#include "main/imports.h"
#include "main/mtypes.h"
#include "main/macros.h"
+#include "main/streaming-load-memcpy.h"
#include "main/bufferobj.h"
+#include "x86/common_x86_asm.h"
#include "brw_context.h"
-#include "intel_blit.h"
+#include "brw_blorp.h"
#include "intel_buffer_objects.h"
#include "intel_batchbuffer.h"
-
-/**
- * Map a buffer object; issue performance warnings if mapping causes stalls.
- *
- * This matches the drm_intel_bo_map API, but takes an additional human-readable
- * name for the buffer object to use in the performance debug message.
- */
-int
-brw_bo_map(struct brw_context *brw,
- drm_intel_bo *bo, int write_enable,
- const char *bo_name)
-{
- if (likely(!brw->perf_debug) || !drm_intel_bo_busy(bo))
- return drm_intel_bo_map(bo, write_enable);
-
- double start_time = get_time();
-
- int ret = drm_intel_bo_map(bo, write_enable);
-
- perf_debug("CPU mapping a busy %s BO stalled and took %.03f ms.\n",
- bo_name, (get_time() - start_time) * 1000);
-
- return ret;
-}
-
-int
-brw_bo_map_gtt(struct brw_context *brw, drm_intel_bo *bo, const char *bo_name)
-{
- if (likely(!brw->perf_debug) || !drm_intel_bo_busy(bo))
- return drm_intel_gem_bo_map_gtt(bo);
-
- double start_time = get_time();
-
- int ret = drm_intel_gem_bo_map_gtt(bo);
-
- perf_debug("GTT mapping a busy %s BO stalled and took %.03f ms.\n",
- bo_name, (get_time() - start_time) * 1000);
-
- return ret;
-}
+#include "intel_tiled_memcpy.h"
static void
mark_buffer_gpu_usage(struct intel_buffer_object *intel_obj,
intel_obj->gpu_active_end = 0;
}
-/** Allocates a new drm_intel_bo to store the data for the buffer object. */
+static void
+mark_buffer_valid_data(struct intel_buffer_object *intel_obj,
+ uint32_t offset, uint32_t size)
+{
+ intel_obj->valid_data_start = MIN2(intel_obj->valid_data_start, offset);
+ intel_obj->valid_data_end = MAX2(intel_obj->valid_data_end, offset + size);
+}
+
+static void
+mark_buffer_invalid(struct intel_buffer_object *intel_obj)
+{
+ intel_obj->valid_data_start = ~0;
+ intel_obj->valid_data_end = 0;
+}
+
+/** Allocates a new brw_bo to store the data for the buffer object. */
static void
alloc_buffer_object(struct brw_context *brw,
struct intel_buffer_object *intel_obj)
{
- intel_obj->buffer = drm_intel_bo_alloc(brw->bufmgr, "bufferobj",
- intel_obj->Base.Size, 64);
+ const struct gl_context *ctx = &brw->ctx;
+
+ uint64_t size = intel_obj->Base.Size;
+ if (ctx->Const.RobustAccess) {
+ /* Pad out buffer objects with an extra 2kB (half a page).
+ *
+ * When pushing UBOs, we need to safeguard against 3DSTATE_CONSTANT_*
+ * reading out of bounds memory. The application might bind a UBO that's
+ * smaller than what the program expects. Ideally, we'd bind an extra
+ * push buffer containing zeros, but we have a limited number of those,
+ * so it's not always viable. Our only safe option is to pad all buffer
+ * objects by the maximum push data length, so that it will never read
+ * past the end of a BO.
+ *
+ * This is unfortunate, but it should result in at most 1 extra page,
+ * which probably isn't too terrible.
