#include "util/u_memory.h"
#include "util/u_inlines.h"
-#include "util/u_format.h"
-#include "util/u_format_s3tc.h"
+#include "util/format/u_format.h"
+#include "util/format/u_format_s3tc.h"
#include "util/u_string.h"
+#include "os/os_mman.h"
+#include "util/os_time.h"
+
#include <stdio.h>
#include <errno.h>
#include <stdlib.h>
-#include "nouveau/nouveau_bo.h"
-#include "nouveau/nouveau_mm.h"
+#include <nouveau_drm.h>
+#include <xf86drm.h>
+
#include "nouveau_winsys.h"
#include "nouveau_screen.h"
+#include "nouveau_context.h"
#include "nouveau_fence.h"
+#include "nouveau_mm.h"
+#include "nouveau_buffer.h"
+
+#include <compiler/glsl_types.h>
/* XXX this should go away */
-#include "state_tracker/drm_driver.h"
-#include "util/u_simple_screen.h"
+#include "frontend/drm_driver.h"
-#include "nouveau_drmif.h"
+/* Even though GPUs might allow addresses with more bits, some engines do not.
+ * Stick with 40 for compatibility.
+ */
+#define NV_GENERIC_VM_LIMIT_SHIFT 39
int nouveau_mesa_debug = 0;
static const char *
nouveau_screen_get_name(struct pipe_screen *pscreen)
{
- struct nouveau_device *dev = nouveau_screen(pscreen)->device;
- static char buffer[128];
+ struct nouveau_device *dev = nouveau_screen(pscreen)->device;
+ static char buffer[128];
- util_snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
- return buffer;
+ snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
+ return buffer;
}
static const char *
nouveau_screen_get_vendor(struct pipe_screen *pscreen)
{
- return "nouveau";
+ return "nouveau";
}
-
-
-struct nouveau_bo *
-nouveau_screen_bo_new(struct pipe_screen *pscreen, unsigned alignment,
- unsigned usage, unsigned bind, unsigned size)
-{
- struct nouveau_device *dev = nouveau_screen(pscreen)->device;
- struct nouveau_bo *bo = NULL;
- uint32_t flags = NOUVEAU_BO_MAP, tile_mode = 0, tile_flags = 0;
- int ret;
-
- if (bind & PIPE_BIND_VERTEX_BUFFER)
- flags |= nouveau_screen(pscreen)->vertex_buffer_flags;
- else if (bind & PIPE_BIND_INDEX_BUFFER)
- flags |= nouveau_screen(pscreen)->index_buffer_flags;
-
- if (bind & (PIPE_BIND_RENDER_TARGET |
- PIPE_BIND_DEPTH_STENCIL |
- PIPE_BIND_SCANOUT |
- PIPE_BIND_DISPLAY_TARGET |
- PIPE_BIND_SAMPLER_VIEW))
- {
- /* TODO: this may be incorrect or suboptimal */
- if (!(bind & PIPE_BIND_SCANOUT))
- flags |= NOUVEAU_BO_GART;
- if (usage != PIPE_USAGE_DYNAMIC)
- flags |= NOUVEAU_BO_VRAM;
-
- if (dev->chipset == 0x50 || dev->chipset >= 0x80) {
- if (bind & PIPE_BIND_DEPTH_STENCIL)
- tile_flags = 0x2800;
- else
- tile_flags = 0x7000;
- }
- }
-
- ret = nouveau_bo_new_tile(dev, flags, alignment, size,
- tile_mode, tile_flags, &bo);
- if (ret)
- return NULL;
-
- return bo;
-}
-
-void *
-nouveau_screen_bo_map(struct pipe_screen *pscreen,
- struct nouveau_bo *bo,
- unsigned map_flags)
+static const char *
+nouveau_screen_get_device_vendor(struct pipe_screen *pscreen)
{
- int ret;
-
- ret = nouveau_bo_map(bo, map_flags);
- if (ret) {
- debug_printf("map failed: %d\n", ret);
- return NULL;
- }
-
- return bo->map;
+ return "NVIDIA";
}
-void *
-nouveau_screen_bo_map_range(struct pipe_screen *pscreen, struct nouveau_bo *bo,
