rpcs3/rpcs3/Emu/RSX/GL/GLDraw.cpp

#include "stdafx.h"
#include "GLGSRender.h"
#include "../rsx_methods.h"
#include "../Common/BufferUtils.h"
#include "../Program/GLSLCommon.h"

#include "Emu/RSX/NV47/HW/context_accessors.define.h"

namespace gl
{
	inline GLenum comparison_op(rsx::comparison_function op)
	{
		return static_cast<GLenum>(op);
	}

	inline GLenum stencil_op(rsx::stencil_op op)
	{
		return static_cast<GLenum>(op);
	}

	GLenum blend_equation(rsx::blend_equation op)
	{
		switch (op)
		{
			// Note : maybe add is signed on gl
		case rsx::blend_equation::add_signed:
			rsx_log.trace("blend equation add_signed used. Emulating using FUNC_ADD");
			[[fallthrough]];
		case rsx::blend_equation::add: return GL_FUNC_ADD;
		case rsx::blend_equation::min: return GL_MIN;
		case rsx::blend_equation::max: return GL_MAX;
		case rsx::blend_equation::subtract: return GL_FUNC_SUBTRACT;
		case rsx::blend_equation::reverse_subtract_signed:
			rsx_log.trace("blend equation reverse_subtract_signed used. Emulating using FUNC_REVERSE_SUBTRACT");
			[[fallthrough]];
		case rsx::blend_equation::reverse_subtract: return GL_FUNC_REVERSE_SUBTRACT;
		case rsx::blend_equation::reverse_add_signed:
		default:
			rsx_log.error("Blend equation 0x%X is unimplemented!", static_cast<u32>(op));
			return GL_FUNC_ADD;
		}
	}

	inline GLenum blend_factor(rsx::blend_factor op)
	{
		return static_cast<GLenum>(op);
	}

	inline GLenum logic_op(rsx::logic_op op)
	{
		return static_cast<GLenum>(op);
	}

	inline GLenum front_face(rsx::front_face op)
	{
		// NOTE: RSX face winding is always based off of upper-left corner like vulkan, but GL is bottom left
		// shader_window_origin register does not affect this
		// verified with Outrun Online Arcade (window_origin::top) and DS2 (window_origin::bottom)
		// correctness of face winding checked using stencil test (GOW collection shadows)
		return static_cast<GLenum>(op) ^ 1u;
	}

	inline GLenum cull_face(rsx::cull_face op)
	{
		return static_cast<GLenum>(op);
	}

	inline GLenum polygon_mode(rsx::polygon_mode mode)
	{
		return static_cast<GLenum>(mode);
	}
}

void GLGSRender::update_draw_state()
{
	m_profiler.start();

	gl_state.enable(GL_SCISSOR_TEST);
	gl_state.enable(rsx::method_registers.dither_enabled(), GL_DITHER);

	if (m_rtts.m_bound_depth_stencil.first)
	{
		// Z-buffer is active.
		gl_state.depth_mask(rsx::method_registers.depth_write_enabled());
		gl_state.stencil_mask(rsx::method_registers.stencil_mask());

		gl_state.enable(rsx::method_registers.depth_clamp_enabled() || !rsx::method_registers.depth_clip_enabled(), GL_DEPTH_CLAMP);

		if (gl_state.enable(rsx::method_registers.depth_test_enabled(), GL_DEPTH_TEST))
		{
			gl_state.depth_func(gl::comparison_op(rsx::method_registers.depth_func()));
		}

		if (gl::get_driver_caps().EXT_depth_bounds_test_supported && (gl_state.enable(rsx::method_registers.depth_bounds_test_enabled(), GL_DEPTH_BOUNDS_TEST_EXT)))
		{
			gl_state.depth_bounds(rsx::method_registers.depth_bounds_min(), rsx::method_registers.depth_bounds_max());
		}

		if (gl::get_driver_caps().NV_depth_buffer_float_supported)
		{
			gl_state.depth_range(rsx::method_registers.clip_min(), rsx::method_registers.clip_max());
		}

