radv/winsys: Add priority handling during submit.
authorBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Mon, 28 Jan 2019 00:23:14 +0000 (01:23 +0100)
committerBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tue, 29 Jan 2019 14:56:52 +0000 (15:56 +0100)
Switched to the raw bo list api to avoid having to use 2 arrays for
everything.

This was introduced in libdrm 2.4.97 which we already depend upon.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c

index 83e6f37f58b08eb536e04e191b2f7a1dcff620ae..79958cdd7cb803bf24b0b39a9e8a84c84b7f02f3 100644 (file)
@@ -395,6 +395,9 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
        bo->is_shared = false;
        bo->priority = priority;
 
+       r = amdgpu_bo_export(buf_handle, amdgpu_bo_handle_type_kms, &bo->bo_handle);
+       assert(!r);
+
        if (initial_domain & RADEON_DOMAIN_VRAM)
                p_atomic_add(&ws->allocated_vram,
                             align64(bo->size, ws->info.gart_page_size));
@@ -503,6 +506,9 @@ radv_amdgpu_winsys_bo_from_ptr(struct radeon_winsys *_ws,
        bo->initial_domain = RADEON_DOMAIN_GTT;
        bo->priority = priority;
 
+       MAYBE_UNUSED int r = amdgpu_bo_export(buf_handle, amdgpu_bo_handle_type_kms, &bo->bo_handle);
+       assert(!r);
+
        p_atomic_add(&ws->allocated_gtt,
                     align64(bo->size, ws->info.gart_page_size));
 
@@ -573,6 +579,9 @@ radv_amdgpu_winsys_bo_from_fd(struct radeon_winsys *_ws,
        bo->priority = priority;
        bo->ref_count = 1;
 
+       r = amdgpu_bo_export(result.buf_handle, amdgpu_bo_handle_type_kms, &bo->bo_handle);
+       assert(!r);
+
        if (bo->initial_domain & RADEON_DOMAIN_VRAM)
                p_atomic_add(&ws->allocated_vram,
                             align64(bo->size, ws->info.gart_page_size));
index 400f79deff762238ba8da4994583154286483891..ad86e03ee00e7e0d2eaf8ebabc12e39ca7e1ca52 100644 (file)
@@ -54,6 +54,7 @@ struct radv_amdgpu_winsys_bo {
                        amdgpu_bo_handle bo;
                        enum radeon_bo_domain initial_domain;
                        bool is_shared;
+                       uint32_t bo_handle;
                        struct list_head global_list_item;
                };
                /* virtual bo */
index 366eb3481a2fbfac6c5b5522de5b8172325c44bf..d3b1e2cd4c66a20939330baa248a5f9f4947b4cb 100644 (file)
@@ -50,7 +50,7 @@ struct radv_amdgpu_cs {
        uint8_t                 *ib_mapped;
        unsigned                    max_num_buffers;
        unsigned                    num_buffers;
-       amdgpu_bo_handle            *handles;
+       struct drm_amdgpu_bo_list_entry *handles;
 
        struct radeon_winsys_bo     **old_ib_buffers;
        unsigned                    num_old_ib_buffers;
@@ -92,17 +92,71 @@ static int ring_to_hw_ip(enum ring_type ring)
        }
 }
 
+struct radv_amdgpu_cs_request {
+       /** Specify flags with additional information */
+       uint64_t flags;
+
+       /** Specify HW IP block type to which to send the IB. */
+       unsigned ip_type;
+
+       /** IP instance index if there are several IPs of the same type. */
+       unsigned ip_instance;
+
+       /**
+        * Specify ring index of the IP. We could have several rings
+        * in the same IP. E.g. 0 for SDMA0 and 1 for SDMA1.
+        */
+       uint32_t ring;
+
+       /**
+        * List handle with resources used by this request. This is a raw
+        * bo list handle used by the kernel.
+        */
+       uint32_t resources;
+
+       /**
+        * Number of dependencies this Command submission needs to
+        * wait for before starting execution.
+        */
+       uint32_t number_of_dependencies;
+
+       /**
+        * Array of dependencies which need to be met before
+        * execution can start.
+        */
+       struct amdgpu_cs_fence *dependencies;
+
+       /** Number of IBs to submit in the field ibs. */
+       uint32_t number_of_ibs;
+
+       /**
+        * IBs to submit. Those IBs will be submit together as single entity
+        */
+       struct amdgpu_cs_ib_info *ibs;
+
+       /**
+        * The returned sequence number for the command submission
+        */
+       uint64_t seq_no;
+
+       /**
+        * The fence information
+        */
+       struct amdgpu_cs_fence_info fence_info;
+};
+
+
 static int radv_amdgpu_signal_sems(struct radv_amdgpu_ctx *ctx,
                                   uint32_t ip_type,
                                   uint32_t ring,
                                   struct radv_winsys_sem_info *sem_info);
 static int radv_amdgpu_cs_submit(struct radv_amdgpu_ctx *ctx,
-                                struct amdgpu_cs_request *request,
+                                struct radv_amdgpu_cs_request *request,
                                 struct radv_winsys_sem_info *sem_info);
 
