.is_valid(), screen_size, !render_data.reflection_probe.is_valid(), clear_color, false); + _setup_environment(&render_data, render_data.reflection_probe.is_valid(), screen_size, flip_y, clear_color, false); _fill_render_list(RENDER_LIST_OPAQUE, &render_data, PASS_MODE_COLOR); render_list[RENDER_LIST_OPAQUE].sort_by_key(); @@ -1792,7 +1807,7 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_ Projection projection = render_data.cam_projection; if (render_data.reflection_probe.is_valid()) { Projection correction; - correction.set_depth_correction(true); + correction.columns[1][1] = -1.0; projection = correction * render_data.cam_projection; } @@ -1811,7 +1826,7 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_ } } - glBindFramebuffer(GL_FRAMEBUFFER, rb->framebuffer); + glBindFramebuffer(GL_FRAMEBUFFER, rt->fbo); glViewport(0, 0, rb->width, rb->height); // Do depth prepass if it's explicitly enabled @@ -1917,6 +1932,11 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_ _render_list_template<PASS_MODE_COLOR_TRANSPARENT>(&render_list_params_alpha, &render_data, 0, render_list[RENDER_LIST_ALPHA].elements.size(), true); + if (!flip_y) { + // Restore the default winding order. + glFrontFace(GL_CCW); + } + if (rb.is_valid()) { _render_buffers_debug_draw(rb, p_shadow_atlas, p_occluder_debug_tex); } @@ -1973,6 +1993,8 @@ void RasterizerSceneGLES3::_render_list_template(RenderListParameters *p_params, } } + bool should_request_redraw = false; + for (uint32_t i = p_from_element; i < p_to_element; i++) { const GeometryInstanceSurface *surf = p_params->elements[i]; GeometryInstanceGLES3 *inst = surf->owner; @@ -2003,6 +2025,11 @@ void RasterizerSceneGLES3::_render_list_template(RenderListParameters *p_params, continue; } + //request a redraw if one of the shaders uses TIME + if (shader->uses_time) { + should_request_redraw = true; + } + if constexpr (p_pass_mode == PASS_MODE_COLOR_TRANSPARENT) { if (scene_state.current_depth_test != shader->depth_test) { if (shader->depth_test == GLES3::SceneShaderData::DEPTH_TEST_DISABLED) { @@ -2227,6 +2254,11 @@ void RasterizerSceneGLES3::_render_list_template(RenderListParameters *p_params, glDisableVertexAttribArray(15); } } + + // Make the actual redraw request + if (should_request_redraw) { + RenderingServerDefault::redraw_request(); + } } void RasterizerSceneGLES3::render_material(const Transform3D &p_cam_transform, const Projection &p_cam_projection, bool p_cam_orthogonal, const PagedArray<RenderGeometryInstance *> &p_instances, RID p_framebuffer, const Rect2i &p_region) { diff --git a/drivers/gles3/shader_gles3.h b/drivers/gles3/shader_gles3.h index 760b5e5ddb..3ab7642357 100644 --- a/drivers/gles3/shader_gles3.h +++ b/drivers/gles3/shader_gles3.h @@ -75,10 +75,10 @@ private: CharString general_defines; // A version is a high-level construct which is a combination of built-in and user-defined shader code, Each user-created Shader makes one version - // Variants use #ifdefs to toggle behaviour on and off to change behaviour of the shader + // Variants use #ifdefs to toggle behavior on and off to change behavior of the shader // All variants are compiled each time a new version is created - // Specializations use #ifdefs to toggle behaviour on and off for performance, on supporting hardware, they will compile a version with everything enabled, and then compile more copies to improve performance - // Use specializations to enable and disabled advanced features, use variants to toggle behaviour when different data may be used (e.g. using a samplerArray vs a sampler, or doing a depth prepass vs a color pass) + // Specializations use #ifdefs to toggle behavior on and off for performance, on supporting hardware, they will compile a version with everything enabled, and then compile more copies to improve performance + // Use specializations to enable and disabled advanced features, use variants to toggle behavior when different data may be used (e.g. using a samplerArray vs a sampler, or doing a depth prepass vs a color pass) struct Version { Vector<StringName> texture_uniforms; CharString uniforms; @@ -208,8 +208,10 @@ protected: spec = version->variants[p_variant].lookup_ptr(specialization_default_mask); } - ERR_FAIL_COND(!spec); // Should never happen - ERR_FAIL_COND(!spec->ok); // Should never happen + if (!spec || !spec->ok) { + WARN_PRINT_ONCE("shader failed to compile, unable to bind shader."); + return; + } glUseProgram(spec->id); current_shader = spec; diff --git a/drivers/gles3/shaders/SCsub b/drivers/gles3/shaders/SCsub index 83ffe8b1e1..b8bb08ec34 100644 --- a/drivers/gles3/shaders/SCsub +++ b/drivers/gles3/shaders/SCsub @@ -17,3 +17,5 @@ if "GLES3_GLSL" in env["BUILDERS"]: env.GLES3_GLSL("scene.glsl") env.GLES3_GLSL("sky.glsl") env.GLES3_GLSL("cubemap_filter.glsl") + env.GLES3_GLSL("canvas_occlusion.glsl") + env.GLES3_GLSL("canvas_sdf.glsl") diff --git a/drivers/gles3/shaders/canvas.glsl b/drivers/gles3/shaders/canvas.glsl index 23db41802e..ca806304c5 100644 --- a/drivers/gles3/shaders/canvas.glsl +++ b/drivers/gles3/shaders/canvas.glsl @@ -10,6 +10,7 @@ mode_instanced = #define USE_ATTRIBUTES \n#define USE_INSTANCING #[specializations] DISABLE_LIGHTING = false +USE_RGBA_SHADOWS = false #[vertex] @@ -213,8 +214,8 @@ void main() { #ifndef DISABLE_LIGHTING uniform sampler2D atlas_texture; //texunit:-2 +uniform sampler2D shadow_atlas_texture; //texunit:-3 #endif // DISABLE_LIGHTING -//uniform sampler2D shadow_atlas_texture; //texunit:-3 uniform sampler2D screen_texture; //texunit:-4 uniform sampler2D sdf_texture; //texunit:-5 uniform sampler2D normal_texture; //texunit:-6 @@ -245,6 +246,35 @@ layout(std140) uniform MaterialUniforms{ #endif #GLOBALS + +float vec4_to_float(vec4 p_vec) { + return dot(p_vec, vec4(1.0 / (255.0 * 255.0 * 255.0), 1.0 / (255.0 * 255.0), 1.0 / 255.0, 1.0)) * 2.0 - 1.0; +} + +vec2 screen_uv_to_sdf(vec2 p_uv) { + return screen_to_sdf * p_uv; +} + +float texture_sdf(vec2 p_sdf) { + vec2 uv = p_sdf * sdf_to_tex.xy + sdf_to_tex.zw; + float d = vec4_to_float(texture(sdf_texture, uv)); + d *= SDF_MAX_LENGTH; + return d * tex_to_sdf; +} + +vec2 texture_sdf_normal(vec2 p_sdf) { + vec2 uv = p_sdf * sdf_to_tex.xy + sdf_to_tex.zw; + + const float EPSILON = 0.001; + return normalize(vec2( + vec4_to_float(texture(sdf_texture, uv + vec2(EPSILON, 0.0))) - vec4_to_float(texture(sdf_texture, uv - vec2(EPSILON, 0.0))), + vec4_to_float(texture(sdf_texture, uv + vec2(0.0, EPSILON))) - vec4_to_float(texture(sdf_texture, uv - vec2(0.0, EPSILON))))); +} + +vec2 sdf_to_screen_uv(vec2 p_sdf) { + return p_sdf * sdf_to_screen; +} + #ifndef DISABLE_LIGHTING #ifdef LIGHT_CODE_USED @@ -299,6 +329,70 @@ vec3 light_normal_compute(vec3 light_vec, vec3 normal, vec3 base_color, vec3 lig } } +#ifdef USE_RGBA_SHADOWS + +#define SHADOW_DEPTH(m_uv) (dot(textureLod(shadow_atlas_texture, (m_uv), 0.0), vec4(1.0 / (255.0 * 255.0 * 255.0), 1.0 / (255.0 * 255.0), 1.0 / 255.0, 1.0)) * 2.0 - 1.0) + +#else + +#define SHADOW_DEPTH(m_uv) (textureLod(shadow_atlas_texture, (m_uv), 0.0).r) + +#endif + +#define SHADOW_TEST(m_uv) \ + { \ + highp float sd = SHADOW_DEPTH(m_uv); \ + shadow += step(sd, shadow_uv.z / shadow_uv.w); \ + } + +//float distance = length(shadow_pos); +vec4 light_shadow_compute(uint light_base, vec4 light_color, vec4 shadow_uv +#ifdef LIGHT_CODE_USED + , + vec3 shadow_modulate +#endif +) { + float shadow = 0.0; + uint shadow_mode = light_array[light_base].flags & LIGHT_FLAGS_FILTER_MASK; + + if (shadow_mode == LIGHT_FLAGS_SHADOW_NEAREST) { + SHADOW_TEST(shadow_uv.xy); + } else if (shadow_mode == LIGHT_FLAGS_SHADOW_PCF5) { + vec2 shadow_pixel_size = vec2(light_array[light_base].shadow_pixel_size, 0.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 2.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size); + SHADOW_TEST(shadow_uv.xy); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 2.0); + shadow /= 5.0; + } else { //PCF13 + vec2 shadow_pixel_size = vec2(light_array[light_base].shadow_pixel_size, 0.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 6.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 5.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 4.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 3.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size * 2.0); + SHADOW_TEST(shadow_uv.xy - shadow_pixel_size); + SHADOW_TEST(shadow_uv.xy); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 2.0); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 3.0); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 4.0); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 5.0); + SHADOW_TEST(shadow_uv.xy + shadow_pixel_size * 6.0); + shadow /= 13.0; + } + + vec4 shadow_color = unpackUnorm4x8(light_array[light_base].shadow_color); +#ifdef LIGHT_CODE_USED + shadow_color.rgb *= shadow_modulate; +#endif + + shadow_color.a *= light_color.a; //respect light alpha + + return mix(light_color, shadow_color, shadow); +} + void light_blend_compute(uint light_base, vec4 light_color, inout vec3 color) { uint blend_mode = light_array[light_base].flags & LIGHT_FLAGS_BLEND_MASK; @@ -527,6 +621,19 @@ void main() { } #endif + if (bool(light_array[light_base].flags & LIGHT_FLAGS_HAS_SHADOW)) { + vec2 shadow_pos = (vec4(shadow_vertex, 0.0, 1.0) * mat4(light_array[light_base].shadow_matrix[0], light_array[light_base].shadow_matrix[1], vec4(0.0, 0.0, 1.0, 0.0), vec4(0.0, 0.0, 0.0, 1.0))).xy; //multiply inverse given its transposed. Optimizer removes useless operations. + + vec4 shadow_uv = vec4(shadow_pos.x, light_array[light_base].shadow_y_ofs, shadow_pos.y * light_array[light_base].shadow_zfar_inv, 1.0); + + light_color = light_shadow_compute(light_base, light_color, shadow_uv +#ifdef LIGHT_CODE_USED + , + shadow_modulate.rgb +#endif + ); + } + light_blend_compute(light_base, light_color, color.rgb); } @@ -584,6 +691,46 @@ void main() { light_color.a = 0.0; } + if (bool(light_array[light_base].flags & LIGHT_FLAGS_HAS_SHADOW)) { + vec2 shadow_pos = (vec4(shadow_vertex, 0.0, 1.0) * mat4(light_array[light_base].shadow_matrix[0], light_array[light_base].shadow_matrix[1], vec4(0.0, 0.0, 1.0, 0.0), vec4(0.0, 0.0, 0.0, 1.0))).xy; //multiply inverse given its transposed. Optimizer removes useless operations. + + vec2 pos_norm = normalize(shadow_pos); + vec2 pos_abs = abs(pos_norm); + vec2 pos_box = pos_norm / max(pos_abs.x, pos_abs.y); + vec2 pos_rot = pos_norm * mat2(vec2(0.7071067811865476, -0.7071067811865476), vec2(0.7071067811865476, 0.7071067811865476)); //is there a faster way to 45 degrees rot? + float tex_ofs; + float dist; + if (pos_rot.y > 0.0) { + if (pos_rot.x > 0.0) { + tex_ofs = pos_box.y * 0.125 + 0.125; + dist = shadow_pos.x; + } else { + tex_ofs = pos_box.x * -0.125 + (0.25 + 0.125); + dist = shadow_pos.y; + } + } else { + if (pos_rot.x < 0.0) { + tex_ofs = pos_box.y * -0.125 + (0.5 + 0.125); + dist = -shadow_pos.x; + } else { + tex_ofs = pos_box.x * 0.125 + (0.75 + 0.125); + dist = -shadow_pos.y; + } + } + + dist *= light_array[light_base].shadow_zfar_inv; + + //float distance = length(shadow_pos); + vec4 shadow_uv = vec4(tex_ofs, light_array[light_base].shadow_y_ofs, dist, 1.0); + + light_color = light_shadow_compute(light_base, light_color, shadow_uv +#ifdef LIGHT_CODE_USED + , + shadow_modulate.rgb +#endif + ); + } + light_blend_compute(light_base, light_color, color.rgb); } #endif diff --git a/drivers/gles3/shaders/canvas_occlusion.glsl b/drivers/gles3/shaders/canvas_occlusion.glsl new file mode 100644 index 0000000000..512800839a --- /dev/null +++ b/drivers/gles3/shaders/canvas_occlusion.glsl @@ -0,0 +1,68 @@ +/* clang-format off */ +#[modes] + +mode_sdf = +mode_shadow = #define MODE_SHADOW +mode_shadow_RGBA = #define MODE_SHADOW \n#define USE_RGBA_SHADOWS + +#[specializations] + +#[vertex] + +layout(location = 0) in vec3 vertex; + +uniform highp mat4 projection; +uniform highp vec4 modelview1; +uniform highp vec4 modelview2; +uniform highp vec2 direction; +uniform highp float z_far; + +#ifdef MODE_SHADOW +out float depth; +#endif + +void main() { + highp vec4 vtx = vec4(vertex, 1.0) * mat4(modelview1, modelview2, vec4(0.0, 0.0, 1.0, 0.0), vec4(0.0, 0.0, 0.0, 1.0)); + +#ifdef MODE_SHADOW + depth = dot(direction, vtx.xy); +#endif + gl_Position = projection * vtx; +} + +#[fragment] + + +uniform highp mat4 projection; +uniform highp vec4 modelview1; +uniform highp vec4 modelview2; +uniform highp vec2 direction; +uniform highp float z_far; + +#ifdef MODE_SHADOW +in highp float depth; +#endif + +#ifdef USE_RGBA_SHADOWS +layout(location = 0) out lowp vec4 out_buf; +#else +layout(location = 0) out highp float out_buf; +#endif + +void main() { + float out_depth = 1.0; + +#ifdef MODE_SHADOW + out_depth = depth / z_far; +#endif + +#ifdef USE_RGBA_SHADOWS + out_depth = clamp(out_depth, -1.0, 1.0); + out_depth = out_depth * 0.5 + 0.5; + highp vec4 comp = fract(out_depth * vec4(255.0 * 255.0 * 255.0, 255.0 * 255.0, 255.0, 1.0)); + comp -= comp.xxyz * vec4(0.0, 1.0 / 255.0, 1.0 / 255.0, 1.0 / 255.0); + out_buf = comp; +#else + out_buf = out_depth; +#endif +} diff --git a/drivers/gles3/shaders/canvas_sdf.glsl b/drivers/gles3/shaders/canvas_sdf.glsl new file mode 100644 index 0000000000..424ec22457 --- /dev/null +++ b/drivers/gles3/shaders/canvas_sdf.glsl @@ -0,0 +1,205 @@ +/* clang-format off */ +#[modes] + +mode_load = #define MODE_LOAD +mode_load_shrink = #define MODE_LOAD_SHRINK +mode_process = #define MODE_PROCESS +mode_store = #define MODE_STORE +mode_store_shrink = #define MODE_STORE_SHRINK + +#[specializations] + +#[vertex] + +layout(location = 0) in vec2 vertex_attrib; + +/* clang-format on */ + +uniform ivec2 size; +uniform int stride; +uniform int shift; +uniform ivec2 base_size; + +void main() { + gl_Position = vec4(vertex_attrib, 1.0, 1.0); +} + +/* clang-format off */ +#[fragment] + +#define SDF_MAX_LENGTH 16384.0 + +#if defined(MODE_LOAD) || defined(MODE_LOAD_SHRINK) +uniform lowp sampler2D src_pixels;//texunit:0 +#else +uniform highp isampler2D src_process;//texunit:0 +#endif + +uniform ivec2 size; +uniform int stride; +uniform int shift; +uniform ivec2 base_size; + +#if defined(MODE_LOAD) || defined(MODE_LOAD_SHRINK) || defined(MODE_PROCESS) +layout(location = 0) out ivec4 distance_field; +#else +layout(location = 0) out vec4 distance_field; +#endif + +vec4 float_to_vec4(float p_float) { + highp vec4 comp = fract(p_float * vec4(255.0 * 255.0 * 255.0, 255.0 * 255.0, 255.0, 1.0)); + comp -= comp.xxyz * vec4(0.0, 1.0 / 255.0, 1.0 / 255.0, 1.0 / 255.0); + return comp; +} + +void main() { + ivec2 pos = ivec2(gl_FragCoord.xy); + +#ifdef MODE_LOAD + + bool solid = texelFetch(src_pixels, pos, 0).r > 0.5; + distance_field = solid ? ivec4(ivec2(-32767), 0, 0) : ivec4(ivec2(32767), 0, 0); +#endif + +#ifdef MODE_LOAD_SHRINK + + int s = 1 << shift; + ivec2 base = pos << shift; + ivec2 center = base + ivec2(shift); + + ivec2 rel = ivec2(32767); + float d = 1e20; + int found = 0; + int solid_found = 0; + for (int i = 0; i < s; i++) { + for (int j = 0; j < s; j++) { + ivec2 src_pos = base + ivec2(i, j); + if (any(greaterThanEqual(src_pos, base_size))) { + continue; + } + bool solid = texelFetch(src_pixels, src_pos, 0).r > 0.5; + if (solid) { + float dist = length(vec2(src_pos - center)); + if (dist < d) { + d = dist; + rel = src_pos; + } + solid_found++; + } + found++; + } + } + + if (solid_found == found) { + //mark solid only if all are solid + rel = ivec2(-32767); + } + + distance_field = ivec4(rel, 0, 0); +#endif + +#ifdef MODE_PROCESS + + ivec2 base = pos << shift; + ivec2 center = base + ivec2(shift); + + ivec2 rel = texelFetch(src_process, pos, 0).xy; + + bool solid = rel.x < 0; + + if (solid) { + rel = -rel - ivec2(1); + } + + if (center != rel) { + //only process if it does not point to itself + const int ofs_table_size = 8; + const ivec2 ofs_table[ofs_table_size] = ivec2[]( + ivec2(-1, -1), + ivec2(0, -1), + ivec2(+1, -1), + + ivec2(-1, 0), + ivec2(+1, 0), + + ivec2(-1, +1), + ivec2(0, +1), + ivec2(+1, +1)); + + float dist = length(vec2(rel - center)); + for (int i = 0; i < ofs_table_size; i++) { + ivec2 src_pos = pos + ofs_table[i] * stride; + if (any(lessThan(src_pos, ivec2(0))) || any(greaterThanEqual(src_pos, size))) { + continue; + } + ivec2 src_rel = texelFetch(src_process, src_pos, 0).xy; + bool src_solid = src_rel.x < 0; + if (src_solid) { + src_rel = -src_rel - ivec2(1); + } + + if (src_solid != solid) { + src_rel = ivec2(src_pos << shift); //point to itself if of different type + } + + float src_dist = length(vec2(src_rel - center)); + if (src_dist < dist) { + dist = src_dist; + rel = src_rel; + } + } + } + + if (solid) { + rel = -rel - ivec2(1); + } + + distance_field = ivec4(rel, 0, 0); +#endif + +#ifdef MODE_STORE + + ivec2 rel = texelFetch(src_process, pos, 0).xy; + + bool solid = rel.x < 0; + + if (solid) { + rel = -rel - ivec2(1); + } + + float d = length(vec2(rel - pos)); + + if (solid) { + d = -d; + } + + d /= SDF_MAX_LENGTH; + d = clamp(d, -1.0, 1.0); + distance_field = float_to_vec4(d*0.5+0.5); + +#endif + +#ifdef MODE_STORE_SHRINK + + ivec2 base = pos << shift; + ivec2 center = base + ivec2(shift); + + ivec2 rel = texelFetch(src_process, pos, 0).xy; + + bool solid = rel.x < 0; + + if (solid) { + rel = -rel - ivec2(1); + } + + float d = length(vec2(rel - center)); + + if (solid) { + d = -d; + } + d /= SDF_MAX_LENGTH; + d = clamp(d, -1.0, 1.0); + distance_field = float_to_vec4(d*0.5+0.5); + +#endif +} diff --git a/drivers/gles3/shaders/canvas_shadow.glsl b/drivers/gles3/shaders/canvas_shadow.glsl deleted file mode 100644 index 94485abd11..0000000000 --- a/drivers/gles3/shaders/canvas_shadow.glsl +++ /dev/null @@ -1,60 +0,0 @@ -/* clang-format off */ -[vertex] - -#ifdef USE_GLES_OVER_GL -#define lowp -#define mediump -#define highp -#else -precision highp float; -precision highp int; -#endif - -layout(location = 0) in highp vec3 vertex; - -uniform highp mat4 projection_matrix; -/* clang-format on */ -uniform highp mat4 light_matrix; -uniform highp mat4 model_matrix; -uniform highp float distance_norm; - -out highp vec4 position_interp; - -void main() { - gl_Position = projection_matrix * (light_matrix * (model_matrix * vec4(vertex, 1.0))); - position_interp = gl_Position; -} - -/* clang-format off */ -[fragment] - -#ifdef USE_GLES_OVER_GL -#define lowp -#define mediump -#define highp -#else -#if defined(USE_HIGHP_PRECISION) -precision highp float; -precision highp int; -#else -precision mediump float; -precision mediump int; -#endif -#endif - -in highp vec4 position_interp; -/* clang-format on */ - -void main() { - highp float depth = ((position_interp.z / position_interp.w) + 1.0) * 0.5 + 0.0; // bias - -#ifdef USE_RGBA_SHADOWS - - highp vec4 comp = fract(depth * vec4(255.0 * 255.0 * 255.0, 255.0 * 255.0, 255.0, 1.0)); - comp -= comp.xxyz * vec4(0.0, 1.0 / 255.0, 1.0 / 255.0, 1.0 / 255.0); - frag_color = comp; -#else - - frag_color = vec4(depth); -#endif -} diff --git a/drivers/gles3/storage/config.cpp b/drivers/gles3/storage/config.cpp index 943df3b156..9b496c0999 100644 --- a/drivers/gles3/storage/config.cpp +++ b/drivers/gles3/storage/config.cpp @@ -71,7 +71,7 @@ Config::Config() { s3tc_supported = true; rgtc_supported = true; //RGTC - core since OpenGL version 3.0 #else - float_texture_supported = extensions.has("GL_ARB_texture_float") || extensions.has("GL_OES_texture_float"); + float_texture_supported = extensions.has("GL_EXT_color_buffer_float"); etc2_supported = true; #if defined(ANDROID_ENABLED) || defined(IOS_ENABLED) // Some Android devices report support for S3TC but we don't expect that and don't export the textures. @@ -84,28 +84,18 @@ Config::Config() { rgtc_supported = extensions.has("GL_EXT_texture_compression_rgtc") || extensions.has("GL_ARB_texture_compression_rgtc") || extensions.has("EXT_texture_compression_rgtc"); #endif -#ifdef GLES_OVER_GL - use_rgba_2d_shadows = false; -#else - use_rgba_2d_shadows = !(float_texture_supported && extensions.has("GL_EXT_texture_rg")); -#endif - glGetIntegerv(GL_MAX_VERTEX_TEXTURE_IMAGE_UNITS, &max_vertex_texture_image_units); glGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &max_texture_image_units); glGetIntegerv(GL_MAX_TEXTURE_SIZE, &max_texture_size); glGetIntegerv(GL_MAX_UNIFORM_BLOCK_SIZE, &max_uniform_buffer_size); - glGetIntegerv(GL_MAX_VIEWPORT_DIMS, &max_viewport_size); + glGetIntegerv(GL_MAX_VIEWPORT_DIMS, max_viewport_size); glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &uniform_buffer_offset_alignment); - // the use skeleton software path should be used if either float texture is not supported, - // OR max_vertex_texture_image_units is zero - use_skeleton_software = (float_texture_supported == false) || (max_vertex_texture_image_units == 0); - support_anisotropic_filter = extensions.has("GL_EXT_texture_filter_anisotropic"); if (support_anisotropic_filter) { glGetFloatv(_GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT, &anisotropic_level); - anisotropic_level = MIN(float(1 << int(ProjectSettings::get_singleton()->get("rendering/textures/default_filters/anisotropic_filtering_level"))), anisotropic_level); + anisotropic_level = MIN(float(1 << int(GLOBAL_GET("rendering/textures/default_filters/anisotropic_filtering_level"))), anisotropic_level); } multiview_supported = extensions.has("GL_OVR_multiview2") || extensions.has("GL_OVR_multiview"); diff --git a/drivers/gles3/storage/config.h b/drivers/gles3/storage/config.h index d4b38acd18..87202fde84 100644 --- a/drivers/gles3/storage/config.h +++ b/drivers/gles3/storage/config.h @@ -56,15 +56,13 @@ private: public: bool use_nearest_mip_filter = false; - bool use_skeleton_software = false; bool use_depth_prepass = true; - bool use_rgba_2d_shadows = false; int max_vertex_texture_image_units = 0; int max_texture_image_units = 0; int max_texture_size = 0; + int max_viewport_size[2] = { 0, 0 }; int max_uniform_buffer_size = 0; - int max_viewport_size = 0; int max_renderable_elements = 0; int max_renderable_lights = 0; int max_lights_per_object = 0; diff --git a/drivers/gles3/storage/material_storage.cpp b/drivers/gles3/storage/material_storage.cpp index f5241e33ab..2d6a793c9e 100644 --- a/drivers/gles3/storage/material_storage.cpp +++ b/drivers/gles3/storage/material_storage.cpp @@ -89,7 +89,7 @@ _FORCE_INLINE_ static void _fill_std140_variant_ubo_value(ShaderLanguage::DataTy gui[j + 3] = 0; // ignored } } else { - int v = value; + uint32_t v = value; gui[0] = v & 1 ? 1 : 0; gui[1] = v & 2 ? 1 : 0; } @@ -116,7 +116,7 @@ _FORCE_INLINE_ static void _fill_std140_variant_ubo_value(ShaderLanguage::DataTy gui[j + 3] = 0; // ignored } } else { - int v = value; + uint32_t v = value; gui[0] = (v & 1) ? 1 : 0; gui[1] = (v & 2) ? 1 : 0; gui[2] = (v & 4) ? 1 : 0; @@ -145,7 +145,7 @@ _FORCE_INLINE_ static void _fill_std140_variant_ubo_value(ShaderLanguage::DataTy } } } else { - int v = value; + uint32_t v = value; gui[0] = (v & 1) ? 1 : 0; gui[1] = (v & 2) ? 1 : 0; gui[2] = (v & 4) ? 1 : 0; @@ -728,7 +728,7 @@ _FORCE_INLINE_ static void _fill_std140_ubo_value(ShaderLanguage::DataType type, switch (type) { case ShaderLanguage::TYPE_BOOL: { uint32_t *gui = (uint32_t *)data; - *gui = value[0].boolean ? 1 : 0; + gui[0] = value[0].boolean ? 