radv/amdgpu: Add winsys implementation of virtual buffers.

v2: - Added comments.
    - Fixed a double unmap bug.
    - Actually unmap the non-edge old ranges.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

src/amd/vulkan/radv_radeon_winsys.h

@@ -47,6 +47,7 @@ enum radeon_bo_flag { /* bitfield */
 	RADEON_FLAG_GTT_WC =        (1 << 0),
 	RADEON_FLAG_CPU_ACCESS =    (1 << 1),
 	RADEON_FLAG_NO_CPU_ACCESS = (1 << 2),
+	RADEON_FLAG_VIRTUAL =       (1 << 3)
 };
 
 enum radeon_bo_usage { /* bitfield */
@@ -289,6 +290,10 @@ struct radeon_winsys {
 
 	void (*buffer_set_metadata)(struct radeon_winsys_bo *bo,
 				    struct radeon_bo_metadata *md);
+
+	void (*buffer_virtual_bind)(struct radeon_winsys_bo *parent,
+	                            uint64_t offset, uint64_t size,
+	                            struct radeon_winsys_bo *bo, uint64_t bo_offset);
 	struct radeon_winsys_ctx *(*ctx_create)(struct radeon_winsys *ws);
 	void (*ctx_destroy)(struct radeon_winsys_ctx *ctx);
 

