i965/gen8: add debug code to show FS disasm with jump locations
[mesa.git] / src / mesa / drivers / dri / i965 / intel_buffer_objects.c
index d286a9fef608a8991ee8b710dd21fed56022a535..96dacde04fba60e800e7680f40e225bb725cf9c1 100644 (file)
@@ -82,7 +82,8 @@ brw_bo_map_gtt(struct brw_context *brw, drm_intel_bo *bo, const char *bo_name)
 }
 
 static GLboolean
-intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj);
+intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj,
+                      gl_map_buffer_index index);
 
 static void
 intel_bufferobj_mark_gpu_usage(struct intel_buffer_object *intel_obj,
@@ -159,8 +160,7 @@ intel_bufferobj_free(struct gl_context * ctx, struct gl_buffer_object *obj)
     * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
     * (though it does if you call glDeleteBuffers)
     */
-   if (obj->Pointer)
-      intel_bufferobj_unmap(ctx, obj);
+   _mesa_buffer_unmap_all_mappings(ctx, obj);
 
    drm_intel_bo_unreference(intel_obj->buffer);
    free(intel_obj);
@@ -182,7 +182,9 @@ intel_bufferobj_data(struct gl_context * ctx,
                      GLenum target,
                      GLsizeiptrARB size,
                      const GLvoid * data,
-                     GLenum usage, struct gl_buffer_object *obj)
+                     GLenum usage,
+                     GLbitfield storageFlags,
+                     struct gl_buffer_object *obj)
 {
    struct brw_context *brw = brw_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
@@ -193,8 +195,10 @@ intel_bufferobj_data(struct gl_context * ctx,
 
    intel_obj->Base.Size = size;
    intel_obj->Base.Usage = usage;
+   intel_obj->Base.StorageFlags = storageFlags;
 
-   assert(!obj->Pointer); /* Mesa should have unmapped it */
+   assert(!obj->Mappings[MAP_USER].Pointer); /* Mesa should have unmapped it */
+   assert(!obj->Mappings[MAP_INTERNAL].Pointer);
 
    if (intel_obj->buffer != NULL)
       release_buffer(intel_obj);
@@ -348,7 +352,8 @@ intel_bufferobj_get_subdata(struct gl_context * ctx,
 static void *
 intel_bufferobj_map_range(struct gl_context * ctx,
                          GLintptr offset, GLsizeiptr length,
-                         GLbitfield access, struct gl_buffer_object *obj)
+                         GLbitfield access, struct gl_buffer_object *obj,
+                          gl_map_buffer_index index)
 {
    struct brw_context *brw = brw_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
@@ -358,12 +363,12 @@ intel_bufferobj_map_range(struct gl_context * ctx,
    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
     * internally uses our functions directly.
     */
-   obj->Offset = offset;
-   obj->Length = length;
-   obj->AccessFlags = access;
+   obj->Mappings[index].Offset = offset;
+   obj->Mappings[index].Length = length;
+   obj->Mappings[index].AccessFlags = access;
 
    if (intel_obj->buffer == NULL) {
-      obj->Pointer = NULL;
+      obj->Mappings[index].Pointer = NULL;
       return NULL;
    }
 
@@ -396,39 +401,40 @@ intel_bufferobj_map_range(struct gl_context * ctx,
     * doesn't require the current contents of that range, make a new
     * BO, and we'll copy what they put in there out at unmap or
     * FlushRange time.
+    *
+    * That is, unless they're looking for a persistent mapping -- we would
+    * need to do blits in the MemoryBarrier call, and it's easier to just do a
+    * GPU stall and do a mapping.
     */
-   if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
+   if (!(access & (GL_MAP_UNSYNCHRONIZED_BIT | GL_MAP_PERSISTENT_BIT)) &&
        (access & GL_MAP_INVALIDATE_RANGE_BIT) &&
        drm_intel_bo_busy(intel_obj->buffer)) {
       /* Ensure that the base alignment of the allocation meets the alignment
        * guarantees the driver has advertised to the application.
        */
       const unsigned alignment = ctx->Const.MinMapBufferAlignment;
-      const unsigned extra = (uintptr_t) offset % alignment;
 