- unsigned offset, unsigned length, unsigned flags)
+static uint64_t
+nouveau_screen_get_timestamp(struct pipe_screen *pscreen)
{
- int ret;
-
- ret = nouveau_bo_map_range(bo, offset, length, flags);
- if (ret) {
- nouveau_bo_unmap(bo);
- if (!(flags & NOUVEAU_BO_NOWAIT) || ret != -EBUSY)
- debug_printf("map_range failed: %d\n", ret);
- return NULL;
- }
+ int64_t cpu_time = os_time_get() * 1000;
- return (char *)bo->map - offset; /* why gallium? why? */
-}
+ /* getparam of PTIMER_TIME takes about x10 as long (several usecs) */
-void
-nouveau_screen_bo_map_flush_range(struct pipe_screen *pscreen, struct nouveau_bo *bo,
- unsigned offset, unsigned length)
-{
- nouveau_bo_map_flush(bo, offset, length);
+ return cpu_time + nouveau_screen(pscreen)->cpu_gpu_time_delta;
}
-void
-nouveau_screen_bo_unmap(struct pipe_screen *pscreen, struct nouveau_bo *bo)
+static struct disk_cache *
+nouveau_screen_get_disk_shader_cache(struct pipe_screen *pscreen)
{
- nouveau_bo_unmap(bo);
-}
-
-void
-nouveau_screen_bo_release(struct pipe_screen *pscreen, struct nouveau_bo *bo)
-{
- nouveau_bo_ref(NULL, &bo);
+ return nouveau_screen(pscreen)->disk_shader_cache;
}
static void
nouveau_screen_fence_ref(struct pipe_screen *pscreen,
- struct pipe_fence_handle **ptr,
- struct pipe_fence_handle *pfence)
-{
- nouveau_fence_ref(nouveau_fence(pfence), (struct nouveau_fence **)ptr);
-}
-
-static boolean
-nouveau_screen_fence_signalled(struct pipe_screen *screen,
- struct pipe_fence_handle *pfence)
+ struct pipe_fence_handle **ptr,
+ struct pipe_fence_handle *pfence)
{
- return nouveau_fence_signalled(nouveau_fence(pfence));
+ nouveau_fence_ref(nouveau_fence(pfence), (struct nouveau_fence **)ptr);
}
-static boolean
+static bool
nouveau_screen_fence_finish(struct pipe_screen *screen,
- struct pipe_fence_handle *pfence,
+ struct pipe_context *ctx,
+ struct pipe_fence_handle *pfence,
uint64_t timeout)
{
- return nouveau_fence_wait(nouveau_fence(pfence));
+ if (!timeout)
+ return nouveau_fence_signalled(nouveau_fence(pfence));
+
+ return nouveau_fence_wait(nouveau_fence(pfence), NULL);
}
struct nouveau_bo *
nouveau_screen_bo_from_handle(struct pipe_screen *pscreen,
- struct winsys_handle *whandle,
- unsigned *out_stride)
+ struct winsys_handle *whandle,
+ unsigned *out_stride)
{
- struct nouveau_device *dev = nouveau_screen(pscreen)->device;
- struct nouveau_bo *bo = 0;
- int ret;
-
- ret = nouveau_bo_handle_ref(dev, whandle->handle, &bo);
- if (ret) {
- debug_printf("%s: ref name 0x%08x failed with %d\n",
- __FUNCTION__, whandle->handle, ret);
- return NULL;
- }
-
- *out_stride = whandle->stride;
- return bo;
+ struct nouveau_device *dev = nouveau_screen(pscreen)->device;
+ struct nouveau_bo *bo = 0;
+ int ret;
+
+ if (whandle->offset != 0) {
+ debug_printf("%s: attempt to import unsupported winsys offset %d\n",
+ __FUNCTION__, whandle->offset);
+ return NULL;
+ }
+
+ if (whandle->type != WINSYS_HANDLE_TYPE_SHARED &&
+ whandle->type != WINSYS_HANDLE_TYPE_FD) {
+ debug_printf("%s: attempt to import unsupported handle type %d\n",
+ __FUNCTION__, whandle->type);
+ return NULL;