src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c

@@ -34,19 +34,201 @@
 #include <amdgpu_drm.h>
 #include <inttypes.h>
 
+static void
+radv_amdgpu_winsys_virtual_map(struct radv_amdgpu_winsys_bo *bo,
+                               const struct radv_amdgpu_map_range *range)
+{
+	assert(range->size);
+
+	if (!range->bo)
+		return; /* TODO: PRT mapping */
+
+	int r = amdgpu_bo_va_op(range->bo->bo, range->bo_offset, range->size,
+	                        range->offset + bo->va, 0, AMDGPU_VA_OP_MAP);
+	if (r)
+		abort();
+}
+
+static void
+radv_amdgpu_winsys_virtual_unmap(struct radv_amdgpu_winsys_bo *bo,
+                                 const struct radv_amdgpu_map_range *range)
+{
+	assert(range->size);
+
+	if (!range->bo)
+		return; /* TODO: PRT mapping */
+
+	int r = amdgpu_bo_va_op(range->bo->bo, range->bo_offset, range->size,
+	                        range->offset + bo->va, 0, AMDGPU_VA_OP_UNMAP);
+	if (r)
+		abort();
+}
+
+static void
+radv_amdgpu_winsys_rebuild_bo_list(struct radv_amdgpu_winsys_bo *bo)
+{
+	bo->bo_count = 0;
+	for (uint32_t i = 0; i < bo->range_count; ++i) {
+		bool found = false;
+		if (!bo->ranges[i].bo)
+			continue;
+
+		for(uint32_t j = 0; j <  bo->bo_count; ++j) {
+			if (bo->bos[j] == bo->ranges[i].bo) {
+				found = true;
+				break;
+			}
+		}
+
+		if (!found) {
+			if (bo->bo_capacity == bo->bo_count) {
+				bo->bos = realloc(bo->bos,
+				                  (bo->bo_capacity + 1) * sizeof(struct radv_amdgpu_bo *));
+				++bo->bo_capacity;
+			}
+			bo->bos[bo->bo_count++] = bo->ranges[i].bo;
+		}
+	}
+}
+
+static void
+radv_amdgpu_winsys_bo_virtual_bind(struct radeon_winsys_bo *_parent,
+                                   uint64_t offset, uint64_t size,
+                                   struct radeon_winsys_bo *_bo, uint64_t bo_offset)
+{
+	struct radv_amdgpu_winsys_bo *parent = (struct radv_amdgpu_winsys_bo *)_parent;
+	struct radv_amdgpu_winsys_bo *bo = (struct radv_amdgpu_winsys_bo*)_bo;
+	int range_count_delta, new_idx;
+	int first = 0, last;
+	struct radv_amdgpu_map_range new_first, new_last;
+
+	assert(parent->is_virtual);
+	assert(!bo || !bo->is_virtual);
+
+	if (!size)
+		return;
+
+	/* We have at most 2 new ranges (1 by the bind, and another one by splitting a range that contains the newly bound range). */
+	if (parent->range_capacity - parent->range_count < 2) {
+		parent->range_capacity += 2;
+		parent->ranges = realloc(parent->ranges,
+		                         parent->range_capacity * sizeof(struct radv_amdgpu_map_range));
+	}
+
+	/*
+	 * [first, last] is exactly the range of ranges that either overlap the
+	 * new parent, or are adjacent to it. This corresponds to the bind ranges
+	 * that may change.
+	 */
+	while(first + 1 < parent->range_count && parent->ranges[first].offset + parent->ranges[first].size < offset)
+		++first;
+
+	last = first;
+	while(last + 1 < parent->range_count && parent->ranges[last].offset <= offset + size)
+		++last;
+
+	/* Whether the first or last range are going to be totally removed or just
+	 * resized/left alone. Note that in the case of first == last, we will split
+	 * this into a part before and after the new range. The remove flag is then
+	 * whether to not create the corresponding split part. */
+	bool remove_first = parent->ranges[first].offset == offset;
+	bool remove_last = parent->ranges[last].offset + parent->ranges[last].size == offset + size;
+	bool unmapped_first = false;
+
+	assert(parent->ranges[first].offset <= offset);
+	assert(parent->ranges[last].offset + parent->ranges[last].size >= offset + size);
+
+	/* Try to merge the new range with the first range. */
+	if (parent->ranges[first].bo == bo && (!bo || offset - bo_offset == parent->ranges[first].offset - parent->ranges[first].bo_offset)) {
+		size += offset - parent->ranges[first].offset;
+		offset = parent->ranges[first].offset;
+		remove_first = true;
+	}
+
+	/* Try to merge the new range with the last range. */
+	if (parent->ranges[last].bo == bo && (!bo || offset - bo_offset == parent->ranges[last].offset - parent->ranges[last].bo_offset)) {
+		size = parent->ranges[last].offset + parent->ranges[last].size - offset;
+		remove_last = true;
+	}
+
+	range_count_delta = 1 - (last - first + 1) + !remove_first + !remove_last;
+	new_idx = first + !remove_first;
+
+	/* Any range between first and last is going to be entirely covered by the new range so just unmap them. */
+	for (int i = first + 1; i < last; ++i)
+		radv_amdgpu_winsys_virtual_unmap(parent, parent->ranges + i);
+
+	/* If the first/last range are not left alone we unmap then and optionally map
+	 * them again after modifications. Not that this implicitly can do the splitting
+	 * if first == last. */
+	new_first = parent->ranges[first];
+	new_last = parent->ranges[last];
+
+	if (parent->ranges[first].offset + parent->ranges[first].size > offset || remove_first) {
+		radv_amdgpu_winsys_virtual_unmap(parent, parent->ranges + first);
+		unmapped_first = true;
+
+		if (!remove_first) {
+			new_first.size = offset - new_first.offset;
+			radv_amdgpu_winsys_virtual_map(parent, &new_first);
+		}
+	}
+
+	if (parent->ranges[last].offset < offset + size || remove_last) {
+		if (first != last || !unmapped_first)
+			radv_amdgpu_winsys_virtual_unmap(parent, parent->ranges + last);
+
+		if (!remove_last) {
+			new_last.size -= offset + size - new_last.offset;
+			new_last.offset = offset + size;
+			radv_amdgpu_winsys_virtual_map(parent, &new_last);
+		}
+	}
+
+	/* Moves the range list after last to account for the changed number of ranges. */
+	memmove(parent->ranges + last + 1 + range_count_delta, parent->ranges + last + 1,
+	        sizeof(struct radv_amdgpu_map_range) * (parent->range_count - last - 1));
+
+	if (!remove_first)
+		parent->ranges[first] = new_first;
+
+	if (!remove_last)
+		parent->ranges[new_idx + 1] = new_last;
+
+	/* Actually set up the new range. */
+	parent->ranges[new_idx].offset = offset;
+	parent->ranges[new_idx].size = size;
+	parent->ranges[new_idx].bo = bo;
+	parent->ranges[new_idx].bo_offset = bo_offset;
+
+	radv_amdgpu_winsys_virtual_map(parent, parent->ranges + new_idx);
+
+	parent->range_count += range_count_delta;
+
+	radv_amdgpu_winsys_rebuild_bo_list(parent);
+}
+
 static void radv_amdgpu_winsys_bo_destroy(struct radeon_winsys_bo *_bo)
 {
 	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
 
