iris: ref/unref the GLSL type singleton in screen_create/destroy
[mesa.git] / src / gallium / drivers / iris / iris_screen.c
index f1bb03de8f94ef570ea6d7a025343059a7365e2a..3b391f1db5d309177dd03ee9994347c8289d6c22 100644 (file)
 #include "iris_pipe.h"
 #include "iris_resource.h"
 #include "iris_screen.h"
+#include "compiler/glsl_types.h"
 #include "intel/compiler/brw_compiler.h"
 #include "intel/common/gen_gem.h"
+#include "intel/common/gen_l3_config.h"
 #include "iris_monitor.h"
 
 static void
@@ -80,30 +82,15 @@ iris_get_name(struct pipe_screen *pscreen)
 {
    struct iris_screen *screen = (struct iris_screen *)pscreen;
    static char buf[128];
-   const char *chipset;
+   const char *name = gen_get_device_name(screen->pci_id);
 
-   switch (screen->pci_id) {
-#undef CHIPSET
-#define CHIPSET(id, symbol, str) case id: chipset = str; break;
-#include "pci_ids/i965_pci_ids.h"
-#include "pci_ids/iris_pci_ids.h"
-   default:
-      chipset = "Unknown Intel Chipset";
-      break;
-   }
+   if (!name)
+      name = "Intel Unknown";
 
-   snprintf(buf, sizeof(buf), "Mesa %s", chipset);
+   snprintf(buf, sizeof(buf), "Mesa %s", name);
    return buf;
 }
 
-static uint64_t
-get_aperture_size(int fd)
-{
-   struct drm_i915_gem_get_aperture aperture = {};
-   gen_ioctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture);
-   return aperture.aper_size;
-}
-
 static int
 iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
 {
@@ -123,6 +110,7 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_FRAGMENT_SHADER_DERIVATIVES:
    case PIPE_CAP_VERTEX_SHADER_SATURATE:
    case PIPE_CAP_PRIMITIVE_RESTART:
+   case PIPE_CAP_PRIMITIVE_RESTART_FIXED_INDEX:
    case PIPE_CAP_INDEP_BLEND_ENABLE:
    case PIPE_CAP_INDEP_BLEND_FUNC:
    case PIPE_CAP_RGB_OVERRIDE_DST_ALPHA_BLEND:
@@ -185,6 +173,7 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_TGSI_BALLOT:
    case PIPE_CAP_MULTISAMPLE_Z_RESOLVE:
    case PIPE_CAP_CLEAR_TEXTURE:
+   case PIPE_CAP_CLEAR_SCISSORED:
    case PIPE_CAP_TGSI_VOTE:
    case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
    case PIPE_CAP_TEXTURE_GATHER_SM5:
@@ -204,6 +193,8 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_GL_SPIRV:
    case PIPE_CAP_GL_SPIRV_VARIABLE_POINTERS:
    case PIPE_CAP_DEMOTE_TO_HELPER_INVOCATION:
+   case PIPE_CAP_NATIVE_FENCE_FD:
+   case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
       return true;
    case PIPE_CAP_FBFETCH:
       return BRW_MAX_DRAW_BUFFERS;
@@ -215,6 +206,8 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_FRAGMENT_SHADER_INTERLOCK:
    case PIPE_CAP_ATOMIC_FLOAT_MINMAX:
       return devinfo->gen >= 9;
+   case PIPE_CAP_DEPTH_BOUNDS_TEST:
+      return devinfo->gen >= 12;
    case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
       return 1;
    case PIPE_CAP_MAX_RENDER_TARGETS:
@@ -284,7 +277,7 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
        * flushing, etc.  That's the big cliff apps will care about.
        */
       const unsigned gpu_mappable_megabytes =
-         (screen->aperture_bytes * 3 / 4) / (1024 * 1024);
+         (devinfo->aperture_bytes * 3 / 4) / (1024 * 1024);
 
       const long system_memory_pages = sysconf(_SC_PHYS_PAGES);
       const long system_page_size = sysconf(_SC_PAGE_SIZE);
@@ -320,6 +313,9 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
              PIPE_CONTEXT_PRIORITY_MEDIUM |
              PIPE_CONTEXT_PRIORITY_HIGH;
 