 static void radv_amdgpu_request_to_fence(struct radv_amdgpu_ctx *ctx,
                                         struct radv_amdgpu_fence *fence,
-                                        struct amdgpu_cs_request *req)
+                                        struct radv_amdgpu_cs_request *req)
 {
        fence->fence.context = ctx->ctx;
        fence->fence.ip_type = req->ip_type;
@@ -414,8 +468,8 @@ static void radv_amdgpu_cs_reset(struct radeon_cmdbuf *_cs)
        cs->failed = false;
 
        for (unsigned i = 0; i < cs->num_buffers; ++i) {
-               unsigned hash = ((uintptr_t)cs->handles[i] >> 6) &
-                                (ARRAY_SIZE(cs->buffer_hash_table) - 1);
+               unsigned hash = cs->handles[i].bo_handle &
+                               (ARRAY_SIZE(cs->buffer_hash_table) - 1);
                cs->buffer_hash_table[hash] = -1;
        }
 
@@ -450,19 +504,19 @@ static void radv_amdgpu_cs_reset(struct radeon_cmdbuf *_cs)
 }
 
 static int radv_amdgpu_cs_find_buffer(struct radv_amdgpu_cs *cs,
-                                     amdgpu_bo_handle bo)
+                                     uint32_t bo)
 {
-       unsigned hash = ((uintptr_t)bo >> 6) & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
+       unsigned hash = bo & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
        int index = cs->buffer_hash_table[hash];
 
        if (index == -1)
                return -1;
 
-       if (cs->handles[index] == bo)
+       if (cs->handles[index].bo_handle == bo)
                return index;
 
        for (unsigned i = 0; i < cs->num_buffers; ++i) {
-               if (cs->handles[i] == bo) {
+               if (cs->handles[i].bo_handle == bo) {
                        cs->buffer_hash_table[hash] = i;
                        return i;
                }
@@ -472,7 +526,7 @@ static int radv_amdgpu_cs_find_buffer(struct radv_amdgpu_cs *cs,
 }
 
 static void radv_amdgpu_cs_add_buffer_internal(struct radv_amdgpu_cs *cs,
-                                              amdgpu_bo_handle bo)
+                                              uint32_t bo, uint8_t priority)
 {
        unsigned hash;
        int index = radv_amdgpu_cs_find_buffer(cs, bo);
@@ -482,11 +536,12 @@ static void radv_amdgpu_cs_add_buffer_internal(struct radv_amdgpu_cs *cs,
 
        if (cs->num_buffers == cs->max_num_buffers) {
                unsigned new_count = MAX2(1, cs->max_num_buffers * 2);
-               cs->handles = realloc(cs->handles, new_count * sizeof(amdgpu_bo_handle));
+               cs->handles = realloc(cs->handles, new_count * sizeof(struct drm_amdgpu_bo_list_entry));
                cs->max_num_buffers = new_count;
        }
 
-       cs->handles[cs->num_buffers] = bo;
+       cs->handles[cs->num_buffers].bo_handle = bo;
+       cs->handles[cs->num_buffers].bo_priority = priority;
 
        hash = ((uintptr_t)bo >> 6) & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
        cs->buffer_hash_table[hash] = cs->num_buffers;
@@ -546,7 +601,7 @@ static void radv_amdgpu_cs_add_buffer(struct radeon_cmdbuf *_cs,
        if (bo->base.is_local)
                return;
 
-       radv_amdgpu_cs_add_buffer_internal(cs, bo->bo);
+       radv_amdgpu_cs_add_buffer_internal(cs, bo->bo_handle, bo->priority);
 }
 
 static void radv_amdgpu_cs_execute_secondary(struct radeon_cmdbuf *_parent,
@@ -556,7 +611,9 @@ static void radv_amdgpu_cs_execute_secondary(struct radeon_cmdbuf *_parent,
        struct radv_amdgpu_cs *child = radv_amdgpu_cs(_child);
 
        for (unsigned i = 0; i < child->num_buffers; ++i) {
-               radv_amdgpu_cs_add_buffer_internal(parent, child->handles[i]);
+               radv_amdgpu_cs_add_buffer_internal(parent,
+                                                  child->handles[i].bo_handle,
+                                                  child->handles[i].bo_priority);
        }
 