		if (gl_state.enable(rsx::method_registers.stencil_test_enabled(), GL_STENCIL_TEST))
		{
			gl_state.stencil_func(gl::comparison_op(rsx::method_registers.stencil_func()),
				rsx::method_registers.stencil_func_ref(),
				rsx::method_registers.stencil_func_mask());

			gl_state.stencil_op(gl::stencil_op(rsx::method_registers.stencil_op_fail()), gl::stencil_op(rsx::method_registers.stencil_op_zfail()),
				gl::stencil_op(rsx::method_registers.stencil_op_zpass()));

			if (rsx::method_registers.two_sided_stencil_test_enabled())
			{
				gl_state.stencil_back_mask(rsx::method_registers.back_stencil_mask());

				gl_state.stencil_back_func(gl::comparison_op(rsx::method_registers.back_stencil_func()),
					rsx::method_registers.back_stencil_func_ref(), rsx::method_registers.back_stencil_func_mask());

				gl_state.stencil_back_op(gl::stencil_op(rsx::method_registers.back_stencil_op_fail()),
					gl::stencil_op(rsx::method_registers.back_stencil_op_zfail()), gl::stencil_op(rsx::method_registers.back_stencil_op_zpass()));
			}
		}
	}

	if (m_rtts.get_color_surface_count())
	{
		// Color buffer is active
		const auto host_write_mask = rsx::get_write_output_mask(rsx::method_registers.surface_color());
		for (int index = 0; index < m_rtts.get_color_surface_count(); ++index)
		{
			bool color_mask_b = rsx::method_registers.color_mask_b(index);
			bool color_mask_g = rsx::method_registers.color_mask_g(index);
			bool color_mask_r = rsx::method_registers.color_mask_r(index);
			bool color_mask_a = rsx::method_registers.color_mask_a(index);

			switch (rsx::method_registers.surface_color())
			{
			case rsx::surface_color_format::b8:
				rsx::get_b8_colormask(color_mask_r, color_mask_g, color_mask_b, color_mask_a);
				break;
			case rsx::surface_color_format::g8b8:
				rsx::get_g8b8_r8g8_colormask(color_mask_r, color_mask_g, color_mask_b, color_mask_a);
				break;
			default:
				break;
			}

			gl_state.color_maski(
				index,
				color_mask_r && host_write_mask[0],
				color_mask_g && host_write_mask[1],
				color_mask_b && host_write_mask[2],
				color_mask_a && host_write_mask[3]);
		}

		// LogicOp and Blend are mutually exclusive. If both are enabled, LogicOp takes precedence.
		// In OpenGL, this behavior is enforced in spec, but let's enforce it at renderer level as well.
		if (gl_state.enable(rsx::method_registers.logic_op_enabled(), GL_COLOR_LOGIC_OP))
		{
			gl_state.logic_op(gl::logic_op(rsx::method_registers.logic_operation()));

			gl_state.enablei(GL_FALSE, GL_BLEND, 0);
			gl_state.enablei(GL_FALSE, GL_BLEND, 1);
			gl_state.enablei(GL_FALSE, GL_BLEND, 2);
			gl_state.enablei(GL_FALSE, GL_BLEND, 3);
		}
		else
		{
			bool mrt_blend_enabled[] =
			{
				rsx::method_registers.blend_enabled(),
				rsx::method_registers.blend_enabled_surface_1(),
				rsx::method_registers.blend_enabled_surface_2(),
				rsx::method_registers.blend_enabled_surface_3()
			};

			if (mrt_blend_enabled[0] || mrt_blend_enabled[1] || mrt_blend_enabled[2] || mrt_blend_enabled[3])
			{
				glBlendFuncSeparate(gl::blend_factor(rsx::method_registers.blend_func_sfactor_rgb()),
					gl::blend_factor(rsx::method_registers.blend_func_dfactor_rgb()),
					gl::blend_factor(rsx::method_registers.blend_func_sfactor_a()),
					gl::blend_factor(rsx::method_registers.blend_func_dfactor_a()));

				auto blend_colors = rsx::get_constant_blend_colors();
				glBlendColor(blend_colors[0], blend_colors[1], blend_colors[2], blend_colors[3]);

				glBlendEquationSeparate(gl::blend_equation(rsx::method_registers.blend_equation_rgb()),
					gl::blend_equation(rsx::method_registers.blend_equation_a()));
			}

			gl_state.enablei(mrt_blend_enabled[0], GL_BLEND, 0);
			gl_state.enablei(mrt_blend_enabled[1], GL_BLEND, 1);
			gl_state.enablei(mrt_blend_enabled[2], GL_BLEND, 2);
			gl_state.enablei(mrt_blend_enabled[3], GL_BLEND, 3);
		}