-	if (bo->ws->debug_all_bos) {
+	if (bo->is_virtual) {
-		pthread_mutex_lock(&bo->ws->global_bo_list_lock);
+		for (uint32_t i = 0; i < bo->range_count; ++i) {
-		LIST_DEL(&bo->global_list_item);
+			radv_amdgpu_winsys_virtual_unmap(bo, bo->ranges + i);
-		bo->ws->num_buffers--;
+		}
-		pthread_mutex_unlock(&bo->ws->global_bo_list_lock);
+		free(bo->bos);
+		free(bo->ranges);
+	} else {
+		if (bo->ws->debug_all_bos) {
+			pthread_mutex_lock(&bo->ws->global_bo_list_lock);
+			LIST_DEL(&bo->global_list_item);
+			bo->ws->num_buffers--;
+			pthread_mutex_unlock(&bo->ws->global_bo_list_lock);
+		}
+		amdgpu_bo_va_op(bo->bo, 0, bo->size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
+		amdgpu_bo_free(bo->bo);
 	}
-	amdgpu_bo_va_op(bo->bo, 0, bo->size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
 	amdgpu_va_range_free(bo->va_handle);
-	amdgpu_bo_free(bo->bo);
 	FREE(bo);
 }
 
@@ -81,6 +263,31 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
 		return NULL;
 	}
 
+	r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+				  size, alignment, 0, &va, &va_handle, 0);
+	if (r)
+		goto error_va_alloc;
+
+	bo->va = va;
+	bo->va_handle = va_handle;
+	bo->size = size;
+	bo->ws = ws;
+	bo->is_virtual = !!(flags & RADEON_FLAG_VIRTUAL);
+
+	if (flags & RADEON_FLAG_VIRTUAL) {
+		bo->ranges = realloc(NULL, sizeof(struct radv_amdgpu_map_range));
+		bo->range_count = 1;
+		bo->range_capacity = 1;
+
+		bo->ranges[0].offset = 0;
+		bo->ranges[0].size = size;
+		bo->ranges[0].bo = NULL;
+		bo->ranges[0].bo_offset = 0;
+
+		radv_amdgpu_winsys_virtual_map(bo, bo->ranges);
+		return (struct radeon_winsys_bo *)bo;
+	}
+
 	request.alloc_size = size;
 	request.phys_alignment = alignment;
 
@@ -105,31 +312,22 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
 		goto error_bo_alloc;
 	}
 
-	r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
-				  size, alignment, 0, &va, &va_handle, 0);
-	if (r)
-		goto error_va_alloc;
-
 	r = amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP);
 	if (r)
 		goto error_va_map;
 
 	bo->bo = buf_handle;
-	bo->va = va;
-	bo->va_handle = va_handle;
 	bo->initial_domain = initial_domain;
-	bo->size = size;
 	bo->is_shared = false;
-	bo->ws = ws;
 	radv_amdgpu_add_buffer_to_global_list(bo);
 	return (struct radeon_winsys_bo *)bo;
 error_va_map:
-	amdgpu_va_range_free(va_handle);
-
-error_va_alloc:
 	amdgpu_bo_free(buf_handle);
 
 error_bo_alloc:
+	amdgpu_va_range_free(va_handle);
+
+error_va_alloc:
 	FREE(bo);
 	return NULL;
 }
@@ -296,4 +494,5 @@ void radv_amdgpu_bo_init_functions(struct radv_amdgpu_winsys *ws)
 	ws->base.buffer_from_fd = radv_amdgpu_winsys_bo_from_fd;
 	ws->base.buffer_get_fd = radv_amdgpu_winsys_get_fd;
 	ws->base.buffer_set_metadata = radv_amdgpu_winsys_bo_set_metadata;
+	ws->base.buffer_virtual_bind = radv_amdgpu_winsys_bo_virtual_bind;
 }

src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h

@@ -31,17 +31,40 @@
 
 #include "radv_amdgpu_winsys.h"
 
-struct radv_amdgpu_winsys_bo {
-	amdgpu_bo_handle bo;
-	amdgpu_va_handle va_handle;
 
-	uint64_t va;
+struct radv_amdgpu_map_range {
-	enum radeon_bo_domain initial_domain;
+	uint64_t offset;
 	uint64_t size;
-	bool is_shared;
+	struct radv_amdgpu_winsys_bo *bo;
+	uint64_t bo_offset;
+};
 