-      if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
-         intel_obj->range_map_buffer = _mesa_align_malloc(length + extra,
+      intel_obj->map_extra[index] = (uintptr_t) offset % alignment;
+      intel_obj->range_map_bo[index] = drm_intel_bo_alloc(brw->bufmgr,
+                                                          "BO blit temp",
+                                                          length +
+                                                          intel_obj->map_extra[index],
                                                           alignment);
-         obj->Pointer = intel_obj->range_map_buffer + extra;
+      if (brw->has_llc) {
+         drm_intel_bo_map(intel_obj->range_map_bo[index],
+                          (access & GL_MAP_WRITE_BIT) != 0);
       } else {
-        intel_obj->range_map_bo = drm_intel_bo_alloc(brw->bufmgr,
-                                                     "range map",
-                                                      length + extra,
-                                                      alignment);
-        if (!(access & GL_MAP_READ_BIT)) {
-           drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
-        } else {
-           drm_intel_bo_map(intel_obj->range_map_bo,
-                            (access & GL_MAP_WRITE_BIT) != 0);
-        }
-        obj->Pointer = intel_obj->range_map_bo->virtual + extra;
+         drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo[index]);
       }
-      return obj->Pointer;
+      obj->Mappings[index].Pointer =
+         intel_obj->range_map_bo[index]->virtual + intel_obj->map_extra[index];
+      return obj->Mappings[index].Pointer;
    }
 
    if (access & GL_MAP_UNSYNCHRONIZED_BIT)
       drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
-   else if (!(access & GL_MAP_READ_BIT)) {
+   else if (!brw->has_llc && (!(access & GL_MAP_READ_BIT) ||
+                              (access & GL_MAP_PERSISTENT_BIT))) {
       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
       intel_bufferobj_mark_inactive(intel_obj);
    } else {
@@ -436,8 +442,8 @@ intel_bufferobj_map_range(struct gl_context * ctx,
       intel_bufferobj_mark_inactive(intel_obj);
    }
 
-   obj->Pointer = intel_obj->buffer->virtual + offset;
-   return obj->Pointer;
+   obj->Mappings[index].Pointer = intel_obj->buffer->virtual + offset;
+   return obj->Mappings[index].Pointer;
 }
 
 /**
@@ -456,36 +462,56 @@ intel_bufferobj_map_range(struct gl_context * ctx,
 static void
 intel_bufferobj_flush_mapped_range(struct gl_context *ctx,
                                   GLintptr offset, GLsizeiptr length,
-                                  struct gl_buffer_object *obj)
+                                  struct gl_buffer_object *obj,
+                                   gl_map_buffer_index index)
 {
    struct brw_context *brw = brw_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
-   drm_intel_bo *temp_bo;
+   GLbitfield access = obj->Mappings[index].AccessFlags;
 
-   /* Unless we're in the range map using a temporary system buffer,
-    * there's no work to do.
+   assert(access & GL_MAP_FLUSH_EXPLICIT_BIT);
+
+   /* If we gave a direct mapping of the buffer instead of using a temporary,
+    * then there's nothing to do.
     */
-   if (intel_obj->range_map_buffer == NULL)
+   if (intel_obj->range_map_bo[index] == NULL)
       return;
 
    if (length == 0)
       return;
 