+ }
+
+ if (whandle->type == WINSYS_HANDLE_TYPE_SHARED)
+ ret = nouveau_bo_name_ref(dev, whandle->handle, &bo);
+ else
+ ret = nouveau_bo_prime_handle_ref(dev, whandle->handle, &bo);
+
+ if (ret) {
+ debug_printf("%s: ref name 0x%08x failed with %d\n",
+ __FUNCTION__, whandle->handle, ret);
+ return NULL;
+ }
+
+ *out_stride = whandle->stride;
+ return bo;
}
-boolean
+bool
nouveau_screen_bo_get_handle(struct pipe_screen *pscreen,
- struct nouveau_bo *bo,
- unsigned stride,
- struct winsys_handle *whandle)
+ struct nouveau_bo *bo,
+ unsigned stride,
+ struct winsys_handle *whandle)
{
- whandle->stride = stride;
-
- if (whandle->type == DRM_API_HANDLE_TYPE_SHARED) {
- return nouveau_bo_handle_get(bo, &whandle->handle) == 0;
- } else if (whandle->type == DRM_API_HANDLE_TYPE_KMS) {
- whandle->handle = bo->handle;
- return TRUE;
- } else {
- return FALSE;
- }
+ whandle->stride = stride;
+
+ if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
+ return nouveau_bo_name_get(bo, &whandle->handle) == 0;
+ } else if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
+ whandle->handle = bo->handle;
+ return true;
+ } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
+ return nouveau_bo_set_prime(bo, (int *)&whandle->handle) == 0;
+ } else {
+ return false;
+ }
+}
+
+static void
+nouveau_disk_cache_create(struct nouveau_screen *screen)
+{
+ struct mesa_sha1 ctx;
+ unsigned char sha1[20];
+ char cache_id[20 * 2 + 1];
+ uint64_t driver_flags = 0;
+
+ _mesa_sha1_init(&ctx);
+ if (!disk_cache_get_function_identifier(nouveau_disk_cache_create,
+ &ctx))
+ return;
+
+ _mesa_sha1_final(&ctx, sha1);
+ disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
+
+ if (screen->prefer_nir)
+ driver_flags |= NOUVEAU_SHADER_CACHE_FLAGS_IR_NIR;
+ else
+ driver_flags |= NOUVEAU_SHADER_CACHE_FLAGS_IR_TGSI;
+
+ screen->disk_shader_cache =
+ disk_cache_create(nouveau_screen_get_name(&screen->base),
+ cache_id, driver_flags);
+}
+
+static void*
+reserve_vma(uintptr_t start, uint64_t reserved_size)
+{
+ void *reserved = os_mmap((void*)start, reserved_size, PROT_NONE,
+ MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+ if (reserved == MAP_FAILED)
+ return NULL;
+ return reserved;
}
int
nouveau_screen_init(struct nouveau_screen *screen, struct nouveau_device *dev)
{
- struct pipe_screen *pscreen = &screen->base;
- int ret;
+ struct pipe_screen *pscreen = &screen->base;
+ struct nv04_fifo nv04_data = { .vram = 0xbeef0201, .gart = 0xbeef0202 };
+ struct nvc0_fifo nvc0_data = { };
+ uint64_t time;
+ int size, ret;
+ void *data;
+ union nouveau_bo_config mm_config;
+
+ char *nv_dbg = getenv("NOUVEAU_MESA_DEBUG");
+ if (nv_dbg)
+ nouveau_mesa_debug = atoi(nv_dbg);
+
+ if (dev->chipset < 0x140)
+ screen->prefer_nir = debug_get_bool_option("NV50_PROG_USE_NIR", false);
+ else
+ screen->prefer_nir = true;
+
+ screen->force_enable_cl = debug_get_bool_option("NOUVEAU_ENABLE_CL", false);
+ if (screen->force_enable_cl)
+ glsl_type_singleton_init_or_ref();
+
+ /* These must be set before any failure is possible, as the cleanup
+ * paths assume they're responsible for deleting them.