+ */
+ size += 64 * 32; /* max read length of 64 256-bit units */
+ }
+ intel_obj->buffer = brw_bo_alloc(brw->bufmgr, "bufferobj", size, 64);
/* the buffer might be bound as a uniform buffer, need to update it
*/
if (intel_obj->Base.UsageHistory & USAGE_TEXTURE_BUFFER)
brw->ctx.NewDriverState |= BRW_NEW_TEXTURE_BUFFER;
if (intel_obj->Base.UsageHistory & USAGE_ATOMIC_COUNTER_BUFFER)
- brw->ctx.NewDriverState |= BRW_NEW_ATOMIC_BUFFER;
+ brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
mark_buffer_inactive(intel_obj);
+ mark_buffer_invalid(intel_obj);
}
static void
release_buffer(struct intel_buffer_object *intel_obj)
{
- drm_intel_bo_unreference(intel_obj->buffer);
+ brw_bo_unreference(intel_obj->buffer);
intel_obj->buffer = NULL;
}
struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
if (!obj) {
_mesa_error_no_memory(__func__);
+ return NULL;
}
_mesa_initialize_buffer_object(ctx, &obj->Base, name);
*/
_mesa_buffer_unmap_all_mappings(ctx, obj);
- drm_intel_bo_unreference(intel_obj->buffer);
- free(intel_obj);
+ brw_bo_unreference(intel_obj->buffer);
+ _mesa_delete_buffer_object(ctx, obj);
}
(void) target;
intel_obj->Base.Size = size;
- intel_obj->Base.BufferRangeSize = size;
intel_obj->Base.Usage = usage;
intel_obj->Base.StorageFlags = storageFlags;
if (!intel_obj->buffer)
return false;
- if (data != NULL)
- drm_intel_bo_subdata(intel_obj->buffer, 0, size, data);
+ if (data != NULL) {
+ brw_bo_subdata(intel_obj->buffer, 0, size, data);
+ mark_buffer_valid_data(intel_obj, 0, size);
+ }
}
return true;
* up with blitting all the time, at the cost of bandwidth)
*/
if (offset + size <= intel_obj->gpu_active_start ||
- intel_obj->gpu_active_end <= offset) {
- if (brw->has_llc) {
- drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
- memcpy(intel_obj->buffer->virtual + offset, data, size);
- drm_intel_bo_unmap(intel_obj->buffer);
-
- if (intel_obj->gpu_active_end > intel_obj->gpu_active_start)
- intel_obj->prefer_stall_to_blit = true;
- return;
- } else {
- perf_debug("BufferSubData could be unsynchronized, but !LLC doesn't support it yet\n");
- }
+ intel_obj->gpu_active_end <= offset ||
+ offset + size <= intel_obj->valid_data_start ||
+ intel_obj->valid_data_end <= offset) {
+ void *map = brw_bo_map(brw, intel_obj->buffer, MAP_WRITE | MAP_ASYNC);
+ memcpy(map + offset, data, size);
+ brw_bo_unmap(intel_obj->buffer);
+
+ if (intel_obj->gpu_active_end > intel_obj->gpu_active_start)
+ intel_obj->prefer_stall_to_blit = true;
+
+ mark_buffer_valid_data(intel_obj, offset, size);
+ return;
}
busy =
- drm_intel_bo_busy(intel_obj->buffer) ||
- drm_intel_bo_references(brw->batch.bo, intel_obj->buffer);
+ brw_bo_busy(intel_obj->buffer) ||
+ brw_batch_references(&brw->batch, intel_obj->buffer);
if (busy) {
- if (size == intel_obj->Base.Size) {
- /* Replace the current busy bo so the subdata doesn't stall. */
- drm_intel_bo_unreference(intel_obj->buffer);
- alloc_buffer_object(brw, intel_obj);
+ if (size == intel_obj->Base.Size ||
+ (intel_obj->valid_data_start >= offset &&
+ intel_obj->valid_data_end <= offset + size)) {
+ /* Replace the current busy bo so the subdata doesn't stall. */
+ brw_bo_unreference(intel_obj->buffer);
+ alloc_buffer_object(brw, intel_obj);
} else if (!intel_obj->prefer_stall_to_blit) {
perf_debug("Using a blit copy to avoid stalling on "
"glBufferSubData(%ld, %ld) (%ldkb) to a busy "
- "(%d-%d) buffer object.\n",
+ "(%d-%d) / valid (%d-%d) buffer object.\n",
(long)offset, (long)offset + size, (long)(size/1024),
intel_obj->gpu_active_start,
- intel_obj->gpu_active_end);
- drm_intel_bo *temp_bo =
- drm_intel_bo_alloc(brw->bufmgr, "subdata temp", size, 64);
-
- drm_intel_bo_subdata(temp_bo, 0, size, data);
-
- intel_emit_linear_blit(brw,
- intel_obj->buffer, offset,
- temp_bo, 0,
- size);
-
- drm_intel_bo_unreference(temp_bo);
+ intel_obj->gpu_active_end,
+ intel_obj->valid_data_start,
+ intel_obj->valid_data_end);
+ struct brw_bo *temp_bo =
+ brw_bo_alloc(brw->bufmgr, "subdata temp", size, 64);
+
+ brw_bo_subdata(temp_bo, 0, size, data);
+
+ brw_blorp_copy_buffers(brw,
+ temp_bo, 0,
+ intel_obj->buffer, offset,
+ size);
+ brw_emit_mi_flush(brw);
+
+ brw_bo_unreference(temp_bo);
+ mark_buffer_valid_data(intel_obj, offset, size);
return;
} else {
perf_debug("Stalling on glBufferSubData(%ld, %ld) (%ldkb) to a busy "
}
}
- drm_intel_bo_subdata(intel_obj->buffer, offset, size, data);
+ brw_bo_subdata(intel_obj->buffer, offset, size, data);
mark_buffer_inactive(intel_obj);
+ mark_buffer_valid_data(intel_obj, offset, size);
}
struct brw_context *brw = brw_context(ctx);
assert(intel_obj);
- if (drm_intel_bo_references(brw->batch.bo, intel_obj->buffer)) {
+ if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
intel_batchbuffer_flush(brw);
}
- drm_intel_bo_get_subdata(intel_obj->buffer, offset, size, data);
+
+ unsigned int map_flags = MAP_READ;
+ mem_copy_fn memcpy_fn = memcpy;
+#ifdef USE_SSE41
+ if (!intel_obj->buffer->cache_coherent && cpu_has_sse4_1) {
+ /* Rather than acquire a new WB mmaping of the buffer object and pull
+ * it into the CPU cache, keep using the WC mmap that we have for writes,
+ * and use the magic movntd instructions instead.