+struct radv_amdgpu_winsys_bo {
+	amdgpu_va_handle va_handle;
+	uint64_t va;
+	uint64_t size;
 	struct radv_amdgpu_winsys *ws;
-	struct list_head global_list_item;
+	bool is_virtual;
+
+	union {
+		/* physical bo */
+		struct {
+			amdgpu_bo_handle bo;
+			enum radeon_bo_domain initial_domain;
+			bool is_shared;
+			struct list_head global_list_item;
+		};
+		/* virtual bo */
+		struct {
+			struct radv_amdgpu_map_range *ranges;
+			uint32_t range_count;
+			uint32_t range_capacity;
+
+			struct radv_amdgpu_winsys_bo **bos;
+			uint32_t bo_count;
+			uint32_t bo_capacity;
+		};
+	};
 };
 
 static inline

src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c

@@ -34,6 +34,11 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"
 
+
+enum {
+	VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
+};
+
 struct radv_amdgpu_cs {
 	struct radeon_winsys_cs base;
 	struct radv_amdgpu_winsys *ws;
@@ -56,6 +61,12 @@ struct radv_amdgpu_cs {
 
 	int                         buffer_hash_table[1024];
 	unsigned                    hw_ip;
+
+	unsigned                    num_virtual_buffers;
+	unsigned                    max_num_virtual_buffers;
+	struct radeon_winsys_bo     **virtual_buffers;
+	uint8_t                     *virtual_buffer_priorities;
+	int                         *virtual_buffer_hash_table;
 };
 
 static inline struct radv_amdgpu_cs *
@@ -141,6 +152,9 @@ static void radv_amdgpu_cs_destroy(struct radeon_winsys_cs *rcs)
 		cs->ws->base.buffer_destroy(cs->old_ib_buffers[i]);
 
 	free(cs->old_ib_buffers);
+	free(cs->virtual_buffers);
+	free(cs->virtual_buffer_priorities);
+	free(cs->virtual_buffer_hash_table);
 	free(cs->handles);
 	free(cs->priorities);
 	free(cs);
@@ -315,7 +329,13 @@ static void radv_amdgpu_cs_reset(struct radeon_winsys_cs *_cs)
 		cs->buffer_hash_table[hash] = -1;
 	}
 
+	for (unsigned i = 0; i < cs->num_virtual_buffers; ++i) {
+		unsigned hash = ((uintptr_t)cs->virtual_buffers[i] >> 6) & (VIRTUAL_BUFFER_HASH_TABLE_SIZE - 1);
+		cs->virtual_buffer_hash_table[hash] = -1;
+	}
+
 	cs->num_buffers = 0;
+	cs->num_virtual_buffers = 0;
 
 	if (cs->ws->use_ib_bos) {
 		cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer, 8);
@@ -380,6 +400,49 @@ static void radv_amdgpu_cs_add_buffer_internal(struct radv_amdgpu_cs *cs,
 	++cs->num_buffers;
 }
 