        for (unsigned i = 0; i < child->num_virtual_buffers; ++i) {
@@ -587,13 +644,13 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                                      unsigned num_extra_bo,
                                      struct radeon_cmdbuf *extra_cs,
                                      const struct radv_winsys_bo_list *radv_bo_list,
-                                     amdgpu_bo_list_handle *bo_list)
+                                     uint32_t *bo_list)
 {
        int r = 0;
 
        if (ws->debug_all_bos) {
                struct radv_amdgpu_winsys_bo *bo;
-               amdgpu_bo_handle *handles;
+               struct drm_amdgpu_bo_list_entry *handles;
                unsigned num = 0;
 
                pthread_mutex_lock(&ws->global_bo_list_lock);
@@ -606,12 +663,12 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 
                LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, global_list_item) {
                        assert(num < ws->num_buffers);
-                       handles[num++] = bo->bo;
+                       handles[num].bo_handle = bo->bo_handle;
+                       handles[num].bo_priority = bo->priority;
                }
 
-               r = amdgpu_bo_list_create(ws->dev, ws->num_buffers,
-                                         handles, NULL,
-                                         bo_list);
+               r = amdgpu_bo_list_create_raw(ws->dev, ws->num_buffers,
+                                             handles, bo_list);
                free(handles);
                pthread_mutex_unlock(&ws->global_bo_list_lock);
        } else if (count == 1 && !num_extra_bo && !extra_cs && !radv_bo_list &&
@@ -621,8 +678,8 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                        *bo_list = 0;
                        return 0;
                }
-               r = amdgpu_bo_list_create(ws->dev, cs->num_buffers, cs->handles,
-                                         NULL, bo_list);
+               r = amdgpu_bo_list_create_raw(ws->dev, cs->num_buffers, cs->handles,
+                                             bo_list);
        } else {
                unsigned total_buffer_count = num_extra_bo;
                unsigned unique_bo_count = num_extra_bo;
@@ -645,14 +702,15 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                        *bo_list = 0;
                        return 0;
                }
-               amdgpu_bo_handle *handles = malloc(sizeof(amdgpu_bo_handle) * total_buffer_count);
+               struct drm_amdgpu_bo_list_entry *handles = malloc(sizeof(struct drm_amdgpu_bo_list_entry) * total_buffer_count);
                if (!handles) {
                        free(handles);
                        return -ENOMEM;
                }
 
                for (unsigned i = 0; i < num_extra_bo; i++) {
-                       handles[i] = extra_bo_array[i]->bo;
+                       handles[i].bo_handle = extra_bo_array[i]->bo_handle;
+                       handles[i].bo_priority = extra_bo_array[i]->priority;
                }
 
                for (unsigned i = 0; i < count + !!extra_cs; ++i) {
@@ -667,7 +725,7 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                                continue;
 
                        if (unique_bo_count == 0 && !cs->num_virtual_buffers) {
-                               memcpy(handles, cs->handles, cs->num_buffers * sizeof(amdgpu_bo_handle));
+                               memcpy(handles, cs->handles, cs->num_buffers * sizeof(struct drm_amdgpu_bo_list_entry));
                                unique_bo_count = cs->num_buffers;
                                continue;
                        }
@@ -675,7 +733,7 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                        for (unsigned j = 0; j < cs->num_buffers; ++j) {
                                bool found = false;
                                for (unsigned k = 0; k < unique_bo_so_far; ++k) {
-                                       if (handles[k] == cs->handles[j]) {
+                                       if (handles[k].bo_handle == cs->handles[j].bo_handle) {
                                                found = true;
                                                break;
                                        }
@@ -691,13 +749,14 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                                        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) {
+                                               if (handles[m].bo_handle == bo->bo_handle) {
                                                        found = true;
                                                        break;
                                                }
                                        }
                                        if (!found) {
-                                               handles[unique_bo_count] = bo->bo;
+                                               handles[unique_bo_count].bo_handle = bo->bo_handle;
+                                               handles[unique_bo_count].bo_priority = bo->priority;
                                                ++unique_bo_count;
                                        }
                                }
@@ -710,21 +769,22 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
                                struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(radv_bo_list->bos[i]);
                                bool found = false;
                                for (unsigned j = 0; j < unique_bo_so_far; ++j) {
-                                       if (bo->bo == handles[j]) {
+                                       if (bo->bo_handle == handles[j].bo_handle) {
                                                found = true;
                                                break;
                                        }
                                }
                                if (!found) {
-                                       handles[unique_bo_count] = bo->bo;
+                                       handles[unique_bo_count].bo_handle = bo->bo_handle;
+                                       handles[unique_bo_count].bo_priority = bo->priority;
                                        ++unique_bo_count;
                                }
                        }
                }
 