		// Antialias control
		if (backend_config.supports_hw_msaa)
		{
			gl_state.enable(/*REGS(m_ctx)->msaa_enabled()*/GL_MULTISAMPLE);

			gl_state.enable(GL_SAMPLE_MASK);
			gl_state.sample_mask(REGS(m_ctx)->msaa_sample_mask());

			gl_state.enable(GL_SAMPLE_SHADING);
			gl_state.min_sample_shading_rate(1.f);

			gl_state.enable(GL_SAMPLE_COVERAGE);
			gl_state.sample_coverage(1.f);
		}

		if (backend_config.supports_hw_a2c)
		{
			const bool hw_enable = backend_config.supports_hw_a2c_1spp || REGS(m_ctx)->surface_antialias() != rsx::surface_antialiasing::center_1_sample;
			gl_state.enable(hw_enable && REGS(m_ctx)->msaa_alpha_to_coverage_enabled(), GL_SAMPLE_ALPHA_TO_COVERAGE);
		}

		if (backend_config.supports_hw_a2one)
		{
			gl_state.enable(REGS(m_ctx)->msaa_alpha_to_one_enabled(), GL_SAMPLE_ALPHA_TO_ONE);
		}
	}

	switch (rsx::method_registers.current_draw_clause.primitive)
	{
	case rsx::primitive_type::lines:
	case rsx::primitive_type::line_loop:
	case rsx::primitive_type::line_strip:
		gl_state.line_width(rsx::method_registers.line_width() * resolution_scaling_config.scale_factor());
		gl_state.enable(rsx::method_registers.line_smooth_enabled(), GL_LINE_SMOOTH);
		break;
	default:
		gl_state.enable(rsx::method_registers.poly_offset_point_enabled(), GL_POLYGON_OFFSET_POINT);
		gl_state.enable(rsx::method_registers.poly_offset_line_enabled(), GL_POLYGON_OFFSET_LINE);
		gl_state.enable(rsx::method_registers.poly_offset_fill_enabled(), GL_POLYGON_OFFSET_FILL);

		// offset_bias is the constant factor, multiplied by the implementation factor R
		// offset_scale is the slope factor, multiplied by the triangle slope factor M
		const auto poly_offset_scale = rsx::method_registers.poly_offset_scale();
		auto poly_offset_bias = rsx::method_registers.poly_offset_bias();

		if (auto ds = m_rtts.m_bound_depth_stencil.second;
			ds && ds->get_internal_format() == gl::texture::internal_format::depth24_stencil8)
		{
			// Check details in VKDraw.cpp about behaviour of RSX vs desktop D24X8 implementations
			// TLDR, RSX expects R = 16,777,215 (2^24 - 1)
			const auto& caps = gl::get_driver_caps();
			if (caps.vendor_NVIDIA || caps.vendor_MESA)
			{
				// R derived to be 8388607 (2^23 - 1)
				poly_offset_bias *= 0.5f;
			}
			else if (caps.vendor_AMD)
			{
				// R derived to be 4194303 (2^22 - 1)
				poly_offset_bias *= 0.25f;
			}
		}
		gl_state.polygon_offset(poly_offset_scale, poly_offset_bias);

		if (gl_state.enable(rsx::method_registers.cull_face_enabled(), GL_CULL_FACE))
		{
			gl_state.cull_face(gl::cull_face(rsx::method_registers.cull_face_mode()));
		}

		gl_state.front_face(gl::front_face(rsx::method_registers.front_face_mode()));
		break;
	}

	// Clip planes
	gl_state.clip_planes((current_vertex_program.output_mask >> CELL_GCM_ATTRIB_OUTPUT_UC0) & 0x3F);

