server: fix software rendering

It was broken

server/src/retro_thread.rs

@@ -4,6 +4,7 @@ use std::{
     time::Duration,
 };
 
+use cudarc::driver::sys::CUmemorytype;
 use tokio::sync::{mpsc, oneshot};
 
 use anyhow::Result;
@@ -185,9 +186,12 @@ impl RetroState {
 
         let has_disconnected_pitch = pitch != size.width as u32;
 
-        for y in 0..size.height {
+        for _y in 0..size.height {
+            // Trick to flip the buffer the way OpenGL likes.
+            let y = (size.height - 1) - _y;
+
             let src_line_off = (y as u32 * pitch) as usize;
-            let mut dest_line_off = (y as u32 * size.width) as usize;
+            let mut dest_line_off = (_y as u32 * size.width) as usize;
 
             // copy only
             if has_disconnected_pitch {
@@ -224,7 +228,10 @@ impl FrontendInterface for RetroState {
                 .lock()
                 .expect("Failed to lock CUDA resource");
 
-            locked.device().bind_to_thread().expect("Failed to bind CUDA device to thread");
+            locked
+                .device()
+                .bind_to_thread()
+                .expect("Failed to bind CUDA device to thread");
 
             locked
                 .register(self.gl_framebuffer.texture_id(), gl::TEXTURE_2D)
@@ -242,18 +249,78 @@ impl FrontendInterface for RetroState {
         let size = self.software_framebuffer.size.clone();
 
         // upload texture to GPU
-        unsafe {
+        /*unsafe {
+            tracing::info!("bind tex");
+            gl::BindTexture(gl::TEXTURE_2D, self.gl_framebuffer.texture_id());
+
+            tracing::info!("upload");
             gl::TexImage2D(
                 gl::TEXTURE_2D,
                 0,
-                gl::RGBA as i32,
+                gl::RGBA8 as i32,
                 size.width as i32,
                 size.height as i32,
                 0,
-                gl::RGBA_INTEGER,
-                gl::UNSIGNED_INT_8_8_8_8,
+                gl::RGBA,
+                gl::UNSIGNED_BYTE,
                 self.software_framebuffer.get_buffer().as_mut_ptr() as *const _,
             );
+
+            tracing::info!("unbind tex");
+            gl::BindTexture(gl::TEXTURE_2D, 0);
+        }*/
+
+        // TODO: Figure out a way to do this with standard OpenGL primitives.
+        // (the above does not work *at all*. Don't ask how I know)
+        unsafe {
+            let mut cuda_resource = self
+                .cuda_resource
+                .lock()
+                .expect("Failed to lock CUDA resource");
+
+            cuda_resource
+                .device()
+                .bind_to_thread()
+                .expect("Failed to bind CUDA device to thread");
+
+            let mut mapped_cuda_resource = cuda_resource
+                .map()
+                .expect("Failed to map CUDA resource");
+
+            let array = mapped_cuda_resource
+                .get_mapped_array()
+                .expect("Failed to get CUarray from CUDA resource");
+
+            let mut memcpy = cudarc::driver::sys::CUDA_MEMCPY2D_st::default();
+
+            // src
+            memcpy.srcXInBytes = 0;
+            memcpy.srcY = 0;
+            memcpy.srcMemoryType = CUmemorytype::CU_MEMORYTYPE_HOST;
+            memcpy.srcHost = self.software_framebuffer.get_buffer().as_mut_ptr() as *const _;
+
+            // dest
+            memcpy.dstXInBytes = 0;
+            memcpy.dstY = 0;
+            memcpy.dstMemoryType = CUmemorytype::CU_MEMORYTYPE_ARRAY;
+            memcpy.dstArray = array;
+            
+            memcpy.WidthInBytes = (size.width * 4) as usize;
+            memcpy.Height = size.height as usize;
+            memcpy.dstPitch = (size.width * 4) as usize;
+
+            // kick it off
+            cudarc::driver::sys::lib()
+                .cuMemcpy2DAsync_v2(&memcpy, std::ptr::null_mut())
+                .result()
+                .expect("cuMemcpy2D fail epic");
+
+            cudarc::driver::sys::lib()
+                .cuStreamSynchronize(std::ptr::null_mut())
+                .result()
+                .expect("fucking");
+
+                mapped_cuda_resource.unmap().expect("fuck you asshole");
         }
 
         let _ = self.event_tx.blocking_send(RetroEvent::Frame);
@@ -353,6 +420,9 @@ impl FrontendInterface for RetroState {
     }
 
     fn hw_gl_init(&mut self) -> Option<HwGlInitData> {
+        // test SW-only even if a core wants to upgrade
+        //return None;
+
         // Only create a new EGL/OpenGL context if we have to.
         let context = self.egl_context.lock().expect("fuck you!");
         let extensions = gpu::egl_helpers::get_extensions(context.get_display());