+ */
+ map_flags |= MAP_COHERENT;
+ memcpy_fn = (mem_copy_fn) _mesa_streaming_load_memcpy;
+ }
+#endif
+
+ void *map = brw_bo_map(brw, intel_obj->buffer, map_flags);
+ if (unlikely(!map)) {
+ _mesa_error_no_memory(__func__);
+ return;
+ }
+ memcpy_fn(data, map + offset, size);
+ brw_bo_unmap(intel_obj->buffer);
mark_buffer_inactive(intel_obj);
}
assert(intel_obj);
+ STATIC_ASSERT(GL_MAP_UNSYNCHRONIZED_BIT == MAP_ASYNC);
+ STATIC_ASSERT(GL_MAP_WRITE_BIT == MAP_WRITE);
+ STATIC_ASSERT(GL_MAP_READ_BIT == MAP_READ);
+ STATIC_ASSERT(GL_MAP_PERSISTENT_BIT == MAP_PERSISTENT);
+ STATIC_ASSERT(GL_MAP_COHERENT_BIT == MAP_COHERENT);
+ assert((access & MAP_INTERNAL_MASK) == 0);
+
/* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
* internally uses our functions directly.
*/
* achieve the required synchronization.
*/
if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
- if (drm_intel_bo_references(brw->batch.bo, intel_obj->buffer)) {
- if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
- drm_intel_bo_unreference(intel_obj->buffer);
- alloc_buffer_object(brw, intel_obj);
- } else {
+ if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
+ if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
+ brw_bo_unreference(intel_obj->buffer);
+ alloc_buffer_object(brw, intel_obj);
+ } else {
perf_debug("Stalling on the GPU for mapping a busy buffer "
"object\n");
- intel_batchbuffer_flush(brw);
- }
- } else if (drm_intel_bo_busy(intel_obj->buffer) &&
- (access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
- drm_intel_bo_unreference(intel_obj->buffer);
- alloc_buffer_object(brw, intel_obj);
+ intel_batchbuffer_flush(brw);
+ }
+ } else if (brw_bo_busy(intel_obj->buffer) &&
+ (access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
+ brw_bo_unreference(intel_obj->buffer);
+ alloc_buffer_object(brw, intel_obj);
}
}
+ if (access & MAP_WRITE)
+ mark_buffer_valid_data(intel_obj, offset, length);
+
/* If the user is mapping a range of an active buffer object but
* 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
*/
if (!(access & (GL_MAP_UNSYNCHRONIZED_BIT | GL_MAP_PERSISTENT_BIT)) &&
(access & GL_MAP_INVALIDATE_RANGE_BIT) &&
- drm_intel_bo_busy(intel_obj->buffer)) {
+ brw_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;
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);
- if (brw->has_llc) {
- brw_bo_map(brw, intel_obj->range_map_bo[index],
- (access & GL_MAP_WRITE_BIT) != 0, "range-map");
- } else {
- drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo[index]);
- }
- obj->Mappings[index].Pointer =
- intel_obj->range_map_bo[index]->virtual + intel_obj->map_extra[index];
+ intel_obj->range_map_bo[index] =
+ brw_bo_alloc(brw->bufmgr, "BO blit temp",
+ length + intel_obj->map_extra[index], alignment);
+ void *map = brw_bo_map(brw, intel_obj->range_map_bo[index], access);
+ obj->Mappings[index].Pointer = map + intel_obj->map_extra[index];
return obj->Mappings[index].Pointer;
}
- if (access & GL_MAP_UNSYNCHRONIZED_BIT) {
- if (!brw->has_llc && brw->perf_debug &&
- drm_intel_bo_busy(intel_obj->buffer)) {
- perf_debug("MapBufferRange with GL_MAP_UNSYNCHRONIZED_BIT stalling (it's actually synchronized on non-LLC platforms)\n");
- }
- drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
- } else if (!brw->has_llc && (!(access & GL_MAP_READ_BIT) ||
- (access & GL_MAP_PERSISTENT_BIT))) {
- drm_intel_gem_bo_map_gtt(intel_obj->buffer);
- mark_buffer_inactive(intel_obj);
- } else {
- brw_bo_map(brw, intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0,
- "MapBufferRange");
+ void *map = brw_bo_map(brw, intel_obj->buffer, access);
+ if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
mark_buffer_inactive(intel_obj);
}
- obj->Mappings[index].Pointer = intel_obj->buffer->virtual + offset;
+ obj->Mappings[index].Pointer = map + offset;
return obj->Mappings[index].Pointer;
}
* another blit of that area and the complete newer data will land the
* second time.