	// Polygon mode. We can only have one polygon mode active at one time, so we need to determine if any face is currently culled.
	const bool show_back = REGS(m_ctx)->cull_face_enabled() && REGS(m_ctx)->cull_face_mode() == rsx::cull_face::front;
	gl_state.polygon_mode(gl::polygon_mode(show_back ? REGS(m_ctx)->polygon_mode_back() : REGS(m_ctx)->polygon_mode_front()));

	//TODO
	//NV4097_SET_ANISO_SPREAD
	//NV4097_SET_SPECULAR_ENABLE
	//NV4097_SET_TWO_SIDE_LIGHT_EN
	//NV4097_SET_FLAT_SHADE_OP
	//NV4097_SET_EDGE_FLAG
	//NV4097_SET_COLOR_KEY_COLOR
	//NV4097_SET_CLIP_ID_TEST_ENABLE

	// For OGL Z range is updated every draw as it is separate from viewport config
	m_graphics_state.clear(rsx::pipeline_state::zclip_config_state_dirty);

	m_frame_stats.setup_time += m_profiler.duration();
}

void GLGSRender::load_texture_env()
{
	// Load textures
	gl::command_context cmd{ gl_state };
	std::lock_guard lock(m_sampler_mutex);

	for (u32 textures_ref = current_fp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
	{
		if (!(textures_ref & 1))
		{
			continue;
		}

		if (!fs_sampler_state[i])
		{
			fs_sampler_state[i] = std::make_unique<gl::texture_cache::sampled_image_descriptor>();
		}

		auto sampler_state = static_cast<gl::texture_cache::sampled_image_descriptor*>(fs_sampler_state[i].get());
		const auto& tex = rsx::method_registers.fragment_textures[i];
		const auto previous_format_class = sampler_state->format_class;

		if (!m_samplers_dirty &&
			!m_textures_dirty[i] &&
			!m_gl_texture_cache.test_if_descriptor_expired(cmd, m_rtts, sampler_state, tex))
		{
			continue;
		}

		const bool is_sampler_dirty = m_textures_dirty[i];
		m_textures_dirty[i] = false;

		if (!tex.enabled())
		{
			*sampler_state = {};
			continue;
		}

		*sampler_state = m_gl_texture_cache.upload_texture(cmd, tex, m_rtts);
		if (!sampler_state->validate())
		{
			continue;
		}

		if (!is_sampler_dirty)
		{
			if (sampler_state->format_class != previous_format_class)
			{
				// Host details changed but RSX is not aware
				m_graphics_state |= rsx::fragment_program_state_dirty;
			}

			if (sampler_state->format_ex)
			{
				// Nothing to change, use cached sampler
				continue;
			}
		}

		sampler_state->format_ex = tex.format_ex();

		if (sampler_state->format_ex.texel_remap_control &&
			sampler_state->image_handle &&
			sampler_state->upload_context == rsx::texture_upload_context::shader_read &&
			(current_fp_metadata.bx2_texture_reads_mask & (1u << i)) == 0 &&
			!g_cfg.video.disable_hardware_texel_remapping) [[ unlikely ]]
		{
			// Check if we need to override the view format
			const auto gl_format = sampler_state->image_handle->view_format();
			GLenum format_override = gl_format;
			rsx::flags32_t flags_to_erase = 0u;
			rsx::flags32_t host_flags_to_set = 0u;

			if (sampler_state->format_ex.hw_SNORM_possible())
			{
				format_override = gl::get_compatible_snorm_format(gl_format);
				flags_to_erase = rsx::texture_control_bits::SEXT_MASK;
				host_flags_to_set = rsx::RSX_HOST_FORMAT_FEATURE_SNORM;
			}
			else if (sampler_state->format_ex.hw_SRGB_possible())
			{
				format_override = gl::get_compatible_srgb_format(gl_format);
				flags_to_erase = rsx::texture_control_bits::GAMMA_CTRL_MASK;
				host_flags_to_set = rsx::RSX_HOST_FORMAT_FEATURE_SRGB;
			}

			if (format_override != GL_NONE && format_override != gl_format)
			{
				sampler_state->image_handle = sampler_state->image_handle->as(format_override);
				sampler_state->format_ex.texel_remap_control &= (~flags_to_erase);
				sampler_state->format_ex.host_features |= host_flags_to_set;
			}
		}

