From dd4225f27e6b43e85694096d96adb211efa7aedd Mon Sep 17 00:00:00 2001
From: Misaki Kasumi <misakikasumi@outlook.com>
Date: Fri, 17 May 2024 14:46:53 +0800
Subject: [PATCH] dispatch: add option "no_compute"

---
 src/dispatch.c                    |  2 +-
 src/include/libplacebo/dispatch.h |  3 +++
 src/include/libplacebo/renderer.h |  2 +-
 src/renderer.c                    | 11 +++++++++--
 4 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/src/dispatch.c b/src/dispatch.c
index ce4e667d..ee76ddf2 100644
--- a/src/dispatch.c
+++ b/src/dispatch.c
@@ -1208,7 +1208,7 @@ bool pl_dispatch_finish(pl_dispatch dp, const struct pl_dispatch_params *params)
     }
 
     const struct pl_gpu_limits *limits = &dp->gpu->limits;
-    bool can_compute = tpars->storable;
+    bool can_compute = tpars->storable && !params->no_compute;
     if (can_compute && params->blend_params)
         can_compute = tpars->format->caps & PL_FMT_CAP_READWRITE;
 
diff --git a/src/include/libplacebo/dispatch.h b/src/include/libplacebo/dispatch.h
index 141055c2..f41fea9a 100644
--- a/src/include/libplacebo/dispatch.h
+++ b/src/include/libplacebo/dispatch.h
@@ -115,6 +115,9 @@ struct pl_dispatch_params {
     // execution time of the shader, which means `pl_dispatch_info.samples` may
     // be empty as a result.
     pl_timer timer;
+
+    // Disable the use of compute shaders.
+    bool no_compute;
 };
 
 #define pl_dispatch_params(...) (&(struct pl_dispatch_params) { __VA_ARGS__ })
diff --git a/src/include/libplacebo/renderer.h b/src/include/libplacebo/renderer.h
index f52a0ea6..eca034e2 100644
--- a/src/include/libplacebo/renderer.h
+++ b/src/include/libplacebo/renderer.h
@@ -679,7 +679,7 @@ static inline void pl_frame_clear(pl_gpu gpu, const struct pl_frame *frame,
 
 // Helper function to clear a frame to a fully tiled background.
 PL_API void pl_frame_clear_tiles(pl_gpu gpu, const struct pl_frame *frame,
-                                 const float tile_colors[2][3], int tile_size);
+                                 const float tile_colors[2][3], int tile_size, bool no_compute);
 
 // Helper functions to return the fixed/inferred pl_frame parameters used
 // for rendering internally. Mutates `image` and `target` in-place to hold
diff --git a/src/renderer.c b/src/renderer.c
index 7f518f6c..c3eb675a 100644
--- a/src/renderer.c
+++ b/src/renderer.c
@@ -518,6 +518,7 @@ static pl_tex _img_tex(struct pass_state *pass, struct img *img, pl_debug_tag ta
     bool ok = pl_dispatch_finish(rr->dp, pl_dispatch_params(
         .shader = &img->sh,
         .target = tex,
+        .no_compute = pass->params->no_compute,
     ));
 
     const char *err_msg = img->err_msg;
@@ -2064,6 +2065,7 @@ static pl_tex get_feature_map(struct pass_state *pass)
     bool ok = pl_dispatch_finish(rr->dp, pl_dispatch_params(
         .shader = &sh,
         .target = inter_tex,
+        .no_compute = pass->params->no_compute,
     ));
     if (!ok)
         goto error;
@@ -2082,6 +2084,7 @@ static pl_tex get_feature_map(struct pass_state *pass)
     ok = pl_dispatch_finish(rr->dp, pl_dispatch_params(
         .shader = &sh,
         .target = out_tex,
+        .no_compute = pass->params->no_compute,
     ));
     if (!ok)
         goto error;
@@ -2263,6 +2266,7 @@ static bool pass_error_diffusion(struct pass_state *pass, pl_shader *sh,
     bool ok = pl_dispatch_finish(rr->dp, pl_dispatch_params(
         .shader = sh,
         .target = edpars.input_tex,
+        .no_compute = pass->params->no_compute,
     ));
 
     if (ok) {
@@ -2303,7 +2307,7 @@ static void clear_target(pl_renderer rr, const struct pl_frame *target,
         pl_frame_clear_rgba(rr->gpu, target, CLEAR_COL(params));
         break;
     case PL_CLEAR_TILES:
-        pl_frame_clear_tiles(rr->gpu, target, params->tile_colors, params->tile_size);
+        pl_frame_clear_tiles(rr->gpu, target, params->tile_colors, params->tile_size, params->no_compute);
         break;
     case PL_CLEAR_SKIP: break;
     case PL_CLEAR_MODE_COUNT: pl_unreachable();
@@ -2622,6 +2626,7 @@ static bool pass_output_target(struct pass_state *pass)
             .target = plane->texture,
             .blend_params = params->blend_params,
             .rect = plane_rect,
+            .no_compute = pass->params->no_compute,
         ));
 
         if (!ok)
@@ -3540,6 +3545,7 @@ bool pl_render_image_mix(pl_renderer rr, const struct pl_frame_mix *images,
             ok = pl_dispatch_finish(rr->dp, pl_dispatch_params(
                 .shader = &inter_pass.img.sh,
                 .target = f->tex,
+                .no_compute = params->no_compute,
             ));
             if (!ok)
                 goto inter_pass_error;
@@ -3785,7 +3791,7 @@ bool pl_frame_is_cropped(const struct pl_frame *frame)
 }
 
 void pl_frame_clear_tiles(pl_gpu gpu, const struct pl_frame *frame,
-                          const float tile_colors[2][3], int tile_size)
+                          const float tile_colors[2][3], int tile_size, bool no_compute)
 {
     struct pl_color_repr repr = frame->repr;
     pl_transform3x3 tr = pl_color_repr_decode(&repr, NULL);
@@ -3832,6 +3838,7 @@ void pl_frame_clear_tiles(pl_gpu gpu, const struct pl_frame *frame,
         pl_dispatch_finish(dp, pl_dispatch_params(
             .shader = &sh,
             .target = plane->texture,
+            .no_compute = no_compute,
         ));
     }
 }