1 : 0; } break; case ShaderLanguage::TYPE_BVEC2: { uint32_t *gui = (uint32_t *)data; @@ -2214,7 +2214,7 @@ void MaterialStorage::global_shader_parameters_load_settings(bool p_load_texture for (const PropertyInfo &E : settings) { if (E.name.begins_with("shader_globals/")) { StringName name = E.name.get_slice("/", 1); - Dictionary d = ProjectSettings::get_singleton()->get(E.name); + Dictionary d = GLOBAL_GET(E.name); ERR_CONTINUE(!d.has("type")); ERR_CONTINUE(!d.has("value")); @@ -2314,7 +2314,7 @@ void MaterialStorage::global_shader_parameters_instance_free(RID p_instance) { global_shader_uniforms.instance_buffer_pos.erase(p_instance); } -void MaterialStorage::global_shader_parameters_instance_update(RID p_instance, int p_index, const Variant &p_value) { +void MaterialStorage::global_shader_parameters_instance_update(RID p_instance, int p_index, const Variant &p_value, int p_flags_count) { if (!global_shader_uniforms.instance_buffer_pos.has(p_instance)) { return; //just not allocated, ignore } @@ -2324,7 +2324,9 @@ void MaterialStorage::global_shader_parameters_instance_update(RID p_instance, i return; //again, not allocated, ignore } ERR_FAIL_INDEX(p_index, ShaderLanguage::MAX_INSTANCE_UNIFORM_INDICES); - ERR_FAIL_COND_MSG(p_value.get_type() > Variant::COLOR, "Unsupported variant type for instance parameter: " + Variant::get_type_name(p_value.get_type())); //anything greater not supported + + Variant::Type value_type = p_value.get_type(); + ERR_FAIL_COND_MSG(p_value.get_type() > Variant::COLOR, "Unsupported variant type for instance parameter: " + Variant::get_type_name(value_type)); //anything greater not supported ShaderLanguage::DataType datatype_from_value[Variant::COLOR + 1] = { ShaderLanguage::TYPE_MAX, //nil @@ -2350,9 +2352,24 @@ void MaterialStorage::global_shader_parameters_instance_update(RID p_instance, i ShaderLanguage::TYPE_VEC4 //color }; - ShaderLanguage::DataType datatype = datatype_from_value[p_value.get_type()]; + ShaderLanguage::DataType datatype = ShaderLanguage::TYPE_MAX; + if (value_type == Variant::INT && p_flags_count > 0) { + switch (p_flags_count) { + case 1: + datatype = ShaderLanguage::TYPE_BVEC2; + break; + case 2: + datatype = ShaderLanguage::TYPE_BVEC3; + break; + case 3: + datatype = ShaderLanguage::TYPE_BVEC4; + break; + } + } else { + datatype = datatype_from_value[value_type]; + } - ERR_FAIL_COND_MSG(datatype == ShaderLanguage::TYPE_MAX, "Unsupported variant type for instance parameter: " + Variant::get_type_name(p_value.get_type())); //anything greater not supported + ERR_FAIL_COND_MSG(datatype == ShaderLanguage::TYPE_MAX, "Unsupported variant type for instance parameter: " + Variant::get_type_name(value_type)); //anything greater not supported pos += p_index; @@ -3341,6 +3358,9 @@ void SceneShaderData::set_code(const String &p_code) { actions.usage_flag_pointers["ALPHA"] = &uses_alpha; actions.usage_flag_pointers["ALPHA_SCISSOR_THRESHOLD"] = &uses_alpha_clip; + // Use alpha clip pipeline for alpha hash/dither. + // This prevents sorting issues inherent to alpha blending and allows such materials to cast shadows. + actions.usage_flag_pointers["ALPHA_HASH_SCALE"] = &uses_alpha_clip; actions.render_mode_flags["depth_prepass_alpha"] = &uses_depth_pre_pass; actions.usage_flag_pointers["SSS_STRENGTH"] = &uses_sss; diff --git a/drivers/gles3/storage/material_storage.h b/drivers/gles3/storage/material_storage.h index 6504c7748c..24d9a0fee1 100644 --- a/drivers/gles3/storage/material_storage.h +++ b/drivers/gles3/storage/material_storage.h @@ -530,7 +530,7 @@ public: virtual int32_t global_shader_parameters_instance_allocate(RID p_instance) override; virtual void global_shader_parameters_instance_free(RID p_instance) override; - virtual void global_shader_parameters_instance_update(RID p_instance, int p_index, const Variant &p_value) override; + virtual void global_shader_parameters_instance_update(RID p_instance, int p_index, const Variant &p_value, int p_flags_count = 0) override; GLuint global_shader_parameters_get_uniform_buffer() const; diff --git a/drivers/gles3/storage/mesh_storage.cpp b/drivers/gles3/storage/mesh_storage.cpp index 11ce31856d..9ec0fc0286 100644 --- a/drivers/gles3/storage/mesh_storage.cpp +++ b/drivers/gles3/storage/mesh_storage.cpp @@ -1004,7 +1004,7 @@ void MeshStorage::multimesh_set_mesh(RID p_multimesh, RID p_mesh) { #define MULTIMESH_DIRTY_REGION_SIZE 512 void MeshStorage::_multimesh_make_local(MultiMesh *multimesh) const { - if (multimesh->data_cache.size() > 0) { + if (multimesh->data_cache.size() > 0 || multimesh->instances == 0) { return; //already local } ERR_FAIL_COND(multimesh->data_cache.size() > 0); @@ -1421,7 +1421,7 @@ Vector<float> MeshStorage::multimesh_get_buffer(RID p_multimesh) const { MultiMesh *multimesh = multimesh_owner.get_or_null(p_multimesh); ERR_FAIL_COND_V(!multimesh, Vector<float>()); Vector<float> ret; - if (multimesh->buffer == 0) { + if (multimesh->buffer == 0 || multimesh->instances == 0) { return Vector<float>(); } else if (multimesh->data_cache.size()) { ret = multimesh->data_cache; diff --git a/drivers/gles3/storage/render_scene_buffers_gles3.cpp b/drivers/gles3/storage/render_scene_buffers_gles3.cpp index b8e4530f56..e0e78de728 100644 --- a/drivers/gles3/storage/render_scene_buffers_gles3.cpp +++ b/drivers/gles3/storage/render_scene_buffers_gles3.cpp @@ -33,17 +33,12 @@ #include "render_scene_buffers_gles3.h" #include "texture_storage.h" -#ifdef ANDROID_ENABLED -#define glFramebufferTextureMultiviewOVR GLES3::Config::get_singleton()->eglFramebufferTextureMultiviewOVR -#endif - RenderSceneBuffersGLES3::~RenderSceneBuffersGLES3() { free_render_buffer_data(); } void RenderSceneBuffersGLES3::configure(RID p_render_target, const Size2i p_internal_size, const Size2i p_target_size, float p_fsr_sharpness, float p_texture_mipmap_bias, RS::ViewportMSAA p_msaa, RenderingServer::ViewportScreenSpaceAA p_screen_space_aa, bool p_use_taa, bool p_use_debanding, uint32_t p_view_count) { GLES3::TextureStorage *texture_storage = GLES3::TextureStorage::get_singleton(); - GLES3::Config *config = GLES3::Config::get_singleton(); //internal_size.x = p_internal_size.x; // ignore for now //internal_size.y = p_internal_size.y; @@ -62,66 +57,9 @@ void RenderSceneBuffersGLES3::configure(RID p_render_target, const Size2i p_inte GLES3::RenderTarget *rt = texture_storage->get_render_target(p_render_target); is_transparent = rt->is_transparent; - - // framebuffer - glGenFramebuffers(1, &framebuffer); - glBindFramebuffer(GL_FRAMEBUFFER, framebuffer); - - if (view_count > 1 && config->multiview_supported) { - glBindTexture(GL_TEXTURE_2D_ARRAY, rt->color); - glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, rt->color, 0, 0, view_count); - } else { - glBindTexture(GL_TEXTURE_2D, rt->color); - glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->color, 0); - } - - glGenTextures(1, &depth_texture); - if (view_count > 1 && config->multiview_supported) { - glBindTexture(GL_TEXTURE_2D_ARRAY, depth_texture); - glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_DEPTH_COMPONENT24, rt->size.x, rt->size.y, view_count, 0, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, nullptr); - - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - } else { - glBindTexture(GL_TEXTURE_2D, depth_texture); - glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH_COMPONENT24, rt->size.x, rt->size.y, 0, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, nullptr); - - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - } - - if (view_count > 1 && config->multiview_supported) { - glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depth_texture, 0, 0, view_count); - } else { - glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, depth_texture, 0); - } - - GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER); - - glBindTexture(GL_TEXTURE_2D, 0); - glBindTexture(GL_TEXTURE_2D_ARRAY, 0); - glBindFramebuffer(GL_FRAMEBUFFER, texture_storage->system_fbo); - - if (status != GL_FRAMEBUFFER_COMPLETE) { - free_render_buffer_data(); - WARN_PRINT("Could not create 3D renderbuffer, status: " + texture_storage->get_framebuffer_error(status)); - return; - } } void RenderSceneBuffersGLES3::free_render_buffer_data() { - if (depth_texture) { - glDeleteTextures(1, &depth_texture); - depth_texture = 0; - } - if (framebuffer) { - glDeleteFramebuffers(1, &framebuffer); - framebuffer = 0; - } } #endif // GLES3_ENABLED diff --git a/drivers/gles3/storage/render_scene_buffers_gles3.h b/drivers/gles3/storage/render_scene_buffers_gles3.h index dbedbd22c3..092c14e1b8 100644 --- a/drivers/gles3/storage/render_scene_buffers_gles3.h +++ b/drivers/gles3/storage/render_scene_buffers_gles3.h @@ -61,9 +61,6 @@ public: bool is_transparent = false; RID render_target; - GLuint internal_texture = 0; // Used for rendering when post effects are enabled - GLuint depth_texture = 0; // Main depth texture - GLuint framebuffer = 0; // Main framebuffer, contains internal_texture and depth_texture or render_target->color and depth_texture //built-in textures used for ping pong image processing and blurring struct Blur { diff --git a/drivers/gles3/storage/texture_storage.cpp b/drivers/gles3/storage/texture_storage.cpp index 524ff14cc9..51e22fe779 100644 --- a/drivers/gles3/storage/texture_storage.cpp +++ b/drivers/gles3/storage/texture_storage.cpp @@ -207,6 +207,11 @@ TextureStorage::TextureStorage() { glBindTexture(GL_TEXTURE_2D, 0); + { + sdf_shader.shader.initialize(); + sdf_shader.shader_version = sdf_shader.shader.version_create(); + } + #ifdef GLES_OVER_GL glEnable(GL_PROGRAM_POINT_SIZE); #endif @@ -222,6 +227,7 @@ TextureStorage::~TextureStorage() { texture_atlas.texture = 0; glDeleteFramebuffers(1, &texture_atlas.framebuffer); texture_atlas.framebuffer = 0; + sdf_shader.shader.version_free(sdf_shader.shader_version); } //TODO, move back to storage @@ -276,55 +282,6 @@ void TextureStorage::canvas_texture_set_texture_repeat(RID p_canvas_texture, RS: ct->texture_repeat = p_repeat; } -/* CANVAS SHADOW */ - -RID TextureStorage::canvas_light_shadow_buffer_create(int p_width) { - Config *config = Config::get_singleton(); - CanvasLightShadow *cls = memnew(CanvasLightShadow); - - if (p_width > config->max_texture_size) { - p_width = config->max_texture_size; - } - - cls->size = p_width; - cls->height = 16; - - glActiveTexture(GL_TEXTURE0); - - glGenFramebuffers(1, &cls->fbo); - glBindFramebuffer(GL_FRAMEBUFFER, cls->fbo); - - glGenRenderbuffers(1, &cls->depth); - glBindRenderbuffer(GL_RENDERBUFFER, cls->depth); - glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT16, cls->size, cls->height); - glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, cls->depth); - - glGenTextures(1, &cls->distance); - glBindTexture(GL_TEXTURE_2D, cls->distance); - if (config->use_rgba_2d_shadows) { - glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, cls->size, cls->height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr); - } else { - glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, cls->size, cls->height, 0, GL_RED, GL_FLOAT, nullptr); - } - - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, cls->distance, 0); - - GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER); - //printf("errnum: %x\n",status); - glBindFramebuffer(GL_FRAMEBUFFER, GLES3::TextureStorage::system_fbo); - - if (status != GL_FRAMEBUFFER_COMPLETE) { - memdelete(cls); - ERR_FAIL_COND_V(status != GL_FRAMEBUFFER_COMPLETE, RID()); - } - - return canvas_light_shadow_owner.make_rid(cls); -} - /* Texture API */ Ref<Image> TextureStorage::_get_gl_image_and_format(const Ref<Image> &p_image, Image::Format p_format, Image::Format &r_real_format, GLenum &r_gl_format, GLenum &r_gl_internal_format, GLenum &r_gl_type, bool &r_compressed, bool p_force_decompress) const { @@ -657,7 +614,9 @@ void TextureStorage::texture_free(RID p_texture) { } if (t->tex_id != 0) { - glDeleteTextures(1, &t->tex_id); + if (!t->is_external) { + glDeleteTextures(1, &t->tex_id); + } t->tex_id = 0; } @@ -723,6 +682,35 @@ void TextureStorage::texture_proxy_initialize(RID p_texture, RID p_base) { texture_owner.initialize_rid(p_texture, proxy_tex); } +RID TextureStorage::texture_create_external(Texture::Type p_type, Image::Format p_format, unsigned int p_image, int p_width, int p_height, int p_depth, int p_layers, RS::TextureLayeredType p_layered_type) { + Texture texture; + texture.active = true; + texture.is_external = true; + texture.type = p_type; + + switch (p_type) { + case Texture::TYPE_2D: { + texture.target = GL_TEXTURE_2D; + } break; + case Texture::TYPE_3D: { + texture.target = GL_TEXTURE_3D; + } break; + case Texture::TYPE_LAYERED: { + texture.target = GL_TEXTURE_2D_ARRAY; + } break; + } + + texture.real_format = texture.format = p_format; + texture.tex_id = p_image; + texture.alloc_width = texture.width = p_width; + texture.alloc_height = texture.height = p_height; + texture.depth = p_depth; + texture.layers = p_layers; + texture.layered_type = p_layered_type; + + return texture_owner.make_rid(texture); +} + void TextureStorage::texture_2d_update(RID p_texture, const Ref<Image> &p_image, int p_layer) { texture_set_data(p_texture, p_image, p_layer); #ifdef TOOLS_ENABLED @@ -733,6 +721,26 @@ void TextureStorage::texture_2d_update(RID p_texture, const Ref<Image> &p_image, } void TextureStorage::texture_proxy_update(RID p_texture, RID p_proxy_to) { + Texture *tex = texture_owner.get_or_null(p_texture); + ERR_FAIL_COND(!tex); + ERR_FAIL_COND(!tex->is_proxy); + Texture *proxy_to = texture_owner.get_or_null(p_proxy_to); + ERR_FAIL_COND(!proxy_to); + ERR_FAIL_COND(proxy_to->is_proxy); + + if (tex->proxy_to.is_valid()) { + Texture *prev_tex = texture_owner.get_or_null(tex->proxy_to); + ERR_FAIL_COND(!prev_tex); + prev_tex->proxies.erase(p_texture); + } + + *tex = *proxy_to; + + tex->proxy_to = p_proxy_to; + tex->is_render_target = false; + tex->is_proxy = true; + tex->proxies.clear(); + proxy_to->proxies.push_back(p_texture); } void TextureStorage::texture_2d_placeholder_initialize(RID p_texture) { @@ -790,6 +798,7 @@ Ref<Image> TextureStorage::texture_2d_get(RID p_texture) const { #ifdef GLES_OVER_GL // OpenGL 3.3 supports glGetTexImage which is faster and simpler than glReadPixels. + // It also allows for reading compressed textures, mipmaps, and more formats. Vector<uint8_t> data; int data_size = Image::get_image_data_size(texture->alloc_width, texture->alloc_height, texture->real_format, texture->mipmaps > 1); @@ -826,8 +835,65 @@ Ref<Image> TextureStorage::texture_2d_get(RID p_texture) const { image->convert(texture->format); } #else - // Support for Web and Mobile will come later. - Ref<Image> image; + + Vector<uint8_t> data; + + // On web and mobile we always read an RGBA8 image with no mipmaps. + int data_size = Image::get_image_data_size(texture->alloc_width, texture->alloc_height, Image::FORMAT_RGBA8, false); + + data.resize(data_size * 2); //add some memory at the end, just in case for buggy drivers + uint8_t *w = data.ptrw(); + + GLuint temp_framebuffer; + glGenFramebuffers(1, &temp_framebuffer); + + GLuint temp_color_texture; + glGenTextures(1, &temp_color_texture); + + glBindFramebuffer(GL_FRAMEBUFFER, temp_framebuffer); + + glBindTexture(GL_TEXTURE_2D, temp_color_texture); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, texture->alloc_width, texture->alloc_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); + + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR); + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, temp_color_texture, 0); + + glDepthMask(GL_FALSE); + glDisable(GL_DEPTH_TEST); + glDisable(GL_CULL_FACE); + glDisable(GL_BLEND); + glDepthFunc(GL_LEQUAL); + glColorMask(1, 1, 1, 1); + glActiveTexture(GL_TEXTURE0); + glBindTexture(GL_TEXTURE_2D, texture->tex_id); + + glViewport(0, 0, texture->alloc_width, texture->alloc_height); + glClearColor(0.0, 0.0, 0.0, 0.0); + glClear(GL_COLOR_BUFFER_BIT); + + CopyEffects::get_singleton()->copy_to_rect(Rect2i(0, 0, 1.0, 1.0)); + + glReadPixels(0, 0, texture->alloc_width, texture->alloc_height, GL_RGBA, GL_UNSIGNED_BYTE, &w[0]); + + glBindFramebuffer(GL_FRAMEBUFFER, 0); + glDeleteTextures(1, &temp_color_texture); + glDeleteFramebuffers(1, &temp_framebuffer); + + data.resize(data_size); + + ERR_FAIL_COND_V(data.size() == 0, Ref<Image>()); + Ref<Image> image = Image::create_from_data(texture->width, texture->height, false, Image::FORMAT_RGBA8, data); + ERR_FAIL_COND_V(image->is_empty(), Ref<Image>()); + + if (texture->format != Image::FORMAT_RGBA8) { + image->convert(texture->format); + } + + if (texture->mipmaps > 1) { + image->generate_mipmaps(); + } + #endif #ifdef TOOLS_ENABLED @@ -976,6 +1042,10 @@ Size2 TextureStorage::texture_size_with_proxy(RID p_texture) { } } +RID TextureStorage::texture_get_rd_texture_rid(RID p_texture, bool p_srgb) const { + return RID(); +} + void TextureStorage::texture_set_data(RID p_texture, const Ref<Image> &p_image, int p_layer) { Texture *texture = texture_owner.get_or_null(p_texture); @@ -1444,43 +1514,74 @@ void TextureStorage::_update_render_target(RenderTarget *rt) { glDepthMask(GL_FALSE); { - /* Front FBO */ + Texture *texture; + bool use_multiview = rt->view_count > 1 && config->multiview_supported; + GLenum texture_target = use_multiview ? GL_TEXTURE_2D_ARRAY : GL_TEXTURE_2D; - Texture *texture = get_texture(rt->texture); - ERR_FAIL_COND(!texture); + /* Front FBO */ - // framebuffer glGenFramebuffers(1, &rt->fbo); glBindFramebuffer(GL_FRAMEBUFFER, rt->fbo); // color - glGenTextures(1, &rt->color); - if (rt->view_count > 1 && config->multiview_supported) { - glBindTexture(GL_TEXTURE_2D_ARRAY, rt->color); - glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, rt->color_internal_format, rt->size.x, rt->size.y, rt->view_count, 0, rt->color_format, rt->color_type, nullptr); - - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + if (rt->overridden.color.is_valid()) { + texture = get_texture(rt->overridden.color); + ERR_FAIL_COND(!texture); + + rt->color = texture->tex_id; + rt->size = Size2i(texture->width, texture->height); } else { - glBindTexture(GL_TEXTURE_2D, rt->color); - glTexImage2D(GL_TEXTURE_2D, 0, rt->color_internal_format, rt->size.x, rt->size.y, 0, rt->color_format, rt->color_type, nullptr); + texture = get_texture(rt->texture); + ERR_FAIL_COND(!texture); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - } + glGenTextures(1, &rt->color); + glBindTexture(texture_target, rt->color); + + if (use_multiview) { + glTexImage3D(texture_target, 0, rt->color_internal_format, rt->size.x, rt->size.y, rt->view_count, 0, rt->color_format, rt->color_type, nullptr); + } else { + glTexImage2D(texture_target, 0, rt->color_internal_format, rt->size.x, rt->size.y, 0, rt->color_format, rt->color_type, nullptr); + } - if (rt->view_count > 1 && config->multiview_supported) { + glTexParameteri(texture_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glTexParameteri(texture_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + } + if (use_multiview) { glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, rt->color, 0, 0, rt->view_count); } else { glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->color, 0); } - GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER); + // depth + if (rt->overridden.depth.is_valid()) { + texture = get_texture(rt->overridden.depth); + ERR_FAIL_COND(!texture); + + rt->depth = texture->tex_id; + } else { + glGenTextures(1, &rt->depth); + glBindTexture(texture_target, rt->depth); + if (use_multiview) { + glTexImage3D(texture_target, 0, GL_DEPTH_COMPONENT24, rt->size.x, rt->size.y, rt->view_count, 0, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, nullptr); + } else { + glTexImage2D(texture_target, 0, GL_DEPTH_COMPONENT24, rt->size.x, rt->size.y, 0, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, nullptr); + } + + glTexParameteri(texture_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glTexParameteri(texture_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + } + if (use_multiview) { + glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, rt->depth, 0, 0, rt->view_count); + } else { + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, rt->depth, 0); + } + + GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER); if (status != GL_FRAMEBUFFER_COMPLETE) { glDeleteFramebuffers(1, &rt->fbo); glDeleteTextures(1, &rt->color); @@ -1488,32 +1589,38 @@ void TextureStorage::_update_render_target(RenderTarget *rt) { rt->size.x = 0; rt->size.y = 0; rt->color = 0; - texture->tex_id = 0; - texture->active = false; + rt->depth = 0; + if (rt->overridden.color.is_null()) { + texture->tex_id = 0; + texture->active = false; + } WARN_PRINT("Could not create render target, status: " + get_framebuffer_error(status)); return; } - texture->format = rt->image_format; - texture->real_format = rt->image_format; - if (rt->view_count > 1 && config->multiview_supported) { - texture->type = Texture::TYPE_LAYERED; - texture->target = GL_TEXTURE_2D_ARRAY; - texture->layers = rt->view_count; + if (rt->overridden.color.is_valid()) { + texture->is_render_target = true; } else { - texture->type = Texture::TYPE_2D; - texture->target = GL_TEXTURE_2D; - texture->layers = 1; + texture->format = rt->image_format; + texture->real_format = rt->image_format; + texture->target = texture_target; + if (rt->view_count > 1 && config->multiview_supported) { + texture->type = Texture::TYPE_LAYERED; + texture->layers = rt->view_count; + } else { + texture->type = Texture::TYPE_2D; + texture->layers = 1; + } + texture->gl_format_cache = rt->color_format; + texture->gl_type_cache = GL_UNSIGNED_BYTE; + texture->gl_internal_format_cache = rt->color_internal_format; + texture->tex_id = rt->color; + texture->width = rt->size.x; + texture->alloc_width = rt->size.x; + texture->height = rt->size.y; + texture->alloc_height = rt->size.y; + texture->active = true; } - texture->gl_format_cache = rt->color_format; - texture->gl_type_cache = GL_UNSIGNED_BYTE; - texture->gl_internal_format_cache = rt->color_internal_format; - texture->tex_id = rt->color; - texture->width = rt->size.x; - texture->alloc_width = rt->size.x; - texture->height = rt->size.y; - texture->alloc_height = rt->size.y; - texture->active = true; } glClearColor(0, 0, 0, 0); @@ -1581,17 +1688,32 @@ void TextureStorage::_clear_render_target(RenderTarget *rt) { if (rt->fbo) { glDeleteFramebuffers(1, &rt->fbo); - glDeleteTextures(1, &rt->color); rt->fbo = 0; + } + + if (rt->overridden.color.is_null()) { + glDeleteTextures(1, &rt->color); rt->color = 0; } - Texture *tex = get_texture(rt->texture); - tex->alloc_height = 0; - tex->alloc_width = 0; - tex->width = 0; - tex->height = 0; - tex->active = false; + if (rt->overridden.depth.is_null()) { + glDeleteTextures(1, &rt->depth); + rt->depth = 0; + } + + if (rt->texture.is_valid()) { + Texture *tex = get_texture(rt->texture); + tex->alloc_height = 0; + tex->alloc_width = 0; + tex->width = 0; + tex->height = 0; + tex->active = false; + } + + if (rt->overridden.color.is_valid()) { + Texture *tex = get_texture(rt->overridden.color); + tex->is_render_target = false; + } if (rt->backbuffer_fbo != 0) { glDeleteFramebuffers(1, &rt->backbuffer_fbo); @@ -1599,6 +1721,16 @@ void TextureStorage::_clear_render_target(RenderTarget *rt) { rt->backbuffer = 0; rt->backbuffer_fbo = 0; } + _render_target_clear_sdf(rt); +} + +void TextureStorage::_clear_render_target_overridden_fbo_cache(RenderTarget *rt) { + // Dispose of the cached fbo's and the allocated textures + for (KeyValue<uint32_t, RenderTarget::RTOverridden::FBOCacheEntry> &E : rt->overridden.fbo_cache) { + glDeleteTextures(E.value.allocated_textures.size(), E.value.allocated_textures.ptr()); + glDeleteFramebuffers(1, &E.value.fbo); + } + rt->overridden.fbo_cache.clear(); } RID TextureStorage::render_target_create() { @@ -1619,11 +1751,14 @@ RID TextureStorage::render_target_create() { void TextureStorage::render_target_free(RID p_rid) { RenderTarget *rt = render_target_owner.get_or_null(p_rid); _clear_render_target(rt); + _clear_render_target_overridden_fbo_cache(rt); Texture *t = get_texture(rt->texture); if (t) { t->is_render_target = false; - texture_free(rt->texture); + if (rt->overridden.color.is_null()) { + texture_free(rt->texture); + } //memdelete(t); } render_target_owner.free(p_rid); @@ -1650,6 +1785,9 @@ void TextureStorage::render_target_set_size(RID p_render_target, int p_width, in if (p_width == rt->size.x && p_height == rt->size.y && p_view_count == rt->view_count) { return; } + if (rt->overridden.color.is_valid()) { + return; + } _clear_render_target(rt); @@ -1667,10 +1805,91 @@ Size2i TextureStorage::render_target_get_size(RID p_render_target) const { return rt->size; } +void TextureStorage::render_target_set_override(RID p_render_target, RID p_color_texture, RID p_depth_texture, RID p_velocity_texture) { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND(!rt); + ERR_FAIL_COND(rt->direct_to_screen); + + rt->overridden.velocity = p_velocity_texture; + + if (rt->overridden.color == p_color_texture && rt->overridden.depth == p_depth_texture) { + return; + } + + if (p_color_texture.is_null() && p_depth_texture.is_null()) { + _clear_render_target(rt); + rt->overridden.is_overridden = false; + rt->overridden.color = RID(); + rt->overridden.depth = RID(); + rt->size = Size2i(); + _clear_render_target_overridden_fbo_cache(rt); + return; + } + + if (!rt->overridden.is_overridden) { + _clear_render_target(rt); + } + + rt->overridden.color = p_color_texture; + rt->overridden.depth = p_depth_texture; + rt->overridden.is_overridden = true; + + uint32_t