		u32 actual_mipcount = 1;
		if (sampler_state->upload_context == rsx::texture_upload_context::shader_read)
		{
			actual_mipcount = tex.get_exact_mipmap_count();
		}
		else if (sampler_state->external_subresource_desc.op == rsx::deferred_request_command::mipmap_gather)
		{
			actual_mipcount = sampler_state->external_subresource_desc.sections_to_copy.size();
		}
		else if (sampler_state->external_subresource_desc.op == rsx::deferred_request_command::cubemap_unwrap)
		{
			actual_mipcount = sampler_state->external_subresource_desc.mipmaps;
		}

		m_fs_sampler_states[i].apply(tex, fs_sampler_state[i].get(), actual_mipcount > 1);

		const auto texture_format = sampler_state->format_ex.format();
		// Depth format redirected to BGRA8 resample stage. Do not filter to avoid bits leaking.
		// If accurate graphics are desired, force a bitcast to COLOR as a workaround.
		const bool is_depth_reconstructed = sampler_state->format_class != rsx::classify_format(texture_format) &&
			(texture_format == CELL_GCM_TEXTURE_A8R8G8B8 || texture_format == CELL_GCM_TEXTURE_D8R8G8B8);
		// SNORM conversion required in shader. Do not interpolate to avoid introducing discontinuities due to how negative numbers work
		const bool is_snorm = (sampler_state->format_ex.texel_remap_control & rsx::texture_control_bits::SEXT_MASK) != 0;

		if (is_depth_reconstructed || is_snorm)
		{
			// Depth format redirected to BGRA8 resample stage. Do not filter to avoid bits leaking.
			m_fs_sampler_states[i].set_parameteri(GL_TEXTURE_MIN_FILTER, GL_NEAREST);
			m_fs_sampler_states[i].set_parameteri(GL_TEXTURE_MAG_FILTER, GL_NEAREST);
		}
	}

	for (u32 textures_ref = current_vp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
	{
		if (!(textures_ref & 1))
		{
			continue;
		}

		if (!vs_sampler_state[i])
		{
			vs_sampler_state[i] = std::make_unique<gl::texture_cache::sampled_image_descriptor>();
		}

		auto sampler_state = static_cast<gl::texture_cache::sampled_image_descriptor*>(vs_sampler_state[i].get());
		const auto& tex = rsx::method_registers.vertex_textures[i];
		const auto previous_format_class = sampler_state->format_class;

		if (!m_samplers_dirty &&
			!m_vertex_textures_dirty[i] &&
			!m_gl_texture_cache.test_if_descriptor_expired(cmd, m_rtts, sampler_state, tex))
		{
			continue;
		}

		const bool is_sampler_dirty = m_vertex_textures_dirty[i];
		m_vertex_textures_dirty[i] = false;

		if (!tex.enabled())
		{
			*sampler_state = {};
			continue;
		}

		*sampler_state = m_gl_texture_cache.upload_texture(cmd, rsx::method_registers.vertex_textures[i], m_rtts);

		if (!sampler_state->validate())
		{
			continue;
		}

		if (is_sampler_dirty)
		{
			m_vs_sampler_states[i].apply(tex, vs_sampler_state[i].get());
		}
		else if (sampler_state->format_class != previous_format_class)
		{
			m_graphics_state |= rsx::vertex_program_state_dirty;
		}
	}

	m_samplers_dirty.store(false);
}

void GLGSRender::bind_texture_env()
{
	// Bind textures and resolve external copy operations
	gl::command_context cmd{ gl_state };

	for (u32 textures_ref = current_fp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
	{
		if (!(textures_ref & 1))
		{
			continue;
		}

		gl::texture_view* view = nullptr;
		auto sampler_state = static_cast<gl::texture_cache::sampled_image_descriptor*>(fs_sampler_state[i].get());

		if (rsx::method_registers.fragment_textures[i].enabled() &&
			sampler_state->validate())
		{
			if (view = sampler_state->image_handle; !view) [[unlikely]]
			{
				view = m_gl_texture_cache.create_temporary_subresource(cmd, sampler_state->external_subresource_desc);
			}
		}

