gpu: add pl_{opengl,vulkan,d3d11}_params.no_compute

May be useful for debugging.

meson.build

@@ -12,6 +12,7 @@ project('libplacebo', ['c', 'cpp'],
     7,
     # API version
     {
+      '350': 'add pl_{opengl,vulkan,d3d11}_params.no_compute',
       '349': 'add pl_color_{primaries,system,transfer}_name(s)',
       '348': 'add pl_color_linearize and pl_color_delinearize',
       '347': 'add pl_gpu_limits.max_mapped_vram',

src/d3d11/context.c

@@ -472,7 +472,7 @@ pl_d3d11 pl_d3d11_create(pl_log log, const struct pl_d3d11_params *params)
         PL_MSG(ctx, level, "Using a software adapter");
     }
 
-    d3d11->gpu = pl_gpu_create_d3d11(ctx);
+    d3d11->gpu = pl_gpu_create_d3d11(ctx, params);
     if (!d3d11->gpu)
         goto error;
 

src/d3d11/gpu.c

@@ -387,7 +387,8 @@ static struct pl_gpu_fns pl_fns_d3d11 = {
     .destroy                = d3d11_gpu_destroy,
 };
 
-pl_gpu pl_gpu_create_d3d11(struct d3d11_ctx *ctx)
+pl_gpu pl_gpu_create_d3d11(struct d3d11_ctx *ctx,
+                           const struct pl_d3d11_params *params)
 {
     pl_assert(ctx->dev);
     IDXGIDevice1 *dxgi_dev = NULL;
@@ -506,7 +507,7 @@ pl_gpu pl_gpu_create_d3d11(struct d3d11_ctx *ctx)
     }
 
     if (p->fl >= D3D_FEATURE_LEVEL_11_0) {
-        gpu->glsl.compute = true;
+        gpu->glsl.compute = !params->no_compute;
         gpu->limits.compute_queues = 1;
         // Set `gpu->limits.blittable_1d_3d`, since `pl_tex_blit_compute`, which
         // is used to emulate blits on 11_0 and up, supports 1D and 3D textures

src/d3d11/gpu.h

@@ -29,7 +29,8 @@
 #include "common.h"
 #include "utils.h"
 
-pl_gpu pl_gpu_create_d3d11(struct d3d11_ctx *ctx);
+pl_gpu pl_gpu_create_d3d11(struct d3d11_ctx *ctx,
+                           const struct pl_d3d11_params *params);
 
 // --- pl_gpu internal structs and helpers
 

src/include/libplacebo/d3d11.h

@@ -70,6 +70,11 @@ struct pl_d3d11_params {
     // Also logs IDXGIInfoQueue messages
     bool debug;
 
+    // Disables the use of compute shaders. Some devices/drivers perform better
+    // without them. This may also help prevent image corruption in cases where
+    // the driver is misbehaving. Some features may be disabled if this is set.
+    bool no_compute;
+
     // Extra flags to pass to D3D11CreateDevice (D3D11_CREATE_DEVICE_FLAG).
     // libplacebo should be compatible with any flags passed here.
     UINT flags;

src/include/libplacebo/opengl.h

@@ -72,6 +72,11 @@ struct pl_opengl_params {
     // undesirable when GPU-accelerated processing is expected.
     bool allow_software;
 
+    // Disables the use of compute shaders. Some devices/drivers perform better
+    // without them. This may also help prevent image corruption in cases where
+    // the driver is misbehaving. Some features may be disabled if this is set.
+    bool no_compute;
+
     // Restrict the maximum allowed GLSL version. (Mainly for testing)
     int max_glsl_version;
 

src/include/libplacebo/vulkan.h

@@ -230,6 +230,11 @@ struct pl_vulkan_params {
     // queues for a given QF enabled, regardless of this setting.
     int queue_count;
 
+    // Disables the use of compute shaders. Some devices/drivers perform better
+    // without them. This may also help prevent image corruption in cases where
+    // the driver is misbehaving. Some features may be disabled if this is set.
+    bool no_compute;
+
     // Bitmask of extra queue families to enable. If set, then *all* queue
     // families matching *any* of these flags will be enabled at device
     // creation time. Setting this to VK_QUEUE_FLAG_BITS_MAX_ENUM effectively
@@ -424,6 +429,7 @@ struct pl_vulkan_import_params {
 
     // Restrict specific features to e.g. work around driver bugs, or simply
     // for testing purposes. See `pl_vulkan_params` for a description of these.
+    bool no_compute;
     int max_glsl_version;
     uint32_t max_api_version;
 };

src/opengl/gpu.c

@@ -155,7 +155,7 @@ pl_gpu pl_gpu_create_gl(pl_log log, pl_opengl pl_gl, const struct pl_opengl_para
     }
 
     if (gl_test_ext(gpu, "GL_ARB_compute_shader", 43, 0) && glsl->version >= 420) {
-        glsl->compute = true;
+        glsl->compute = !params->no_compute;
         get(GL_MAX_COMPUTE_SHARED_MEMORY_SIZE, &glsl->max_shmem_size);
         get(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &glsl->max_group_threads);
         for (int i = 0; i < 3; i++)

src/vulkan/context.c

@@ -1366,7 +1366,8 @@ static void unlock_queue(pl_vulkan pl_vk, uint32_t qf, uint32_t qidx)
     vk->unlock_queue(vk->queue_ctx, qf, qidx);
 }
 
-static bool finalize_context(struct pl_vulkan_t *pl_vk, int max_glsl_version)
+static bool finalize_context(struct pl_vulkan_t *pl_vk, int max_glsl_version,
+                             bool no_compute)
 {
     struct vk_ctx *vk = PL_PRIV(pl_vk);
 
@@ -1383,14 +1384,16 @@ static bool finalize_context(struct pl_vulkan_t *pl_vk, int max_glsl_version)
         return false;
 
     // Blacklist / restrict features
+    struct pl_glsl_version *glsl = (struct pl_glsl_version *) &pl_vk->gpu->glsl;
     if (max_glsl_version) {
-        struct pl_glsl_version *glsl = (struct pl_glsl_version *) &pl_vk->gpu->glsl;
         glsl->version = PL_MIN(glsl->version, max_glsl_version);
         glsl->version = PL_MAX(glsl->version, 140); // required for GL_KHR_vulkan_glsl
         PL_INFO(vk, "Restricting GLSL version to %d... new version is %d",
                 max_glsl_version, glsl->version);
     }
 
+    glsl->compute &= !no_compute;
+
     // Expose the resulting vulkan objects
     pl_vk->instance = vk->inst;
     pl_vk->phys_device = vk->physd;
@@ -1522,7 +1525,7 @@ pl_vulkan pl_vulkan_create(pl_log log, const struct pl_vulkan_params *params)
     if (!device_init(vk, params))
         goto error;
 
-    if (!finalize_context(pl_vk, params->max_glsl_version))
+    if (!finalize_context(pl_vk, params->max_glsl_version, params->no_compute))
         goto error;
 
     return pl_vk;
@@ -1691,7 +1694,7 @@ next_qf: ;
         goto error;
     }
 
-    if (!finalize_context(pl_vk, params->max_glsl_version))
+    if (!finalize_context(pl_vk, params->max_glsl_version, params->no_compute))
         goto error;
 
     pl_free(tmp);