-   temp_bo = drm_intel_bo_alloc(brw->bufmgr, "range map flush", length, 64);
-
-   /* Use obj->Pointer instead of intel_obj->range_map_buffer because the
-    * former points to the actual mapping while the latter may be offset to
-    * meet alignment guarantees.
+   /* Note that we're not unmapping our buffer while executing the blit.  We
+    * need to have a mapping still at the end of this call, since the user
+    * gets to make further modifications and glFlushMappedBufferRange() calls.
+    * This is safe, because:
+    *
+    * - On LLC platforms, we're using a CPU mapping that's coherent with the
+    *   GPU (except for the render caches), so the kernel doesn't need to do
+    *   any flushing work for us except for what happens at batch exec time
+    *   anyway.
+    *
+    * - On non-LLC platforms, we're using a GTT mapping that writes directly
+    *   to system memory (except for the chipset cache that gets flushed at
+    *   batch exec time).
+    *
+    * In both cases we don't need to stall for the previous blit to complete
+    * so we can re-map (and we definitely don't want to, since that would be
+    * slow): If the user edits a part of their buffer that's previously been
+    * blitted, then our lack of synchoronization is fine, because either
+    * they'll get some too-new data in the first blit and not do another blit
+    * of that area (but in that case the results are undefined), or they'll do
+    * another blit of that area and the complete newer data will land the
+    * second time.
     */
-   drm_intel_bo_subdata(temp_bo, 0, length, obj->Pointer);
-
    intel_emit_linear_blit(brw,
-                         intel_obj->buffer, obj->Offset + offset,
-                         temp_bo, 0,
+                         intel_obj->buffer,
+                          obj->Mappings[index].Offset + offset,
+                         intel_obj->range_map_bo[index],
+                          intel_obj->map_extra[index] + offset,
                          length);
-   intel_bufferobj_mark_gpu_usage(intel_obj, obj->Offset + offset, length);
-
-   drm_intel_bo_unreference(temp_bo);
+   intel_bufferobj_mark_gpu_usage(intel_obj,
+                                  obj->Mappings[index].Offset + offset,
+                                  length);
 }
 
 
@@ -495,32 +521,26 @@ intel_bufferobj_flush_mapped_range(struct gl_context *ctx,
  * Implements glUnmapBuffer().
  */
 static GLboolean
-intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
+intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj,
+                      gl_map_buffer_index index)
 {
    struct brw_context *brw = brw_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 
    assert(intel_obj);
-   assert(obj->Pointer);
-   if (intel_obj->range_map_buffer != NULL) {
-      /* Since we've emitted some blits to buffers that will (likely) be used
-       * in rendering operations in other cache domains in this batch, emit a
-       * flush.  Once again, we wish for a domain tracker in libdrm to cover
-       * usage inside of a batchbuffer.
-       */
-      intel_batchbuffer_emit_mi_flush(brw);
-      _mesa_align_free(intel_obj->range_map_buffer);
-      intel_obj->range_map_buffer = NULL;
-   } else if (intel_obj->range_map_bo != NULL) {
-      const unsigned extra = obj->Pointer - intel_obj->range_map_bo->virtual;
-
-      drm_intel_bo_unmap(intel_obj->range_map_bo);
-
-      intel_emit_linear_blit(brw,
-                            intel_obj->buffer, obj->Offset,
-                            intel_obj->range_map_bo, extra,
-                            obj->Length);
-      intel_bufferobj_mark_gpu_usage(intel_obj, obj->Offset, obj->Length);
+   assert(obj->Mappings[index].Pointer);
+   if (intel_obj->range_map_bo[index] != NULL) {
+      drm_intel_bo_unmap(intel_obj->range_map_bo[index]);
+
+      if (!(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT)) {
+         intel_emit_linear_blit(brw,
+                                intel_obj->buffer, obj->Mappings[index].Offset,
+                                intel_obj->range_map_bo[index],
+                                intel_obj->map_extra[index],
+                                obj->Mappings[index].Length);
+         intel_bufferobj_mark_gpu_usage(intel_obj, obj->Mappings[index].Offset,
+                                        obj->Mappings[index].Length);
+      }
 
       /* Since we've emitted some blits to buffers that will (likely) be used
        * in rendering operations in other cache domains in this batch, emit a
@@ -529,14 +549,14 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
        */
       intel_batchbuffer_emit_mi_flush(brw);
 
-      drm_intel_bo_unreference(intel_obj->range_map_bo);
-      intel_obj->range_map_bo = NULL;
+      drm_intel_bo_unreference(intel_obj->range_map_bo[index]);
+      intel_obj->range_map_bo[index] = NULL;
    } else if (intel_obj->buffer != NULL) {
       drm_intel_bo_unmap(intel_obj->buffer);
    }
-   obj->Pointer = NULL;
-   obj->Offset = 0;
-   obj->Length = 0;
+   obj->Mappings[index].Pointer = NULL;
+   obj->Mappings[index].Offset = 0;
+   obj->Mappings[index].Length = 0;
 
    return true;
 }