		if (view) [[likely]]
		{
			view->bind(cmd, GL_FRAGMENT_TEXTURES_START + i);

			if (current_fragment_program.texture_state.redirected_textures & (1 << i))
			{
				auto root_texture = static_cast<gl::viewable_image*>(view->image());
				auto stencil_view = root_texture->get_view(rsx::default_remap_vector.with_encoding(gl::GL_REMAP_IDENTITY), gl::image_aspect::stencil);
				stencil_view->bind(cmd, GL_STENCIL_MIRRORS_START + i);
			}
		}
		else
		{
			const auto target = gl::get_target(current_fragment_program.get_texture_dimension(i));
			cmd->bind_texture(GL_FRAGMENT_TEXTURES_START + i, target, m_null_textures[target]->id());

			if (current_fragment_program.texture_state.redirected_textures & (1 << i))
			{
				cmd->bind_texture(GL_STENCIL_MIRRORS_START + i, target, m_null_textures[target]->id());
			}
		}
	}

	for (u32 textures_ref = current_vp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
	{
		if (!(textures_ref & 1))
		{
			continue;
		}

		auto sampler_state = static_cast<gl::texture_cache::sampled_image_descriptor*>(vs_sampler_state[i].get());
		gl::texture_view* view = nullptr;

		if (rsx::method_registers.vertex_textures[i].enabled() &&
			sampler_state->validate())
		{
			if (view = sampler_state->image_handle; !view)
			{
				view = m_gl_texture_cache.create_temporary_subresource(cmd, sampler_state->external_subresource_desc);
			}
		}

		if (view) [[likely]]
		{
			view->bind(cmd, GL_VERTEX_TEXTURES_START + i);
		}
		else
		{
			cmd->bind_texture(GL_VERTEX_TEXTURES_START + i, GL_TEXTURE_2D, GL_NONE);
		}
	}
}

void GLGSRender::emit_geometry(u32 sub_index)
{
	const auto do_heap_cleanup = [this]()
	{
		if (manually_flush_ring_buffers)
		{
			m_attrib_ring_buffer->unmap();
			m_index_ring_buffer->unmap();
		}
		else
		{
			//DMA push; not needed with MAP_COHERENT
			//glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
		}
	};

	m_profiler.start();

	auto& draw_call = rsx::method_registers.current_draw_clause;
	const rsx::flags32_t vertex_state_mask = rsx::vertex_base_changed | rsx::vertex_arrays_changed;
	const rsx::flags32_t vertex_state = (sub_index == 0) ? rsx::vertex_arrays_changed : draw_call.execute_pipeline_dependencies(m_ctx) & vertex_state_mask;

	if (vertex_state & rsx::vertex_arrays_changed)
	{
		m_draw_processor.analyse_inputs_interleaved(m_vertex_layout, current_vp_metadata);
	}
	else if (vertex_state & rsx::vertex_base_changed)
	{
		// Rebase vertex bases instead of
		for (auto& info : m_vertex_layout.interleaved_blocks)
		{
			info->vertex_range.second = 0;
			const auto vertex_base_offset = rsx::method_registers.vertex_data_base_offset();
			info->real_offset_address = rsx::get_address(rsx::get_vertex_offset_from_base(vertex_base_offset, info->base_offset), info->memory_location);
		}
	}
	else
	{
		// Discard cached results
		for (auto& info : m_vertex_layout.interleaved_blocks)
		{
			info->vertex_range.second = 0;
		}
	}

	if (vertex_state && !m_vertex_layout.validate())
	{
		// No vertex inputs enabled
		// Execute remaining pipeline barriers with NOP draw
		do
		{
			draw_call.execute_pipeline_dependencies(m_ctx);
		} while (draw_call.next());

		draw_call.end();
		return;
	}

	if (manually_flush_ring_buffers)
	{
		//Use approximations to reserve space. This path is mostly for debug purposes anyway
		u32 approx_vertex_count = draw_call.get_elements_count();
		u32 approx_working_buffer_size = approx_vertex_count * 256;