*/
- intel_emit_linear_blit(brw,
- intel_obj->buffer,
- obj->Mappings[index].Offset + offset,
- intel_obj->range_map_bo[index],
+ brw_blorp_copy_buffers(brw,
+ intel_obj->range_map_bo[index],
intel_obj->map_extra[index] + offset,
- length);
+ intel_obj->buffer,
+ obj->Mappings[index].Offset + offset,
+ length);
mark_buffer_gpu_usage(intel_obj,
obj->Mappings[index].Offset + offset,
length);
+ brw_emit_mi_flush(brw);
}
assert(intel_obj);
assert(obj->Mappings[index].Pointer);
if (intel_obj->range_map_bo[index] != NULL) {
- drm_intel_bo_unmap(intel_obj->range_map_bo[index]);
+ brw_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,
+ brw_blorp_copy_buffers(brw,
intel_obj->range_map_bo[index],
intel_obj->map_extra[index],
+ intel_obj->buffer, obj->Mappings[index].Offset,
obj->Mappings[index].Length);
mark_buffer_gpu_usage(intel_obj, obj->Mappings[index].Offset,
obj->Mappings[index].Length);
+ brw_emit_mi_flush(brw);
}
/* Since we've emitted some blits to buffers that will (likely) be used
* flush. Once again, we wish for a domain tracker in libdrm to cover
* usage inside of a batchbuffer.
*/
- brw_emit_mi_flush(brw);
- drm_intel_bo_unreference(intel_obj->range_map_bo[index]);
+ brw_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);
+ brw_bo_unmap(intel_obj->buffer);
}
obj->Mappings[index].Pointer = NULL;
obj->Mappings[index].Offset = 0;
* Anywhere that uses buffer objects in the pipeline should be using this to
* mark the range of the buffer that is being accessed by the pipeline.
*/
-drm_intel_bo *
+struct brw_bo *
intel_bufferobj_buffer(struct brw_context *brw,
struct intel_buffer_object *intel_obj,
- uint32_t offset, uint32_t size)
+ uint32_t offset, uint32_t size, bool write)
{
/* This is needed so that things like transform feedback and texture buffer
* objects that need a BO but don't want to check that they exist for
mark_buffer_gpu_usage(intel_obj, offset, size);
+ /* If writing, (conservatively) mark this section as having valid data. */
+ if (write)
+ mark_buffer_valid_data(intel_obj, offset, size);
+
return intel_obj->buffer;
}
struct brw_context *brw = brw_context(ctx);
struct intel_buffer_object *intel_src = intel_buffer_object(src);
struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
- drm_intel_bo *src_bo, *dst_bo;
+ struct brw_bo *src_bo, *dst_bo;
if (size == 0)
return;
- dst_bo = intel_bufferobj_buffer(brw, intel_dst, write_offset, size);
- src_bo = intel_bufferobj_buffer(brw, intel_src, read_offset, size);
+ dst_bo = intel_bufferobj_buffer(brw, intel_dst, write_offset, size, true);
+ src_bo = intel_bufferobj_buffer(brw, intel_src, read_offset, size, false);
- intel_emit_linear_blit(brw,
- dst_bo, write_offset,
- src_bo, read_offset, size);
+ brw_blorp_copy_buffers(brw,
+ src_bo, read_offset,
+ dst_bo, write_offset, size);
/* 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