hash_key = hash_murmur3_one_64(p_color_texture.get_id()); + hash_key = hash_murmur3_one_64(p_depth_texture.get_id(), hash_key); + hash_key = hash_fmix32(hash_key); + + RBMap<uint32_t, RenderTarget::RTOverridden::FBOCacheEntry>::Element *cache; + if ((cache = rt->overridden.fbo_cache.find(hash_key)) != nullptr) { + rt->fbo = cache->get().fbo; + rt->size = cache->get().size; + rt->texture = p_color_texture; + return; + } + + _update_render_target(rt); + + RenderTarget::RTOverridden::FBOCacheEntry new_entry; + new_entry.fbo = rt->fbo; + new_entry.size = rt->size; + // Keep track of any textures we had to allocate because they weren't overridden. + if (p_color_texture.is_null()) { + new_entry.allocated_textures.push_back(rt->color); + } + if (p_depth_texture.is_null()) { + new_entry.allocated_textures.push_back(rt->depth); + } + rt->overridden.fbo_cache.insert(hash_key, new_entry); +} + +RID TextureStorage::render_target_get_override_color(RID p_render_target) const { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, RID()); + + return rt->overridden.color; +} + +RID TextureStorage::render_target_get_override_depth(RID p_render_target) const { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, RID()); + + return rt->overridden.depth; +} + +RID TextureStorage::render_target_get_override_velocity(RID p_render_target) const { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, RID()); + + return rt->overridden.velocity; +} + RID TextureStorage::render_target_get_texture(RID p_render_target) { RenderTarget *rt = render_target_owner.get_or_null(p_render_target); ERR_FAIL_COND_V(!rt, RID()); + if (rt->overridden.color.is_valid()) { + return rt->overridden.color; + } + return rt->texture; } @@ -1680,8 +1899,10 @@ void TextureStorage::render_target_set_transparent(RID p_render_target, bool p_t rt->is_transparent = p_transparent; - _clear_render_target(rt); - _update_render_target(rt); + if (rt->overridden.color.is_null()) { + _clear_render_target(rt); + _update_render_target(rt); + } } bool TextureStorage::render_target_get_transparent(RID p_render_target) const { @@ -1702,6 +1923,11 @@ void TextureStorage::render_target_set_direct_to_screen(RID p_render_target, boo // those functions change how they operate depending on the value of DIRECT_TO_SCREEN _clear_render_target(rt); rt->direct_to_screen = p_direct_to_screen; + if (rt->direct_to_screen) { + rt->overridden.color = RID(); + rt->overridden.depth = RID(); + rt->overridden.velocity = RID(); + } _update_render_target(rt); } @@ -1734,6 +1960,7 @@ void TextureStorage::render_target_set_msaa(RID p_render_target, RS::ViewportMSA } WARN_PRINT("2D MSAA is not yet supported for GLES3."); + _clear_render_target(rt); rt->msaa = p_msaa; _update_render_target(rt); @@ -1784,13 +2011,271 @@ void TextureStorage::render_target_do_clear_request(RID p_render_target) { } void TextureStorage::render_target_set_sdf_size_and_scale(RID p_render_target, RS::ViewportSDFOversize p_size, RS::ViewportSDFScale p_scale) { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND(!rt); + if (rt->sdf_oversize == p_size && rt->sdf_scale == p_scale) { + return; + } + + rt->sdf_oversize = p_size; + rt->sdf_scale = p_scale; + + _render_target_clear_sdf(rt); +} + +Rect2i TextureStorage::_render_target_get_sdf_rect(const RenderTarget *rt) const { + Size2i margin; + int scale; + switch (rt->sdf_oversize) { + case RS::VIEWPORT_SDF_OVERSIZE_100_PERCENT: { + scale = 100; + } break; + case RS::VIEWPORT_SDF_OVERSIZE_120_PERCENT: { + scale = 120; + } break; + case RS::VIEWPORT_SDF_OVERSIZE_150_PERCENT: { + scale = 150; + } break; + case RS::VIEWPORT_SDF_OVERSIZE_200_PERCENT: { + scale = 200; + } break; + default: { + } + } + + margin = (rt->size * scale / 100) - rt->size; + + Rect2i r(Vector2i(), rt->size); + r.position -= margin; + r.size += margin * 2; + + return r; } Rect2i TextureStorage::render_target_get_sdf_rect(RID p_render_target) const { - return Rect2i(); + const RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, Rect2i()); + + return _render_target_get_sdf_rect(rt); } void TextureStorage::render_target_mark_sdf_enabled(RID p_render_target, bool p_enabled) { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND(!rt); + + rt->sdf_enabled = p_enabled; +} + +bool TextureStorage::render_target_is_sdf_enabled(RID p_render_target) const { + const RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, false); + + return rt->sdf_enabled; +} + +GLuint TextureStorage::render_target_get_sdf_texture(RID p_render_target) { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, 0); + if (rt->sdf_texture_read == 0) { + Texture *texture = texture_owner.get_or_null(default_gl_textures[DEFAULT_GL_TEXTURE_BLACK]); + return texture->tex_id; + } + + return rt->sdf_texture_read; +} + +void TextureStorage::_render_target_allocate_sdf(RenderTarget *rt) { + ERR_FAIL_COND(rt->sdf_texture_write_fb != 0); + + Size2i size = _render_target_get_sdf_rect(rt).size; + + glGenTextures(1, &rt->sdf_texture_write); + glActiveTexture(GL_TEXTURE0); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_write); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, size.width, size.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + + glGenFramebuffers(1, &rt->sdf_texture_write_fb); + glBindFramebuffer(GL_FRAMEBUFFER, rt->sdf_texture_write_fb); + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->sdf_texture_write, 0); + + int scale; + switch (rt->sdf_scale) { + case RS::VIEWPORT_SDF_SCALE_100_PERCENT: { + scale = 100; + } break; + case RS::VIEWPORT_SDF_SCALE_50_PERCENT: { + scale = 50; + } break; + case RS::VIEWPORT_SDF_SCALE_25_PERCENT: { + scale = 25; + } break; + default: { + scale = 100; + } break; + } + + rt->process_size = size * scale / 100; + rt->process_size.x = MAX(rt->process_size.x, 1); + rt->process_size.y = MAX(rt->process_size.y, 1); + + glGenTextures(2, rt->sdf_texture_process); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_process[0]); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16I, rt->process_size.width, rt->process_size.height, 0, GL_RG_INTEGER, GL_SHORT, nullptr); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_process[1]); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16I, rt->process_size.width, rt->process_size.height, 0, GL_RG_INTEGER, GL_SHORT, nullptr); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + + glGenTextures(1, &rt->sdf_texture_read); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_read); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, rt->process_size.width, rt->process_size.height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); +} + +void TextureStorage::_render_target_clear_sdf(RenderTarget *rt) { + if (rt->sdf_texture_write_fb != 0) { + glDeleteTextures(1, &rt->sdf_texture_read); + glDeleteTextures(1, &rt->sdf_texture_write); + glDeleteTextures(2, rt->sdf_texture_process); + glDeleteFramebuffers(1, &rt->sdf_texture_write_fb); + rt->sdf_texture_read = 0; + rt->sdf_texture_write = 0; + rt->sdf_texture_process[0] = 0; + rt->sdf_texture_process[1] = 0; + rt->sdf_texture_write_fb = 0; + } +} + +GLuint TextureStorage::render_target_get_sdf_framebuffer(RID p_render_target) { + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND_V(!rt, 0); + + if (rt->sdf_texture_write_fb == 0) { + _render_target_allocate_sdf(rt); + } + + return rt->sdf_texture_write_fb; +} +void TextureStorage::render_target_sdf_process(RID p_render_target) { + CopyEffects *copy_effects = CopyEffects::get_singleton(); + + RenderTarget *rt = render_target_owner.get_or_null(p_render_target); + ERR_FAIL_COND(!rt); + ERR_FAIL_COND(rt->sdf_texture_write_fb == 0); + + Rect2i r = _render_target_get_sdf_rect(rt); + + Size2i size = r.size; + int32_t shift = 0; + + bool shrink = false; + + switch (rt->sdf_scale) { + case RS::VIEWPORT_SDF_SCALE_50_PERCENT: { + size[0] >>= 1; + size[1] >>= 1; + shift = 1; + shrink = true; + } break; + case RS::VIEWPORT_SDF_SCALE_25_PERCENT: { + size[0] >>= 2; + size[1] >>= 2; + shift = 2; + shrink = true; + } break; + default: { + }; + } + + GLuint temp_fb; + glGenFramebuffers(1, &temp_fb); + glBindFramebuffer(GL_FRAMEBUFFER, temp_fb); + + // Load + CanvasSdfShaderGLES3::ShaderVariant variant = shrink ? CanvasSdfShaderGLES3::MODE_LOAD_SHRINK : CanvasSdfShaderGLES3::MODE_LOAD; + sdf_shader.shader.version_bind_shader(sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::BASE_SIZE, r.size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SIZE, size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::STRIDE, 0, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SHIFT, shift, sdf_shader.shader_version, variant); + + glActiveTexture(GL_TEXTURE0); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_write); + + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->sdf_texture_process[0], 0); + glViewport(0, 0, size.width, size.height); + glEnable(GL_SCISSOR_TEST); + glScissor(0, 0, size.width, size.height); + + copy_effects->draw_screen_triangle(); + + // Process + + int stride = nearest_power_of_2_templated(MAX(size.width, size.height) / 2); + + variant = CanvasSdfShaderGLES3::MODE_PROCESS; + sdf_shader.shader.version_bind_shader(sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::BASE_SIZE, r.size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SIZE, size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::STRIDE, stride, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SHIFT, shift, sdf_shader.shader_version, variant); + + bool swap = false; + + //jumpflood + while (stride > 0) { + glBindTexture(GL_TEXTURE_2D, 0); + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->sdf_texture_process[swap ? 0 : 1], 0); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_process[swap ? 1 : 0]); + + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::STRIDE, stride, sdf_shader.shader_version, variant); + + copy_effects->draw_screen_triangle(); + + stride /= 2; + swap = !swap; + } + + // Store + variant = shrink ? CanvasSdfShaderGLES3::MODE_STORE_SHRINK : CanvasSdfShaderGLES3::MODE_STORE; + sdf_shader.shader.version_bind_shader(sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::BASE_SIZE, r.size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SIZE, size, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::STRIDE, stride, sdf_shader.shader_version, variant); + sdf_shader.shader.version_set_uniform(CanvasSdfShaderGLES3::SHIFT, shift, sdf_shader.shader_version, variant); + + glBindTexture(GL_TEXTURE_2D, 0); + glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->sdf_texture_read, 0); + glBindTexture(GL_TEXTURE_2D, rt->sdf_texture_process[swap ? 