		//Allocate 256K heap if we have no approximation at this time (inlined array)
		m_attrib_ring_buffer->reserve_storage_on_heap(std::max(approx_working_buffer_size, 256 * 1024U));
		m_index_ring_buffer->reserve_storage_on_heap(16 * 1024);
	}

	// Do vertex upload before RTT prep / texture lookups to give the driver time to push data
	auto upload_info = set_vertex_buffer();
	do_heap_cleanup();

	if (upload_info.vertex_draw_count == 0)
	{
		// Malformed vertex setup; abort
		return;
	}

	const GLenum draw_mode = gl::draw_mode(draw_call.primitive);
	update_vertex_env(upload_info);

	m_frame_stats.vertex_upload_time += m_profiler.duration();

	gl_state.use_program(m_program->id());

	if (!upload_info.index_info)
	{
		if (draw_call.is_trivial_instanced_draw)
		{
			glDrawArraysInstanced(draw_mode, 0, upload_info.vertex_draw_count, draw_call.pass_count());
		}
		else if (draw_call.is_single_draw())
		{
			glDrawArrays(draw_mode, 0, upload_info.vertex_draw_count);
		}
		else
		{
			const auto& subranges = draw_call.get_subranges();
			const auto draw_count = subranges.size();
			const auto& driver_caps = gl::get_driver_caps();
			bool use_draw_arrays_fallback = false;

			m_scratch_buffer.resize(draw_count * 24);
			GLint* firsts = reinterpret_cast<GLint*>(m_scratch_buffer.data());
			GLsizei* counts = (firsts + draw_count);
			const GLvoid** offsets = utils::bless<const GLvoid*>(counts + draw_count);

			u32 first = 0;
			u32 dst_index = 0;
			for (const auto& range : subranges)
			{
				firsts[dst_index] = first;
				counts[dst_index] = range.count;
				offsets[dst_index++] = reinterpret_cast<const GLvoid*>(u64{first << 2});

				if (driver_caps.vendor_AMD && (first + range.count) > (0x100000 >> 2))
				{
					// Unlikely, but added here in case the identity buffer is not large enough somehow
					use_draw_arrays_fallback = true;
					break;
				}

				first += range.count;
			}

			if (use_draw_arrays_fallback)
			{
				// MultiDrawArrays is broken on some primitive types using AMD. One known type is GL_TRIANGLE_STRIP but there could be more
				for (u32 n = 0; n < draw_count; ++n)
				{
					glDrawArrays(draw_mode, firsts[n], counts[n]);
				}
			}
			else if (driver_caps.vendor_AMD)
			{
				// Use identity index buffer to fix broken vertexID on AMD
				m_identity_index_buffer->bind();
				glMultiDrawElements(draw_mode, counts, GL_UNSIGNED_INT, offsets, static_cast<GLsizei>(draw_count));
			}
			else
			{
				// Normal render
				glMultiDrawArrays(draw_mode, firsts, counts, static_cast<GLsizei>(draw_count));
			}
		}
	}
	else
	{
		const GLenum index_type = std::get<0>(*upload_info.index_info);
		const u32 index_offset = std::get<1>(*upload_info.index_info);
		const bool restarts_valid = gl::is_primitive_native(draw_call.primitive) && !draw_call.is_disjoint_primitive;

		if (gl_state.enable(restarts_valid && rsx::method_registers.restart_index_enabled(), GL_PRIMITIVE_RESTART))
		{
			glPrimitiveRestartIndex((index_type == GL_UNSIGNED_SHORT) ? 0xffff : 0xffffffff);
		}

		m_index_ring_buffer->bind();

		if (draw_call.is_trivial_instanced_draw)
		{
			glDrawElementsInstanced(draw_mode, upload_info.vertex_draw_count, index_type, reinterpret_cast<GLvoid*>(u64{ index_offset }), draw_call.pass_count());
		}
		else if (draw_call.is_single_draw())
		{
			glDrawElements(draw_mode, upload_info.vertex_draw_count, index_type, reinterpret_cast<GLvoid*>(u64{index_offset}));
		}
		else
		{
			const auto subranges = draw_call.get_subranges();
			const auto draw_count = subranges.size();
			const u32 type_scale = (index_type == GL_UNSIGNED_SHORT) ? 1 : 2;
			uptr index_ptr = index_offset;
			m_scratch_buffer.resize(draw_count * 16);