+static void radv_amdgpu_cs_add_virtual_buffer(struct radeon_winsys_cs *_cs,
+                                              struct radeon_winsys_bo *bo,
+                                              uint8_t priority)
+{
+	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
+	unsigned hash = ((uintptr_t)bo >> 6) & (VIRTUAL_BUFFER_HASH_TABLE_SIZE - 1);
+
+
+	if (!cs->virtual_buffer_hash_table) {
+		cs->virtual_buffer_hash_table = malloc(VIRTUAL_BUFFER_HASH_TABLE_SIZE * sizeof(int));
+		for (int i = 0; i < VIRTUAL_BUFFER_HASH_TABLE_SIZE; ++i)
+			cs->virtual_buffer_hash_table[i] = -1;
+	}
+
+	if (cs->virtual_buffer_hash_table[hash] >= 0) {
+		int idx = cs->virtual_buffer_hash_table[hash];
+		if (cs->virtual_buffers[idx] == bo) {
+			cs->virtual_buffer_priorities[idx] = MAX2(cs->virtual_buffer_priorities[idx], priority);
+			return;
+		}
+		for (unsigned i = 0; i < cs->num_virtual_buffers; ++i) {
+			if (cs->virtual_buffers[i] == bo) {
+				cs->virtual_buffer_priorities[i] = MAX2(cs->virtual_buffer_priorities[i], priority);
+				cs->virtual_buffer_hash_table[hash] = i;
+				return;
+			}
+		}
+	}
+
+	if(cs->max_num_virtual_buffers <= cs->num_virtual_buffers) {
+		cs->max_num_virtual_buffers = MAX2(2, cs->max_num_virtual_buffers * 2);
+		cs->virtual_buffers = realloc(cs->virtual_buffers, sizeof(struct radv_amdgpu_virtual_virtual_buffer*) * cs->max_num_virtual_buffers);
+		cs->virtual_buffer_priorities = realloc(cs->virtual_buffer_priorities, sizeof(uint8_t) * cs->max_num_virtual_buffers);
+	}
+
+	cs->virtual_buffers[cs->num_virtual_buffers] = bo;
+	cs->virtual_buffer_priorities[cs->num_virtual_buffers] = priority;
+
+	cs->virtual_buffer_hash_table[hash] = cs->num_virtual_buffers;
+	++cs->num_virtual_buffers;
+
+}
+
 static void radv_amdgpu_cs_add_buffer(struct radeon_winsys_cs *_cs,
 				 struct radeon_winsys_bo *_bo,
 				 uint8_t priority)
@@ -387,6 +450,11 @@ static void radv_amdgpu_cs_add_buffer(struct radeon_winsys_cs *_cs,
 	struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
 	struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
 
+	if (bo->is_virtual)  {
+		radv_amdgpu_cs_add_virtual_buffer(_cs, _bo, priority);
+		return;
+	}
+
 	radv_amdgpu_cs_add_buffer_internal(cs, bo->bo, priority);
 }
 
@@ -401,6 +469,11 @@ static void radv_amdgpu_cs_execute_secondary(struct radeon_winsys_cs *_parent,
 						   child->priorities[i]);
 	}
 
+	for (unsigned i = 0; i < child->num_virtual_buffers; ++i) {
+		radv_amdgpu_cs_add_buffer(&parent->base, child->virtual_buffers[i],
+		                          child->virtual_buffer_priorities[i]);
+	}
+
 	if (parent->ws->use_ib_bos) {
 		if (parent->base.cdw + 4 > parent->base.max_dw)
 			radv_amdgpu_cs_grow(&parent->base, 4);
@@ -449,7 +522,8 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 					  bo_list);
 		free(handles);
 		pthread_mutex_unlock(&ws->global_bo_list_lock);
-	} else if (count == 1 && !extra_bo && !extra_cs) {
+	} else if (count == 1 && !extra_bo && !extra_cs &&
+	           !radv_amdgpu_cs(cs_array[0])->num_virtual_buffers) {
 		struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[0];
 		r = amdgpu_bo_list_create(ws->dev, cs->num_buffers, cs->handles,
 					  cs->priorities, bo_list);
@@ -459,6 +533,8 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 		for (unsigned i = 0; i < count; ++i) {
 			struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[i];
 			total_buffer_count += cs->num_buffers;
+			for (unsigned j = 0; j < cs->num_virtual_buffers; ++j)
+				total_buffer_count += radv_amdgpu_winsys_bo(cs->virtual_buffers[j])->bo_count;
 		}
 
 		if (extra_cs) {
@@ -512,6 +588,26 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 					++unique_bo_count;
 				}
 			}
+			for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
+				struct radv_amdgpu_winsys_bo *virtual_bo = radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
+				for(unsigned k = 0; k < virtual_bo->bo_count; ++k) {
+					struct radv_amdgpu_winsys_bo *bo = virtual_bo->bos[k];
+					bool found = false;
+					for (unsigned m = 0; m < unique_bo_count; ++m) {
+						if (handles[m] == bo->bo) {
+							found = true;
+							priorities[m] = MAX2(priorities[m],
+									cs->virtual_buffer_priorities[j]);
+							break;
+						}
+					}
+					if (!found) {
+						handles[unique_bo_count] = bo->bo;
+						priorities[unique_bo_count] = cs->virtual_buffer_priorities[j];
+						++unique_bo_count;
+					}
+				}
+			}
 		}
 		r = amdgpu_bo_list_create(ws->dev, unique_bo_count, handles,
 					  priorities, bo_list);