+   case PIPE_CAP_FRONTEND_NOOP:
+      return true;
+
    // XXX: don't hardcode 00:00:02.0 PCI here
    case PIPE_CAP_PCI_GROUP:
       return 0;
@@ -330,6 +326,10 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_PCI_FUNCTION:
       return 0;
 
+   case PIPE_CAP_OPENCL_INTEGER_FUNCTIONS:
+   case PIPE_CAP_INTEGER_MULTIPLY_32X16:
+      return true;
+
    default:
       return u_pipe_screen_get_param_defaults(pscreen, param);
    }
@@ -407,6 +407,9 @@ iris_get_shader_param(struct pipe_screen *pscreen,
       return 1;
    case PIPE_SHADER_CAP_INT64_ATOMICS:
    case PIPE_SHADER_CAP_FP16:
+   case PIPE_SHADER_CAP_FP16_DERIVATIVES:
+   case PIPE_SHADER_CAP_INT16:
+   case PIPE_SHADER_CAP_GLSL_16BIT_CONSTS:
       return 0;
    case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
    case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
@@ -420,7 +423,7 @@ iris_get_shader_param(struct pipe_screen *pscreen,
    case PIPE_SHADER_CAP_PREFERRED_IR:
       return PIPE_SHADER_IR_NIR;
    case PIPE_SHADER_CAP_SUPPORTED_IRS:
-      return 1 << PIPE_SHADER_IR_NIR;
+      return (1 << PIPE_SHADER_IR_NIR);
    case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
    case PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED:
       return 1;
@@ -446,6 +449,7 @@ iris_get_compute_param(struct pipe_screen *pscreen,
    struct iris_screen *screen = (struct iris_screen *)pscreen;
    const struct gen_device_info *devinfo = &screen->devinfo;
 
+   /* Limit max_threads to 64 for the GPGPU_WALKER command. */
    const unsigned max_threads = MIN2(64, devinfo->max_cs_threads);
    const uint32_t max_invocations = 32 * max_threads;
 
@@ -456,9 +460,6 @@ iris_get_compute_param(struct pipe_screen *pscreen,
 } while (0)
 
    switch (param) {
-   case PIPE_COMPUTE_CAP_ADDRESS_BITS:
-      RET((uint32_t []){ 32 });
-
    case PIPE_COMPUTE_CAP_IR_TARGET:
       if (ret)
          strcpy(ret, "gen");
@@ -476,25 +477,25 @@ iris_get_compute_param(struct pipe_screen *pscreen,
 
    case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
       /* MaxComputeWorkGroupInvocations */
+   case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
+      /* MaxComputeVariableGroupInvocations */
       RET((uint64_t []) { max_invocations });
 
    case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
       /* MaxComputeSharedMemorySize */
       RET((uint64_t []) { 64 * 1024 });
 
-   case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
-      RET((uint32_t []) { 1 });
-
    case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
       RET((uint32_t []) { BRW_SUBGROUP_SIZE });
 
+   case PIPE_COMPUTE_CAP_ADDRESS_BITS:
+   case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
    case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
    case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
    case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
    case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
    case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
    case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
-   case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
       // XXX: I think these are for Clover...
       return 0;
 
@@ -518,18 +519,24 @@ iris_get_timestamp(struct pipe_screen *pscreen)
    return result;
 }
 