			GLsizei *counts = reinterpret_cast<GLsizei*>(m_scratch_buffer.data());
			const GLvoid** offsets = utils::bless<const GLvoid*>(counts + draw_count);
			int dst_index = 0;

			for (const auto &range : subranges)
			{
				const auto index_size = get_index_count(draw_call.primitive, range.count);
				counts[dst_index] = index_size;
				offsets[dst_index++] = reinterpret_cast<const GLvoid*>(index_ptr);

				index_ptr += (index_size << type_scale);
			}

			glMultiDrawElements(draw_mode, counts, index_type, offsets, static_cast<GLsizei>(draw_count));
		}
	}

	m_frame_stats.draw_exec_time += m_profiler.duration();
}

void GLGSRender::begin()
{
	// Save shader state now before prefetch and loading happens
	m_interpreter_state = (m_graphics_state.load() & rsx::pipeline_state::invalidate_pipeline_bits);

	rsx::thread::begin();

	if (skip_current_frame || cond_render_ctrl.disable_rendering())
	{
		return;
	}

	init_buffers(rsx::framebuffer_creation_context::context_draw);

	if (m_graphics_state & rsx::pipeline_state::invalidate_pipeline_bits)
	{
		// Shaders need to be reloaded.
		m_prev_program = m_program;
		m_program = nullptr;
	}
}

void GLGSRender::end()
{
	m_profiler.start();

	if (skip_current_frame || !m_graphics_state.test(rsx::rtt_config_valid) || cond_render_ctrl.disable_rendering())
	{
		execute_nop_draw();
		rsx::thread::end();
		return;
	}

	analyse_current_rsx_pipeline();

	m_frame_stats.setup_time += m_profiler.duration();

	// Active texture environment is used to decode shaders
	load_texture_env();
	m_frame_stats.textures_upload_time += m_profiler.duration();

	// NOTE: Due to common OpenGL driver architecture, vertex data has to be uploaded as far away from the draw as possible
	// TODO: Implement shaders cache prediction to avoid uploading vertex data if draw is going to skip
	if (!load_program())
	{
		// Program is not ready, skip drawing this
		std::this_thread::yield();
		execute_nop_draw();
		// m_rtts.on_write(); - breaks games for obvious reasons
		rsx::thread::end();
		return;
	}

	// Load program execution environment
	load_program_env();
	m_frame_stats.setup_time += m_profiler.duration();

	bind_texture_env();
	m_gl_texture_cache.release_uncached_temporary_subresources();
	m_frame_stats.textures_upload_time += m_profiler.duration();

	gl::command_context cmd{ gl_state };
	if (auto ds = std::get<1>(m_rtts.m_bound_depth_stencil))
	{
		ds->write_barrier(cmd);
	}

	for (auto &rtt : m_rtts.m_bound_render_targets)
	{
		if (auto surface = std::get<1>(rtt))
		{
			surface->write_barrier(cmd);
		}
	}

	m_graphics_state.clear(rsx::zeta_address_cyclic_barrier);

	update_draw_state();

	if (g_cfg.video.debug_output)
	{
		m_program->validate();
	}

	auto& draw_call = REGS(m_ctx)->current_draw_clause;
	draw_call.begin();
	u32 subdraw = 0u;
	do
	{
		emit_geometry(subdraw++);

		if (draw_call.is_trivial_instanced_draw)
		{
			// We already completed. End the draw.
			draw_call.end();
		}
	}
	while (draw_call.next());

	m_rtts.on_write(m_framebuffer_layout.color_write_enabled, m_framebuffer_layout.zeta_write_enabled);

	m_attrib_ring_buffer->notify();
	m_index_ring_buffer->notify();
	m_fragment_env_buffer->notify();
	m_vertex_env_buffer->notify();
	m_texture_parameters_buffer->notify();
	m_vertex_layout_buffer->notify();
	m_fragment_constants_buffer->notify();
	m_transform_constants_buffer->notify();
	m_instancing_ring_buffer->notify();

	m_frame_stats.setup_time += m_profiler.duration();

	rsx::thread::end();
}