tu: Support softfloat64

DOOM Eternal uses fp64 without checking. Don't expose it, but as a
workaround lower the fp64 operations to software so that we don't choke
on them, similar to anv.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38088>

src/freedreno/vulkan/meson.build

@@ -62,6 +62,21 @@ libtu_files += custom_target(
   depfile : 'float32_spv.h.d',
 )
 
+libtu_files += custom_target(
+  'float64_spv.h',
+  input : [glsl2spirv, float64_glsl_file],
+  output : 'float64_spv.h',
+  command : [
+    prog_python, '@INPUT@', '@OUTPUT@',
+    prog_glslang,
+    '--vn', 'float64_spv_source',
+    '--glsl-version', '450',
+    '-Olib',
+    glslang_depfile,
+  ],
+  depfile : 'float64_spv.h.d',
+)
+
 subdir('bvh')
 
 libtu_includes = [

src/freedreno/vulkan/tu_device.cc

@@ -3144,7 +3144,7 @@ tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
 
    vk_meta_device_finish(&device->vk, &device->meta);
 
-   tu_destroy_softfloat32(device);
+   tu_destroy_softfloat(device);
 
    ir3_compiler_destroy(device->compiler);
 

src/freedreno/vulkan/tu_device.h

@@ -322,6 +322,7 @@ struct tu_device
    struct vk_meta_device meta;
 
    struct nir_shader *float32_shader;
+   struct nir_shader *float64_shader;
    mtx_t softfloat_mutex;
 
    radix_sort_vk_t *radix_sort;

src/freedreno/vulkan/tu_shader.cc

@@ -129,6 +129,8 @@ static const uint32_t float32_spv[] = {
 #include "float32_spv.h"
 };
 
+#include "float64_spv.h"
+
 void
 tu_init_softfloat32(struct tu_device *dev)
 {
@@ -144,10 +146,26 @@ tu_init_softfloat32(struct tu_device *dev)
 }
 
 void
-tu_destroy_softfloat32(struct tu_device *dev)
+tu_init_softfloat64(struct tu_device *dev)
+{
+   if (dev->float64_shader)
+      return;
+
+   mtx_lock(&dev->softfloat_mutex);
+   if (!dev->float64_shader) {
+      dev->float64_shader = tu_spirv_to_nir_library(dev, float64_spv_source,
+                                                    ARRAY_SIZE(float64_spv_source));
+   }
+   mtx_unlock(&dev->softfloat_mutex);
+}
+
+void
+tu_destroy_softfloat(struct tu_device *dev)
 {
    if (dev->float32_shader)
       ralloc_free(dev->float32_shader);
+   if (dev->float64_shader)
+      ralloc_free(dev->float64_shader);
 }
 
 static void
@@ -162,6 +180,19 @@ tu_nir_lower_softfloat32(struct tu_device *dev, nir_shader *nir)
    ir3_optimize_loop(dev->compiler, &optimize_options, nir);
 }
 
+static void
+tu_nir_lower_softfloat64(struct tu_device *dev, nir_shader *nir)
+{
+   tu_init_softfloat64(dev);
+
+   NIR_PASS(_, nir, nir_lower_doubles, dev->float64_shader,
+            nir_lower_fp64_full_software);
+
+   /* Cleanup the result before linking to minimize shader size. */
+   struct ir3_optimize_options optimize_options = {};
+   ir3_optimize_loop(dev->compiler, &optimize_options, nir);
+}
+
 nir_shader *
 tu_spirv_to_nir(struct tu_device *dev,
                 void *mem_ctx,
@@ -252,6 +283,10 @@ tu_spirv_to_nir(struct tu_device *dev,
       tu_nir_lower_softfloat32(dev, nir);
    }
 
+   if (nir->info.bit_sizes_float & 64) {
+      tu_nir_lower_softfloat64(dev, nir);
+   }
+
    return nir;
 }
 

src/freedreno/vulkan/tu_shader.h

@@ -134,7 +134,7 @@ struct tu_shader_key {
 extern const struct vk_pipeline_cache_object_ops tu_shader_ops;
 
 void
-tu_destroy_softfloat32(struct tu_device *device);
+tu_destroy_softfloat(struct tu_device *device);
 
 bool
 tu_nir_lower_multiview(nir_shader *nir, uint32_t mask, struct tu_device *dev);