-static void
-iris_destroy_screen(struct pipe_screen *pscreen)
+void
+iris_screen_destroy(struct iris_screen *screen)
 {
-   struct iris_screen *screen = (struct iris_screen *) pscreen;
+   glsl_type_singleton_decref();
    iris_bo_unreference(screen->workaround_bo);
-   u_transfer_helper_destroy(pscreen->transfer_helper);
-   iris_bufmgr_destroy(screen->bufmgr);
+   u_transfer_helper_destroy(screen->base.transfer_helper);
+   iris_bufmgr_unref(screen->bufmgr);
    disk_cache_destroy(screen->disk_cache);
-   close(screen->fd);
+   close(screen->winsys_fd);
    ralloc_free(screen);
 }
 
+static void
+iris_screen_unref(struct pipe_screen *pscreen)
+{
+   iris_pscreen_unref(pscreen);
+}
+
 static void
 iris_query_memory_info(struct pipe_screen *pscreen,
                        struct pipe_memory_info *info)
@@ -556,27 +563,38 @@ iris_get_disk_shader_cache(struct pipe_screen *pscreen)
 }
 
 static int
-iris_getparam(struct iris_screen *screen, int param, int *value)
+iris_getparam(int fd, int param, int *value)
 {
    struct drm_i915_getparam gp = { .param = param, .value = value };
 
-   if (ioctl(screen->fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1)
+   if (ioctl(fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1)
       return -errno;
 
    return 0;
 }
 
 static int
-iris_getparam_integer(struct iris_screen *screen, int param)
+iris_getparam_integer(int fd, int param)
 {
    int value = -1;
 
-   if (iris_getparam(screen, param, &value) == 0)
+   if (iris_getparam(fd, param, &value) == 0)
       return value;
 
    return -1;
 }
 
+static const struct gen_l3_config *
+iris_get_default_l3_config(const struct gen_device_info *devinfo,
+                           bool compute)
+{
+   bool wants_dc_cache = true;
+   bool has_slm = compute;
+   const struct gen_l3_weights w =
+      gen_get_default_l3_weights(devinfo, wants_dc_cache, has_slm);
+   return gen_get_l3_config(devinfo, w);
+}
+
 static void
 iris_shader_debug_log(void *data, const char *fmt, ...)
 {
@@ -614,28 +632,66 @@ iris_shader_perf_log(void *data, const char *fmt, ...)
    va_end(args);
 }
 
+static void
+iris_detect_kernel_features(struct iris_screen *screen)
+{
+   /* Kernel 5.2+ */
+   if (gen_gem_supports_syncobj_wait(screen->fd))
+      screen->kernel_features |= KERNEL_HAS_WAIT_FOR_SUBMIT;
+}
+
+static bool
+iris_init_identifier_bo(struct iris_screen *screen)
+{
+   void *bo_map;
+
+   bo_map = iris_bo_map(NULL, screen->workaround_bo, MAP_READ | MAP_WRITE);
+   if (!bo_map)
+      return false;
+
+   screen->workaround_bo->kflags |= EXEC_OBJECT_CAPTURE;
+   screen->workaround_address = (struct iris_address) {
+      .bo = screen->workaround_bo,
+      .offset = ALIGN(
+         intel_debug_write_identifiers(bo_map, 4096, "Iris") + 8, 8),
+   };
+
+   iris_bo_unmap(screen->workaround_bo);
+
+   return true;
+}
+
 struct pipe_screen *
 iris_screen_create(int fd, const struct pipe_screen_config *config)
 {
+   /* Here are the i915 features we need for Iris (in chronoligical order) :
+    *    - I915_PARAM_HAS_EXEC_NO_RELOC     (3.10)
+    *    - I915_PARAM_HAS_EXEC_HANDLE_LUT   (3.10)
+    *    - I915_PARAM_HAS_EXEC_BATCH_FIRST  (4.13)
+    *    - I915_PARAM_HAS_EXEC_FENCE_ARRAY  (4.14)
+    *    - I915_PARAM_HAS_CONTEXT_ISOLATION (4.16)
+    *
+    * Checking the last feature availability will include all previous ones.
+    */
+   if (iris_getparam_integer(fd, I915_PARAM_HAS_CONTEXT_ISOLATION) <= 0) {
+      debug_error("Kernel is too old for Iris. Consider upgrading to kernel v4.16.\n");
+      return NULL;
+   }
+
    struct iris_screen *screen = rzalloc(NULL, struct iris_screen);
    if (!screen)
       return NULL;
 