+ */
+ screen->drm = nouveau_drm(&dev->object);
+ screen->device = dev;
+
+ /*
+ * this is initialized to 1 in nouveau_drm_screen_create after screen
+ * is fully constructed and added to the global screen list.
+ */
+ screen->refcount = -1;
+
+ if (dev->chipset < 0xc0) {
+ data = &nv04_data;
+ size = sizeof(nv04_data);
+ } else {
+ data = &nvc0_data;
+ size = sizeof(nvc0_data);
+ }
+
+ screen->has_svm = false;
+ /* we only care about HMM with OpenCL enabled */
+ if (dev->chipset > 0x130 && screen->force_enable_cl) {
+ /* Before being able to enable SVM we need to carve out some memory for
+ * driver bo allocations. Let's just base the size on the available VRAM.
+ *
+ * 40 bit is the biggest we care about and for 32 bit systems we don't
+ * want to allocate all of the available memory either.
+ *
+ * Also we align the size we want to reserve to the next POT to make use
+ * of hugepages.
+ */
+ const int vram_shift = util_logbase2_ceil64(dev->vram_size);
+ const int limit_bit =
+ MIN2(sizeof(void*) * 8 - 1, NV_GENERIC_VM_LIMIT_SHIFT);
+ screen->svm_cutout_size =
+ BITFIELD64_BIT(MIN2(sizeof(void*) == 4 ? 26 : NV_GENERIC_VM_LIMIT_SHIFT, vram_shift));
+
+ size_t start = screen->svm_cutout_size;
+ do {
+ screen->svm_cutout = reserve_vma(start, screen->svm_cutout_size);
+ if (!screen->svm_cutout) {
+ start += screen->svm_cutout_size;
+ continue;
+ }
+
+ struct drm_nouveau_svm_init svm_args = {
+ .unmanaged_addr = (uint64_t)screen->svm_cutout,
+ .unmanaged_size = screen->svm_cutout_size,
+ };
+
+ ret = drmCommandWrite(screen->drm->fd, DRM_NOUVEAU_SVM_INIT,
+ &svm_args, sizeof(svm_args));
+ screen->has_svm = !ret;
+ if (!screen->has_svm)
+ os_munmap(screen->svm_cutout, screen->svm_cutout_size);
+ break;
+ } while ((start + screen->svm_cutout_size) < BITFIELD64_MASK(limit_bit));
+ }
+
+ /*
+ * Set default VRAM domain if not overridden
+ */
+ if (!screen->vram_domain) {
+ if (dev->vram_size > 0)
+ screen->vram_domain = NOUVEAU_BO_VRAM;
+ else
+ screen->vram_domain = NOUVEAU_BO_GART;
+ }
+
+ ret = nouveau_object_new(&dev->object, 0, NOUVEAU_FIFO_CHANNEL_CLASS,
+ data, size, &screen->channel);
+ if (ret)
+ goto err;
+
+ ret = nouveau_client_new(screen->device, &screen->client);
+ if (ret)
+ goto err;
+ ret = nouveau_pushbuf_new(screen->client, screen->channel,
+ 4, 512 * 1024, 1,
+ &screen->pushbuf);
+ if (ret)
+ goto err;
+
+ /* getting CPU time first appears to be more accurate */
+ screen->cpu_gpu_time_delta = os_time_get();
+
+ ret = nouveau_getparam(dev, NOUVEAU_GETPARAM_PTIMER_TIME, &time);
+ if (!ret)
+ screen->cpu_gpu_time_delta = time - screen->cpu_gpu_time_delta * 1000;
+
+ pscreen->get_name = nouveau_screen_get_name;
+ pscreen->get_vendor = nouveau_screen_get_vendor;
+ pscreen->get_device_vendor = nouveau_screen_get_device_vendor;
+ pscreen->get_disk_shader_cache = nouveau_screen_get_disk_shader_cache;
+
+ pscreen->get_timestamp = nouveau_screen_get_timestamp;
+
+ pscreen->fence_reference = nouveau_screen_fence_ref;