1 : 0]); + + copy_effects->draw_screen_triangle(); + + glBindTexture(GL_TEXTURE_2D, 0); + glBindFramebuffer(GL_FRAMEBUFFER, system_fbo); + glDeleteFramebuffers(1, &temp_fb); + glDisable(GL_SCISSOR_TEST); } void TextureStorage::render_target_copy_to_back_buffer(RID p_render_target, const Rect2i &p_region, bool p_gen_mipmaps) { diff --git a/drivers/gles3/storage/texture_storage.h b/drivers/gles3/storage/texture_storage.h index dbe39428ac..7714e72f62 100644 --- a/drivers/gles3/storage/texture_storage.h +++ b/drivers/gles3/storage/texture_storage.h @@ -39,6 +39,8 @@ #include "servers/rendering/renderer_compositor.h" #include "servers/rendering/storage/texture_storage.h" +#include "../shaders/canvas_sdf.glsl.gen.h" + // This must come first to avoid windows.h mess #include "platform_config.h" #ifndef OPENGL_INCLUDE_H @@ -84,18 +86,8 @@ namespace GLES3 { #define _GL_TEXTURE_EXTERNAL_OES 0x8D65 -#ifdef GLES_OVER_GL -#define _GL_HALF_FLOAT_OES 0x140B -#else -#define _GL_HALF_FLOAT_OES 0x8D61 -#endif - #define _EXT_TEXTURE_CUBE_MAP_SEAMLESS 0x884F -#define _RED_OES 0x1903 - -#define _DEPTH_COMPONENT24_OES 0x81A6 - #ifndef GLES_OVER_GL #define glClearDepth glClearDepthf #endif //!GLES_OVER_GL @@ -128,23 +120,13 @@ struct CanvasTexture { RS::CanvasItemTextureRepeat texture_repeat = RS::CANVAS_ITEM_TEXTURE_REPEAT_DEFAULT; }; -/* CANVAS SHADOW */ - -struct CanvasLightShadow { - RID self; - int size; - int height; - GLuint fbo; - GLuint depth; - GLuint distance; //for older devices -}; - struct RenderTarget; struct Texture { RID self; bool is_proxy = false; + bool is_external = false; bool is_render_target = false; RID proxy_to = RID(); @@ -206,6 +188,7 @@ struct Texture { void copy_from(const Texture &o) { proxy_to = o.proxy_to; is_proxy = o.is_proxy; + is_external = o.is_external; width = o.width; height = o.height; alloc_width = o.alloc_width; @@ -329,6 +312,7 @@ struct RenderTarget { RID self; GLuint fbo = 0; GLuint color = 0; + GLuint depth = 0; GLuint backbuffer_fbo = 0; GLuint backbuffer = 0; @@ -337,12 +321,35 @@ struct RenderTarget { GLuint color_type = GL_UNSIGNED_BYTE; Image::Format image_format = Image::FORMAT_RGBA8; + GLuint sdf_texture_write = 0; + GLuint sdf_texture_write_fb = 0; + GLuint sdf_texture_process[2] = { 0, 0 }; + GLuint sdf_texture_read = 0; + RS::ViewportSDFOversize sdf_oversize = RS::VIEWPORT_SDF_OVERSIZE_120_PERCENT; + RS::ViewportSDFScale sdf_scale = RS::VIEWPORT_SDF_SCALE_50_PERCENT; + Size2i process_size; + bool sdf_enabled = false; + bool is_transparent = false; bool direct_to_screen = false; bool used_in_frame = false; RS::ViewportMSAA msaa = RS::VIEWPORT_MSAA_DISABLED; + struct RTOverridden { + bool is_overridden = false; + RID color; + RID depth; + RID velocity; + + struct FBOCacheEntry { + GLuint fbo; + Size2i size; + Vector<GLuint> allocated_textures; + }; + RBMap<uint32_t, FBOCacheEntry> fbo_cache; + } overridden; + RID texture; Color clear_color = Color(1, 1, 1, 1); @@ -362,10 +369,6 @@ private: RID_Owner<CanvasTexture, true> canvas_texture_owner; - /* CANVAS SHADOW */ - - RID_PtrOwner<CanvasLightShadow> canvas_light_shadow_owner; - /* Texture API */ mutable RID_Owner<Texture> texture_owner; @@ -409,8 +412,17 @@ private: mutable RID_Owner<RenderTarget> render_target_owner; void _clear_render_target(RenderTarget *rt); + void _clear_render_target_overridden_fbo_cache(RenderTarget *rt); void _update_render_target(RenderTarget *rt); void _create_render_target_backbuffer(RenderTarget *rt); + void _render_target_allocate_sdf(RenderTarget *rt); + void _render_target_clear_sdf(RenderTarget *rt); + Rect2i _render_target_get_sdf_rect(const RenderTarget *rt) const; + + struct RenderTargetSDF { + CanvasSdfShaderGLES3 shader; + RID shader_version; + } sdf_shader; public: static TextureStorage *get_singleton(); @@ -437,10 +449,6 @@ public: virtual void canvas_texture_set_texture_filter(RID p_item, RS::CanvasItemTextureFilter p_filter) override; virtual void canvas_texture_set_texture_repeat(RID p_item, RS::CanvasItemTextureRepeat p_repeat) override; - /* CANVAS SHADOW */ - - RID canvas_light_shadow_buffer_create(int p_width); - /* Texture API */ Texture *get_texture(RID p_rid) { @@ -464,6 +472,8 @@ public: virtual void texture_3d_initialize(RID p_texture, Image::Format, int p_width, int p_height, int p_depth, bool p_mipmaps, const Vector<Ref<Image>> &p_data) override; virtual void texture_proxy_initialize(RID p_texture, RID p_base) override; //all slices, then all the mipmaps, must be coherent + RID texture_create_external(Texture::Type p_type, Image::Format p_format, unsigned int p_image, int p_width, int p_height, int p_depth, int p_layers, RS::TextureLayeredType p_layered_type = RS::TEXTURE_LAYERED_2D_ARRAY); + virtual void texture_2d_update(RID p_texture, const Ref<Image> &p_image, int p_layer = 0) override; virtual void texture_3d_update(RID p_texture, const Vector<Ref<Image>> &p_data) override{}; virtual void texture_proxy_update(RID p_proxy, RID p_base) override; @@ -494,6 +504,8 @@ public: virtual Size2 texture_size_with_proxy(RID p_proxy) override; + virtual RID texture_get_rd_texture_rid(RID p_texture, bool p_srgb = false) const override; + void texture_set_data(RID p_texture, const Ref<Image> &p_image, int p_layer = 0); void texture_set_data_partial(RID p_texture, const Ref<Image> &p_image, int src_x, int src_y, int src_w, int src_h, int dst_x, int dst_y, int p_dst_mip, int p_layer = 0); //Ref<Image> texture_get_data(RID p_texture, int p_layer = 0) const; @@ -586,9 +598,13 @@ public: void render_target_disable_clear_request(RID p_render_target) override; void render_target_do_clear_request(RID p_render_target) override; - void render_target_set_sdf_size_and_scale(RID p_render_target, RS::ViewportSDFOversize p_size, RS::ViewportSDFScale p_scale) override; - Rect2i render_target_get_sdf_rect(RID p_render_target) const override; - void render_target_mark_sdf_enabled(RID p_render_target, bool p_enabled) override; + virtual void render_target_set_sdf_size_and_scale(RID p_render_target, RS::ViewportSDFOversize p_size, RS::ViewportSDFScale p_scale) override; + virtual Rect2i render_target_get_sdf_rect(RID p_render_target) const override; + GLuint render_target_get_sdf_texture(RID p_render_target); + GLuint render_target_get_sdf_framebuffer(RID p_render_target); + void render_target_sdf_process(RID p_render_target); + virtual void render_target_mark_sdf_enabled(RID p_render_target, bool p_enabled) override; + bool render_target_is_sdf_enabled(RID p_render_target) const; void render_target_copy_to_back_buffer(RID p_render_target, const Rect2i &p_region, bool p_gen_mipmaps); void render_target_clear_back_buffer(RID p_render_target, const Rect2i &p_region, const Color &p_color); @@ -599,12 +615,10 @@ public: virtual void render_target_set_vrs_texture(RID p_render_target, RID p_texture) override {} virtual RID render_target_get_vrs_texture(RID p_render_target) const override { return RID(); } - virtual void render_target_set_override_color(RID p_render_target, RID p_texture) override {} - virtual RID render_target_get_override_color(RID p_render_target) const override { return RID(); } - virtual void render_target_set_override_depth(RID p_render_target, RID p_texture) override {} - virtual RID render_target_get_override_depth(RID p_render_target) const override { return RID(); } - virtual void render_target_set_override_velocity(RID p_render_target, RID p_texture) override {} - virtual RID render_target_get_override_velocity(RID p_render_target) const override { return RID(); } + virtual void render_target_set_override(RID p_render_target, RID p_color_texture, RID p_depth_texture, RID p_velocity_texture) override; + virtual RID render_target_get_override_color(RID p_render_target) const override; + virtual RID render_target_get_override_depth(RID p_render_target) const override; + virtual RID render_target_get_override_velocity(RID p_render_target) const override; virtual RID render_target_get_texture(RID p_render_target) override; diff --git a/drivers/gles3/storage/utilities.cpp b/drivers/gles3/storage/utilities.cpp index 0a04f0d4d0..8e7e218bb9 100644 --- a/drivers/gles3/storage/utilities.cpp +++ b/drivers/gles3/storage/utilities.cpp @@ -38,20 +38,44 @@ #include "particles_storage.h" #include "texture_storage.h" +#include "servers/rendering/rendering_server_globals.h" + using namespace GLES3; Utilities *Utilities::singleton = nullptr; Utilities::Utilities() { singleton = this; + frame = 0; + for (int i = 0; i < FRAME_COUNT; i++) { + frames[i].index = 0; + glGenQueries(max_timestamp_query_elements, frames[i].queries); + + frames[i].timestamp_names.resize(max_timestamp_query_elements); + frames[i].timestamp_cpu_values.resize(max_timestamp_query_elements); + frames[i].timestamp_count = 0; + + frames[i].timestamp_result_names.resize(max_timestamp_query_elements); + frames[i].timestamp_cpu_result_values.resize(max_timestamp_query_elements); + frames[i].timestamp_result_values.resize(max_timestamp_query_elements); + frames[i].timestamp_result_count = 0; + } } Utilities::~Utilities() { singleton = nullptr; + for (int i = 0; i < FRAME_COUNT; i++) { + glDeleteQueries(max_timestamp_query_elements, frames[i].queries); + } } Vector<uint8_t> Utilities::buffer_get_data(GLenum p_target, GLuint p_buffer, uint32_t p_buffer_size) { Vector<uint8_t> ret; + + if (p_buffer_size == 0) { + return ret; + } + ret.resize(p_buffer_size); glBindBuffer(p_target, p_buffer); @@ -213,87 +237,69 @@ void Utilities::visibility_notifier_call(RID p_notifier, bool p_enter, bool p_de /* TIMING */ -//void Utilities::render_info_begin_capture() { -// info.snap = info.render; -//} - -//void Utilities::render_info_end_capture() { -// info.snap.object_count = info.render.object_count - info.snap.object_count; -// info.snap.draw_call_count = info.render.draw_call_count - info.snap.draw_call_count; -// info.snap.material_switch_count = info.render.material_switch_count - info.snap.material_switch_count; -// info.snap.surface_switch_count = info.render.surface_switch_count - info.snap.surface_switch_count; -// info.snap.shader_rebind_count = info.render.shader_rebind_count - info.snap.shader_rebind_count; -// info.snap.vertices_count = info.render.vertices_count - info.snap.vertices_count; -// info.snap._2d_item_count = info.render._2d_item_count - info.snap._2d_item_count; -// info.snap._2d_draw_call_count = info.render._2d_draw_call_count - info.snap._2d_draw_call_count; -//} - -//int Utilities::get_captured_render_info(RS::RenderInfo p_info) { -// switch (p_info) { -// case RS::INFO_OBJECTS_IN_FRAME: { -// return info.snap.object_count; -// } break; -// case RS::INFO_VERTICES_IN_FRAME: { -// return info.snap.vertices_count; -// } break; -// case RS::INFO_MATERIAL_CHANGES_IN_FRAME: { -// return info.snap.material_switch_count; -// } break; -// case RS::INFO_SHADER_CHANGES_IN_FRAME: { -// return info.snap.shader_rebind_count; -// } break; -// case RS::INFO_SURFACE_CHANGES_IN_FRAME: { -// return info.snap.surface_switch_count; -// } break; -// case RS::INFO_DRAW_CALLS_IN_FRAME: { -// return info.snap.draw_call_count; -// } break; -// /* -// case RS::INFO_2D_ITEMS_IN_FRAME: { -// return info.snap._2d_item_count; -// } break; -// case RS::INFO_2D_DRAW_CALLS_IN_FRAME: { -// return info.snap._2d_draw_call_count; -// } break; -// */ -// default: { -// return get_render_info(p_info); -// } -// } -//} - -//int Utilities::get_render_info(RS::RenderInfo p_info) { -// switch (p_info) { -// case RS::INFO_OBJECTS_IN_FRAME: -// return info.render_final.object_count; -// case RS::INFO_VERTICES_IN_FRAME: -// return info.render_final.vertices_count; -// case RS::INFO_MATERIAL_CHANGES_IN_FRAME: -// return info.render_final.material_switch_count; -// case RS::INFO_SHADER_CHANGES_IN_FRAME: -// return info.render_final.shader_rebind_count; -// case RS::INFO_SURFACE_CHANGES_IN_FRAME: -// return info.render_final.surface_switch_count; -// case RS::INFO_DRAW_CALLS_IN_FRAME: -// return info.render_final.draw_call_count; -// /* -// case RS::INFO_2D_ITEMS_IN_FRAME: -// return info.render_final._2d_item_count; -// case RS::INFO_2D_DRAW_CALLS_IN_FRAME: -// return info.render_final._2d_draw_call_count; -//*/ -// case RS::INFO_USAGE_VIDEO_MEM_TOTAL: -// return 0; //no idea -// case RS::INFO_VIDEO_MEM_USED: -// return info.vertex_mem + info.texture_mem; -// case RS::INFO_TEXTURE_MEM_USED: -// return info.texture_mem; -// case RS::INFO_VERTEX_MEM_USED: -// return info.vertex_mem; -// default: -// return 0; //no idea either -// } -//} +void Utilities::capture_timestamps_begin() { + capture_timestamp("Frame Begin"); +} + +void Utilities::capture_timestamp(const String &p_name) { + ERR_FAIL_COND(frames[frame].timestamp_count >= max_timestamp_query_elements); + +#ifdef GLES_OVER_GL + glQueryCounter(frames[frame].queries[frames[frame].timestamp_count], GL_TIMESTAMP); +#endif + + frames[frame].timestamp_names[frames[frame].timestamp_count] = p_name; + frames[frame].timestamp_cpu_values[frames[frame].timestamp_count] = OS::get_singleton()->get_ticks_usec(); + frames[frame].timestamp_count++; +} + +void Utilities::_capture_timestamps_begin() { + // frame is incremented at the end of the frame so this gives us the queries for frame - 2. By then they should be ready. + if (frames[frame].timestamp_count) { +#ifdef GLES_OVER_GL + for (uint32_t i = 0; i < frames[frame].timestamp_count; i++) { + uint64_t temp = 0; + glGetQueryObjectui64v(frames[frame].queries[i], GL_QUERY_RESULT, &temp); + frames[frame].timestamp_result_values[i] = temp; + } +#endif + SWAP(frames[frame].timestamp_names, frames[frame].timestamp_result_names); + SWAP(frames[frame].timestamp_cpu_values, frames[frame].timestamp_cpu_result_values); + } + + frames[frame].timestamp_result_count = frames[frame].timestamp_count; + frames[frame].timestamp_count = 0; + frames[frame].index = Engine::get_singleton()->get_frames_drawn(); + capture_timestamp("Internal Begin"); +} + +void Utilities::capture_timestamps_end() { + capture_timestamp("Internal End"); + frame = (frame + 1) % FRAME_COUNT; +} + +uint32_t Utilities::get_captured_timestamps_count() const { + return frames[frame].timestamp_result_count; +} + +uint64_t Utilities::get_captured_timestamps_frame() const { + return frames[frame].index; +} + +uint64_t Utilities::get_captured_timestamp_gpu_time(uint32_t p_index) const { + ERR_FAIL_UNSIGNED_INDEX_V(p_index, frames[frame].timestamp_result_count, 0); + return frames[frame].timestamp_result_values[p_index]; +} + +uint64_t Utilities::get_captured_timestamp_cpu_time(uint32_t p_index) const { + ERR_FAIL_UNSIGNED_INDEX_V(p_index, frames[frame].timestamp_result_count, 0); + return frames[frame].timestamp_cpu_result_values[p_index]; +} + +String Utilities::get_captured_timestamp_name(uint32_t p_index) const { + ERR_FAIL_UNSIGNED_INDEX_V(p_index, frames[frame].timestamp_result_count, String()); + return frames[frame].timestamp_result_names[p_index]; +} /* MISC */ @@ -362,7 +368,7 @@ Size2i Utilities::get_maximum_viewport_size() const { return Size2i(); } - return Size2i(config->max_viewport_size, config->max_viewport_size); + return Size2i(config->max_viewport_size[0], config->max_viewport_size[1]); } #endif // GLES3_ENABLED diff --git a/drivers/gles3/storage/utilities.h b/drivers/gles3/storage/utilities.h index 936ac35cc9..55a875958e 100644 --- a/drivers/gles3/storage/utilities.h +++ b/drivers/gles3/storage/utilities.h @@ -79,62 +79,35 @@ public: /* TIMING */ - struct Info { - uint64_t texture_mem = 0; - uint64_t vertex_mem = 0; - - struct Render { - uint32_t object_count; - uint32_t draw_call_count; - uint32_t material_switch_count; - uint32_t surface_switch_count; - uint32_t shader_rebind_count; - uint32_t vertices_count; - uint32_t _2d_item_count; - uint32_t _2d_draw_call_count; - - void reset() { - object_count = 0; - draw_call_count = 0; - material_switch_count = 0; - surface_switch_count = 0; - shader_rebind_count = 0; - vertices_count = 0; - _2d_item_count = 0; - _2d_draw_call_count = 0; - } - } render, render_final, snap; - - Info() { - render.reset(); - render_final.reset(); - } - - } info; - - virtual void capture_timestamps_begin() override {} - virtual void capture_timestamp(const String &p_name) override {} - virtual uint32_t get_captured_timestamps_count() const override { - return 0; - } - virtual uint64_t get_captured_timestamps_frame() const override { - return 0; - } - virtual uint64_t get_captured_timestamp_gpu_time(uint32_t p_index) const override { - return 0; - } - virtual uint64_t get_captured_timestamp_cpu_time(uint32_t p_index) const override { - return 0; - } - virtual String get_captured_timestamp_name(uint32_t p_index) const override { - return String(); - } - - // void render_info_begin_capture() override; - // void render_info_end_capture() override; - // int get_captured_render_info(RS::RenderInfo p_info) override; - - // int get_render_info(RS::RenderInfo p_info) override; +#define MAX_QUERIES 256 +#define FRAME_COUNT 3 + + struct Frame { + GLuint queries[MAX_QUERIES]; + TightLocalVector<String> timestamp_names; + TightLocalVector<uint64_t> timestamp_cpu_values; + uint32_t timestamp_count = 0; + TightLocalVector<String> timestamp_result_names; + TightLocalVector<uint64_t> timestamp_cpu_result_values; + TightLocalVector<uint64_t> timestamp_result_values; + uint32_t timestamp_result_count = 0; + uint64_t index = 0; + }; + + const uint32_t max_timestamp_query_elements = MAX_QUERIES; + + Frame frames[FRAME_COUNT]; // Frames for capturing timestamps. We use 3 so we don't need to wait for commands to complete + uint32_t frame = 0; + + virtual void capture_timestamps_begin() override; + virtual void capture_timestamp(const String &p_name) override; + virtual uint32_t get_captured_timestamps_count() const override; + virtual uint64_t get_captured_timestamps_frame() const override; + virtual uint64_t get_captured_timestamp_gpu_time(uint32_t p_index) const override; + virtual uint64_t get_captured_timestamp_cpu_time(uint32_t p_index) const override; + virtual String get_captured_timestamp_name(uint32_t p_index) const override; + void _capture_timestamps_begin(); + void capture_timestamps_end(); /* MISC */ diff --git a/drivers/unix/os_unix.cpp b/drivers/unix/os_unix.cpp index fc06291a3a..161706489f 100644 --- a/drivers/unix/os_unix.cpp +++ b/drivers/unix/os_unix.cpp @@ -505,11 +505,11 @@ bool OS_Unix::set_environment(const String &p_var, const String &p_value) const } String OS_Unix::get_user_data_dir() const { - String appname = get_safe_dir_name(ProjectSettings::get_singleton()->get("application/config/name")); + String appname = get_safe_dir_name(GLOBAL_GET("application/config/name")); if (!appname.is_empty()) { - bool use_custom_dir = ProjectSettings::get_singleton()->get("application/config/use_custom_user_dir"); + bool use_custom_dir = GLOBAL_GET("application/config/use_custom_user_dir"); if (use_custom_dir) { - String custom_dir = get_safe_dir_name(ProjectSettings::get_singleton()->get("application/config/custom_user_dir_name"), true); + String custom_dir = get_safe_dir_name(GLOBAL_GET("application/config/custom_user_dir_name"), true); if (custom_dir.is_empty()) { custom_dir = appname; } diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp index de29363556..0ff4ca31e6 100644 --- a/drivers/vulkan/rendering_device_vulkan.cpp +++ b/drivers/vulkan/rendering_device_vulkan.cpp @@ -2160,14 +2160,35 @@ RID RenderingDeviceVulkan::texture_create_from_extension(TextureType p_type, Dat texture.height = p_height; texture.depth = p_depth; texture.layers = p_layers; - texture.mipmaps = 0; // Maybe make this settable too? + texture.mipmaps = 1; texture.usage_flags = p_flags; texture.base_mipmap = 0; texture.base_layer = 0; texture.allowed_shared_formats.push_back(RD::DATA_FORMAT_R8G8B8A8_UNORM); texture.allowed_shared_formats.push_back(RD::DATA_FORMAT_R8G8B8A8_SRGB); - // Do we need to do something with texture.layout? + // Set base layout based on usage priority. + + if (texture.usage_flags & TEXTURE_USAGE_SAMPLING_BIT) { + // First priority, readable. + texture.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_STORAGE_BIT) { + // Second priority, storage. + + texture.layout = VK_IMAGE_LAYOUT_GENERAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + // Third priority, color or depth. + + texture.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + texture.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; + + } else { + texture.layout = VK_IMAGE_LAYOUT_GENERAL; + } if (texture.usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { texture.read_aspect_mask = VK_IMAGE_ASPECT_DEPTH_BIT; @@ -3401,6 +3422,16 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF LocalVector<int32_t> attachment_last_pass; attachment_last_pass.resize(p_attachments.size()); + if (p_view_count > 1) { + const VulkanContext::MultiviewCapabilities capabilities = context->get_multiview_capabilities(); + + // This only works with multiview! + ERR_FAIL_COND_V_MSG(!capabilities.is_supported, VK_NULL_HANDLE, "Multiview not supported"); + + // Make sure we limit this to the number of views we support. + ERR_FAIL_COND_V_MSG(p_view_count > capabilities.max_view_count, VK_NULL_HANDLE, "Hardware does not support requested number of views for Multiview render pass"); + } + // These are only used if we use multiview but we need to define them in scope. const uint32_t view_mask = (1 << p_view_count) - 1; const uint32_t correlation_mask = (1 << p_view_count) - 1; @@ -3701,7 +3732,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; color_references.push_back(reference); } @@ -3723,7 +3754,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; // TODO: We need to set this here, possibly VK_IMAGE_ASPECT_COLOR_BIT? + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; input_references.push_back(reference); } @@ -3752,7 +3783,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; // VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; resolve_references.push_back(reference); } @@ -3767,7 +3798,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer depth format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); depth_stencil_reference.attachment = attachment_remap[attachment]; depth_stencil_reference.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; - depth_stencil_reference.aspectMask = 0; + depth_stencil_reference.aspectMask = VK_IMAGE_ASPECT_NONE; attachment_last_pass[attachment] = i; if (is_multisample_first) { @@ -3793,7 +3824,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF vrs_reference.pNext = nullptr; vrs_reference.attachment = attachment_remap[attachment]; vrs_reference.layout = VK_IMAGE_LAYOUT_FRAGMENT_SHADING_RATE_ATTACHMENT_OPTIMAL_KHR; - vrs_reference.aspectMask = 0; + vrs_reference.aspectMask = VK_IMAGE_ASPECT_NONE; Size2i texel_size = context->get_vrs_capabilities().max_texel_size; @@ -3934,16 +3965,9 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF Vector<uint32_t> view_masks; VkRenderPassMultiviewCreateInfo render_pass_multiview_create_info; - if (p_view_count > 1) { - // This may no longer be needed with the new settings already including this. - - const VulkanContext::MultiviewCapabilities capabilities = context->get_multiview_capabilities(); - - // For now this only works with multiview! - ERR_FAIL_COND_V_MSG(!capabilities.is_supported, VK_NULL_HANDLE, "Multiview not supported"); - - // Make sure we limit this to the number of views we support. - ERR_FAIL_COND_V_MSG(p_view_count > capabilities.max_view_count, VK_NULL_HANDLE, "Hardware does not support requested number of views for Multiview render pass"); + if ((p_view_count > 1) && !context->supports_renderpass2()) { + // This is only required when using vkCreateRenderPass, we add it if vkCreateRenderPass2KHR is not supported + // resulting this in being passed to our vkCreateRenderPass fallback. // Set view masks for each subpass. for (uint32_t i = 0; i < subpasses.size(); i++) { @@ -7727,7 +7751,7 @@ void RenderingDeviceVulkan::draw_list_bind_index_array(DrawListID p_list, RID p_ dl->validation.index_array_size = index_array->indices; dl->validation.index_array_offset = index_array->offset; - vkCmdBindIndexBuffer(dl->command_buffer, index_array->buffer, index_array->offset, index_array->index_type); + vkCmdBindIndexBuffer(dl->command_buffer, index_array->buffer, 0, index_array->index_type); } void RenderingDeviceVulkan::draw_list_set_line_width(DrawListID p_list, float p_width) { @@ -9363,12 +9387,10 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de } } - // NOTE: If adding new project settings here, also duplicate their definition in - // rendering_server.cpp for headless doctool. - staging_buffer_block_size = GLOBAL_DEF("rendering/rendering_device/staging_buffer/block_size_kb", 256); + staging_buffer_block_size = GLOBAL_GET("rendering/rendering_device/staging_buffer/block_size_kb"); staging_buffer_block_size = MAX(4u, staging_buffer_block_size); staging_buffer_block_size *= 1024; // Kb -> bytes. - staging_buffer_max_size = GLOBAL_DEF("rendering/rendering_device/staging_buffer/max_size_mb", 128); + staging_buffer_max_size = GLOBAL_GET("rendering/rendering_device/staging_buffer/max_size_mb"); staging_buffer_max_size = MAX(1u, staging_buffer_max_size); staging_buffer_max_size *= 1024 * 1024; @@ -9376,7 +9398,7 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de // Validate enough blocks. staging_buffer_max_size = staging_buffer_block_size * 4; } - texture_upload_region_size_px = GLOBAL_DEF("rendering/rendering_device/staging_buffer/texture_upload_region_size_px", 64); + texture_upload_region_size_px = GLOBAL_GET("rendering/rendering_device/staging_buffer/texture_upload_region_size_px"); texture_upload_region_size_px = nearest_power_of_2_templated(texture_upload_region_size_px); frames_drawn = frame_count; // Start from frame count, so everything else is immediately old. @@ -9391,7 +9413,7 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de ERR_CONTINUE(err != OK); } - max_descriptors_per_pool = GLOBAL_DEF("rendering/rendering_device/vulkan/max_descriptors_per_pool", 64); + max_descriptors_per_pool = GLOBAL_GET("rendering/rendering_device/vulkan/max_descriptors_per_pool"); // Check to make sure DescriptorPoolKey is good. static_assert(sizeof(uint64_t) * 3 >= UNIFORM_TYPE_MAX * sizeof(uint16_t)); diff --git a/drivers/vulkan/vulkan_context.cpp b/drivers/vulkan/vulkan_context.cpp index 9ee5a67471..464ab474e1 100644 --- a/drivers/vulkan/vulkan_context.cpp +++ b/drivers/vulkan/vulkan_context.cpp @@ -687,21 +687,43 @@ Error VulkanContext::_check_capabilities() { vkGetPhysicalDeviceFeatures2_func(gpu, &device_features); - vrs_capabilities.pipeline_vrs_supported = vrs_features.pipelineFragmentShadingRate; - vrs_capabilities.primitive_vrs_supported = vrs_features.primitiveFragmentShadingRate; - vrs_capabilities.attachment_vrs_supported = vrs_features.attachmentFragmentShadingRate; + // We must check that the relative extension is present before assuming a + // feature as enabled. Actually, according to the spec we shouldn't add the + // structs in pNext at all, but this works fine. + // See also: https://github.com/godotengine/godot/issues/65409 + for (uint32_t i = 0; i < enabled_extension_count; ++i) { + if (!strcmp(VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME, extension_names[i])) { + vrs_capabilities.pipeline_vrs_supported = vrs_features.pipelineFragmentShadingRate; + vrs_capabilities.primitive_vrs_supported = vrs_features.primitiveFragmentShadingRate; + vrs_capabilities.attachment_vrs_supported = vrs_features.attachmentFragmentShadingRate; - multiview_capabilities.is_supported = multiview_features.multiview; - multiview_capabilities.geometry_shader_is_supported = multiview_features.multiviewGeometryShader; - multiview_capabilities.tessellation_shader_is_supported = multiview_features.multiviewTessellationShader; + continue; + } - shader_capabilities.shader_float16_is_supported = shader_features.shaderFloat16; - shader_capabilities.shader_int8_is_supported = shader_features.shaderInt8; + if (!strcmp(VK_KHR_MULTIVIEW_EXTENSION_NAME, extension_names[i])) { + multiview_capabilities.is_supported = multiview_features.multiview; + multiview_capabilities.geometry_shader_is_supported = multiview_features.multiviewGeometryShader; + multiview_capabilities.tessellation_shader_is_supported = multiview_features.multiviewTessellationShader; - storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported = storage_feature.storageBuffer16BitAccess; - storage_buffer_capabilities.uniform_and_storage_buffer_16_bit_access_is_supported = storage_feature.uniformAndStorageBuffer16BitAccess; - storage_buffer_capabilities.storage_push_constant_16_is_supported = storage_feature.storagePushConstant16; - storage_buffer_capabilities.storage_input_output_16 = storage_feature.storageInputOutput16; + continue; + } + + if (!strcmp(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, extension_names[i])) { + shader_capabilities.shader_float16_is_supported = shader_features.shaderFloat16; + shader_capabilities.shader_int8_is_supported = shader_features.shaderInt8; + + continue; + } + + if (!strcmp(VK_KHR_16BIT_STORAGE_EXTENSION_NAME, extension_names[i])) { + storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported = storage_feature.storageBuffer16BitAccess; + storage_buffer_capabilities.uniform_and_storage_buffer_16_bit_access_is_supported = storage_feature.uniformAndStorageBuffer16BitAccess; + storage_buffer_capabilities.storage_push_constant_16_is_supported = storage_feature.storagePushConstant16; + storage_buffer_capabilities.storage_input_output_16 = storage_feature.storageInputOutput16; + + continue; + } + } } // Check extended properties. @@ -810,7 +832,7 @@ Error VulkanContext::_create_instance() { } } - CharString cs = ProjectSettings::get_singleton()->get("application/config/name").operator String().utf8(); + CharString cs = GLOBAL_GET("application/config/name").operator String().utf8(); const VkApplicationInfo app = { /*sType*/ VK_STRUCTURE_TYPE_APPLICATION_INFO, /*pNext*/ nullptr, @@ -1314,7 +1336,7 @@ Error VulkanContext::_create_device() { vulkan11features.shaderDrawParameters = 0; nextptr = &vulkan11features; } else { - // On Vulkan 1.0 and 1.1 we use our older structs to initialise these features. + // On Vulkan 1.0 and 1.1 we use our older structs to initialize these features. storage_feature.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR; storage_feature.pNext = nextptr; storage_feature.storageBuffer16BitAccess = storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported; @@ -1461,7 +1483,7 @@ Error VulkanContext::_initialize_queues(VkSurfaceKHR p_surface) { color_space = surfFormats[0].colorSpace; } else { // These should be ordered with the ones we want to use on top and fallback modes further down - // we want a 32bit RGBA unsigned normalised buffer or similar. + // we want a 32bit RGBA unsigned normalized buffer or similar. const VkFormat allowed_formats[] = { VK_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_R8G8B8A8_UNORM diff --git a/drivers/vulkan/vulkan_context.h b/drivers/vulkan/vulkan_context.h index b47aec1de1..d6a25c5cd7 100644 --- a/drivers/vulkan/vulkan_context.h +++ b/drivers/vulkan/vulkan_context.h @@ -273,6 +273,7 @@ protected: public: // Extension calls. + bool supports_renderpass2() const { return has_renderpass2_ext; } VkResult vkCreateRenderPass2KHR(VkDevice p_device, const VkRenderPassCreateInfo2 *p_create_info, const VkAllocationCallbacks *p_allocator, VkRenderPass *p_render_pass); uint32_t get_vulkan_major() const { return vulkan_major; }; diff --git a/drivers/wasapi/audio_driver_wasapi.cpp b/drivers/wasapi/audio_driver_wasapi.cpp index d6636606d2..f196530e51 100644 --- a/drivers/wasapi/audio_driver_wasapi.cpp +++ b/drivers/wasapi/audio_driver_wasapi.cpp @@ -127,11 +127,6 @@ static bool default_capture_device_changed = false; #pragma GCC diagnostic ignored "-Wnon-virtual-dtor" #endif -#if defined(__GNUC__) -// Workaround GCC warning from -Wcast-function-type. -#define GetProcAddress (void *)GetProcAddress -#endif - class CMMNotificationClient : public IMMNotificationClient { LONG _cRef = 1; IMMDeviceEnumerator *_pEnumerator = nullptr; @@ -206,21 +201,7 @@ public: static CMMNotificationClient notif_client; -typedef const char *(CDECL *PWineGetVersionPtr)(void); - -bool AudioDriverWASAPI::is_running_on_wine() { - HMODULE nt_lib = LoadLibraryW(L"ntdll.dll"); - if (!nt_lib) { - return false; - } - - PWineGetVersionPtr wine_get_version = (PWineGetVersionPtr)GetProcAddress(nt_lib, "wine_get_version"); - FreeLibrary(nt_lib); - - return (bool)wine_get_version; -} - -Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_capture, bool reinit) { +Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_capture, bool p_reinit, bool p_no_audio_client_3) { WAVEFORMATEX *pwfex; IMMDeviceEnumerator *enumerator = nullptr; IMMDevice *device = nullptr; @@ -286,7 +267,7 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c } } - if (reinit) { + if (p_reinit) { // In case we're trying to re-initialize the device prevent throwing this error on the console, // otherwise if there is currently no device available this will spam the console. if (hr != S_OK) { @@ -304,10 +285,11 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c } using_audio_client_3 = !p_capture; // IID_IAudioClient3 is only used for adjustable output latency (not input) - if (using_audio_client_3 && is_running_on_wine()) { + + if (p_no_audio_client_3) { using_audio_client_3 = false; - print_verbose("WASAPI: Wine detected, falling back to IAudioClient interface"); } + if (using_audio_client_3) { hr = device->Activate(IID_IAudioClient3, CLSCTX_ALL, nullptr, (void **)&p_device->audio_client); if (hr != S_OK) { @@ -325,7 +307,7 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c SAFE_RELEASE(device) - if (reinit) { + if (p_reinit) { if (hr != S_OK) { return ERR_CANT_OPEN; } @@ -436,7 +418,12 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c &fundamental_period_frames, &min_period_frames, &max_period_frames); - ERR_FAIL_COND_V_MSG(hr != S_OK, ERR_CANT_OPEN, "WASAPI: GetSharedModeEnginePeriod failed with error 0x" + String::num_uint64(hr, 16) + "."); + if (hr != S_OK) { + print_verbose("WASAPI: GetSharedModeEnginePeriod failed with error 0x" + String::num_uint64(hr, 16) + ", falling back to IAudioClient."); + CoTaskMemFree(pwfex); + SAFE_RELEASE(device) + return audio_device_init(p_device, p_capture, p_reinit, true); + } // Period frames must be an integral multiple of fundamental_period_frames or IAudioClient3 initialization will fail, // so we need to select the closest multiple to the user-specified latency. @@ -453,12 +440,25 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c buffer_frames = period_frames; hr = device_audio_client_3->InitializeSharedAudioStream(0, period_frames, pwfex, nullptr); - ERR_FAIL_COND_V_MSG(hr != S_OK, ERR_CANT_OPEN, "WASAPI: InitializeSharedAudioStream failed with error 0x" + String::num_uint64(hr, 16) + "."); - uint32_t output_latency_in_frames; - WAVEFORMATEX *current_pwfex; - device_audio_client_3->GetCurrentSharedModeEnginePeriod(¤t_pwfex, &output_latency_in_frames); - real_latency = (float)output_latency_in_frames / (float)current_pwfex->nSamplesPerSec; - CoTaskMemFree(current_pwfex); + if (hr != S_OK) { + print_verbose("WASAPI: InitializeSharedAudioStream failed with error 0x" + String::num_uint64(hr, 16) + ", falling back to IAudioClient."); + CoTaskMemFree(pwfex); + SAFE_RELEASE(device); + return audio_device_init(p_device, p_capture, p_reinit, true); + } else { + uint32_t output_latency_in_frames; + WAVEFORMATEX *current_pwfex; + hr = device_audio_client_3->GetCurrentSharedModeEnginePeriod(¤t_pwfex, &output_latency_in_frames); + if (hr == OK) { + real_latency = (float)output_latency_in_frames / (float)current_pwfex->nSamplesPerSec; + CoTaskMemFree(current_pwfex); + } else { + print_verbose("WASAPI: GetCurrentSharedModeEnginePeriod failed with error 0x" + String::num_uint64(hr, 16) + ", falling back to IAudioClient."); + CoTaskMemFree(pwfex); + SAFE_RELEASE(device); + return audio_device_init(p_device, p_capture, p_reinit, true); + } + } } if (p_capture) { @@ -475,8 +475,8 @@ Error AudioDriverWASAPI::audio_device_init(AudioDeviceWASAPI *p_device, bool p_c return OK; } -Error AudioDriverWASAPI::init_render_device(bool reinit) { - Error err = audio_device_init(&audio_output, false, reinit); +Error AudioDriverWASAPI::init_render_device(bool p_reinit) { + Error err = audio_device_init(&audio_output, false, p_reinit); if (err != OK) { return err; } @@ -507,8 +507,8 @@ Error AudioDriverWASAPI::init_render_device(bool reinit) { return OK; } -Error AudioDriverWASAPI::init_capture_device(bool reinit) { - Error err = audio_device_init(&audio_input, true, reinit); +Error AudioDriverWASAPI::init_capture_device(bool p_reinit) { + Error err = audio_device_init(&audio_input, true, p_reinit); if (err != OK) { return err; } diff --git a/drivers/wasapi/audio_driver_wasapi.h b/drivers/wasapi/audio_driver_wasapi.h index e9f2794e97..fb6a55734d 100644 --- a/drivers/wasapi/audio_driver_wasapi.h +++ b/drivers/wasapi/audio_driver_wasapi.h @@ -79,19 +79,17 @@ class AudioDriverWASAPI : public AudioDriver { SafeFlag exit_thread; - static bool is_running_on_wine(); - static _FORCE_INLINE_ void write_sample(WORD format_tag, int bits_per_sample, BYTE *buffer, int i, int32_t sample); static _FORCE_INLINE_ int32_t read_sample(WORD format_tag, int bits_per_sample, BYTE *buffer, int i); static void thread_func(void *p_udata); - Error init_render_device(bool reinit = false); - Error init_capture_device(bool reinit = false); + Error init_render_device(bool p_reinit = false); + Error init_capture_device(bool p_reinit = false); Error finish_render_device(); Error finish_capture_device(); - Error audio_device_init(AudioDeviceWASAPI *p_device, bool p_capture, bool reinit); + Error audio_device_init(AudioDeviceWASAPI *p_device, bool p_capture, bool p_reinit, bool p_no_audio_client_3 = false); Error audio_device_finish(AudioDeviceWASAPI *p_device); PackedStringArray audio_device_get_list(bool p_capture); |