-   screen->fd = fd;
-
    if (!gen_get_device_info_from_fd(fd, &screen->devinfo))
       return NULL;
    screen->pci_id = screen->devinfo.chipset_id;
    screen->no_hw = screen->devinfo.no_hw;
 
+   p_atomic_set(&screen->refcount, 1);
+
    if (screen->devinfo.gen < 8 || screen->devinfo.is_cherryview)
       return NULL;
 
-   screen->aperture_bytes = get_aperture_size(fd);
-
-   if (getenv("INTEL_NO_HW") != NULL)
-      screen->no_hw = true;
-
    bool bo_reuse = false;
    int bo_reuse_mode = driQueryOptioni(config->options, "bo_reuse");
    switch (bo_reuse_mode) {
@@ -646,15 +702,24 @@ iris_screen_create(int fd, const struct pipe_screen_config *config)
       break;
    }
 
-   screen->bufmgr = iris_bufmgr_init(&screen->devinfo, fd, bo_reuse);
+   screen->bufmgr = iris_bufmgr_get_for_fd(&screen->devinfo, fd, bo_reuse);
    if (!screen->bufmgr)
       return NULL;
 
+   screen->fd = iris_bufmgr_get_fd(screen->bufmgr);
+   screen->winsys_fd = fd;
+
+   if (getenv("INTEL_NO_HW") != NULL)
+      screen->no_hw = true;
+
    screen->workaround_bo =
       iris_bo_alloc(screen->bufmgr, "workaround", 4096, IRIS_MEMZONE_OTHER);
    if (!screen->workaround_bo)
       return NULL;
 
+   if (!iris_init_identifier_bo(screen))
+      return NULL;
+
    brw_process_intel_debug_variable();
 
    screen->driconf.dual_color_blend_by_location =
@@ -673,6 +738,10 @@ iris_screen_create(int fd, const struct pipe_screen_config *config)
    screen->compiler->shader_perf_log = iris_shader_perf_log;
    screen->compiler->supports_pull_constants = false;
    screen->compiler->supports_shader_constants = true;
+   screen->compiler->compact_params = false;
+
+   screen->l3_config_3d = iris_get_default_l3_config(&screen->devinfo, false);
+   screen->l3_config_cs = iris_get_default_l3_config(&screen->devinfo, true);
 
    iris_disk_cache_init(screen);
 
@@ -680,15 +749,17 @@ iris_screen_create(int fd, const struct pipe_screen_config *config)
                       sizeof(struct iris_transfer), 64);
 
    screen->subslice_total =
-      iris_getparam_integer(screen, I915_PARAM_SUBSLICE_TOTAL);
+      iris_getparam_integer(screen->fd, I915_PARAM_SUBSLICE_TOTAL);
    assert(screen->subslice_total >= 1);
 
+   iris_detect_kernel_features(screen);
+
    struct pipe_screen *pscreen = &screen->base;
 
    iris_init_screen_fence_functions(pscreen);
    iris_init_screen_resource_functions(pscreen);
 
-   pscreen->destroy = iris_destroy_screen;
+   pscreen->destroy = iris_screen_unref;
    pscreen->get_name = iris_get_name;
    pscreen->get_vendor = iris_get_vendor;
    pscreen->get_device_vendor = iris_get_device_vendor;
@@ -706,5 +777,7 @@ iris_screen_create(int fd, const struct pipe_screen_config *config)
    pscreen->get_driver_query_group_info = iris_get_monitor_group_info;
    pscreen->get_driver_query_info = iris_get_monitor_info;
 
+   glsl_type_singleton_init_or_ref();
+
    return pscreen;
 }