+ pscreen->fence_finish = nouveau_screen_fence_finish;
+
+ nouveau_disk_cache_create(screen);
+
+ screen->transfer_pushbuf_threshold = 192;
+ screen->lowmem_bindings = PIPE_BIND_GLOBAL; /* gallium limit */
+ screen->vidmem_bindings =
+ PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL |
+ PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
+ PIPE_BIND_CURSOR |
+ PIPE_BIND_SAMPLER_VIEW |
+ PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE |
+ PIPE_BIND_COMPUTE_RESOURCE |
+ PIPE_BIND_GLOBAL;
+ screen->sysmem_bindings =
+ PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_STREAM_OUTPUT |
+ PIPE_BIND_COMMAND_ARGS_BUFFER;
+
+ memset(&mm_config, 0, sizeof(mm_config));
+
+ screen->mm_GART = nouveau_mm_create(dev,
+ NOUVEAU_BO_GART | NOUVEAU_BO_MAP,
+ &mm_config);
+ screen->mm_VRAM = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, &mm_config);
+ return 0;
+
+err:
+ if (screen->svm_cutout)
+ os_munmap(screen->svm_cutout, screen->svm_cutout_size);
+ return ret;
+}
+
+void
+nouveau_screen_fini(struct nouveau_screen *screen)
+{
+ int fd = screen->drm->fd;
- char *nv_dbg = getenv("NOUVEAU_MESA_DEBUG");
- if (nv_dbg)
- nouveau_mesa_debug = atoi(nv_dbg);
+ if (screen->force_enable_cl)
+ glsl_type_singleton_decref();
+ if (screen->has_svm)
+ os_munmap(screen->svm_cutout, screen->svm_cutout_size);
- ret = nouveau_channel_alloc(dev, 0xbeef0201, 0xbeef0202,
- 512*1024, &screen->channel);
- if (ret)
- return ret;
- screen->device = dev;
+ nouveau_mm_destroy(screen->mm_GART);
+ nouveau_mm_destroy(screen->mm_VRAM);
- pscreen->get_name = nouveau_screen_get_name;
- pscreen->get_vendor = nouveau_screen_get_vendor;
+ nouveau_pushbuf_del(&screen->pushbuf);
- pscreen->fence_reference = nouveau_screen_fence_ref;
- pscreen->fence_signalled = nouveau_screen_fence_signalled;
- pscreen->fence_finish = nouveau_screen_fence_finish;
+ nouveau_client_del(&screen->client);
+ nouveau_object_del(&screen->channel);
- util_format_s3tc_init();
+ nouveau_device_del(&screen->device);
+ nouveau_drm_del(&screen->drm);
+ close(fd);
- screen->mm_GART = nouveau_mm_create(dev,
- NOUVEAU_BO_GART | NOUVEAU_BO_MAP,
- 0x000);
- screen->mm_VRAM = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, 0x000);
- return 0;
+ disk_cache_destroy(screen->disk_shader_cache);
}
-void
-nouveau_screen_fini(struct nouveau_screen *screen)
+static void
+nouveau_set_debug_callback(struct pipe_context *pipe,
+ const struct pipe_debug_callback *cb)
{
- nouveau_mm_destroy(screen->mm_GART);
- nouveau_mm_destroy(screen->mm_VRAM);
-
- nouveau_channel_free(&screen->channel);
+ struct nouveau_context *context = nouveau_context(pipe);
- nouveau_device_close(&screen->device);
+ if (cb)
+ context->debug = *cb;
+ else
+ memset(&context->debug, 0, sizeof(context->debug));
}
+void
+nouveau_context_init(struct nouveau_context *context)
+{
+ context->pipe.set_debug_callback = nouveau_set_debug_callback;
+}
projects / mesa.git / blobdiff