                if (unique_bo_count > 0) {
-                       r = amdgpu_bo_list_create(ws->dev, unique_bo_count, handles,
-                                                 NULL, bo_list);
+                       r = amdgpu_bo_list_create_raw(ws->dev, unique_bo_count, handles,
+                                                     bo_list);
                } else {
                        *bo_list = 0;
                }
@@ -746,7 +806,7 @@ static struct amdgpu_cs_fence_info radv_set_cs_fence(struct radv_amdgpu_ctx *ctx
 }
 
 static void radv_assign_last_submit(struct radv_amdgpu_ctx *ctx,
-                                   struct amdgpu_cs_request *request)
+                                   struct radv_amdgpu_cs_request *request)
 {
        radv_amdgpu_request_to_fence(ctx,
                                     &ctx->last_submission[request->ip_type][request->ring],
@@ -767,8 +827,8 @@ static int radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
        struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
        struct radv_amdgpu_fence *fence = (struct radv_amdgpu_fence *)_fence;
        struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[0]);
-       amdgpu_bo_list_handle bo_list;
-       struct amdgpu_cs_request request = {0};
+       uint32_t bo_list;
+       struct radv_amdgpu_cs_request request = {0};
        struct amdgpu_cs_ib_info ibs[2];
        unsigned number_of_ibs = 1;
 
@@ -830,8 +890,7 @@ static int radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
                                        "see dmesg for more information.\n");
        }
 
-       if (bo_list)
-               amdgpu_bo_list_destroy(bo_list);
+       amdgpu_bo_list_destroy_raw(ctx->ws->dev, bo_list);
 
        if (r)
                return r;
@@ -857,8 +916,8 @@ static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
        int r;
        struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
        struct radv_amdgpu_fence *fence = (struct radv_amdgpu_fence *)_fence;
-       amdgpu_bo_list_handle bo_list;
-       struct amdgpu_cs_request request = {};
+       uint32_t bo_list;
+       struct radv_amdgpu_cs_request request = {};
        struct amdgpu_cs_ib_info *ibs;
        struct radv_amdgpu_cs *cs0;
        unsigned number_of_ibs;
@@ -881,8 +940,7 @@ static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
 
        ibs = malloc(number_of_ibs * sizeof(*ibs));
        if (!ibs) {
-               if (bo_list)
-                       amdgpu_bo_list_destroy(bo_list);
+               amdgpu_bo_list_destroy_raw(ctx->ws->dev, bo_list);
                return -ENOMEM;
        }
 
@@ -918,8 +976,7 @@ static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
                                        "see dmesg for more information.\n");
        }
 
-       if (bo_list)
-               amdgpu_bo_list_destroy(bo_list);
+       amdgpu_bo_list_destroy_raw(ctx->ws->dev, bo_list);
        free(ibs);
 
        if (r)
@@ -948,8 +1005,8 @@ static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
        struct radv_amdgpu_fence *fence = (struct radv_amdgpu_fence *)_fence;
        struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[0]);
        struct radeon_winsys *ws = (struct radeon_winsys*)cs0->ws;
-       amdgpu_bo_list_handle bo_list;
-       struct amdgpu_cs_request request;
+       uint32_t bo_list;
+       struct radv_amdgpu_cs_request request;
        uint32_t pad_word = 0xffff1000U;
        bool emit_signal_sem = sem_info->cs_emit_signal;
 
@@ -1112,8 +1169,7 @@ static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
                                                "see dmesg for more information.\n");
                }
 
-               if (bo_list)
-                       amdgpu_bo_list_destroy(bo_list);
+               amdgpu_bo_list_destroy_raw(ctx->ws->dev, bo_list);
 
                for (unsigned j = 0; j < number_of_ibs; j++) {
                        ws->buffer_destroy(bos[j]);
@@ -1342,7 +1398,7 @@ static struct drm_amdgpu_cs_chunk_sem *radv_amdgpu_cs_alloc_syncobj_chunk(struct
 }
 
 static int radv_amdgpu_cs_submit(struct radv_amdgpu_ctx *ctx,
-                                struct amdgpu_cs_request *request,
+                                struct radv_amdgpu_cs_request *request,
                                 struct radv_winsys_sem_info *sem_info)
 {
        int r;
@@ -1444,7 +1500,7 @@ static int radv_amdgpu_cs_submit(struct radv_amdgpu_ctx *ctx,
                num_chunks++;
        }
 
-       r = amdgpu_cs_submit_raw(ctx->ws->dev,
+       r = amdgpu_cs_submit_raw2(ctx->ws->dev,
                                 ctx->ctx,
                                 request->resources,
                                 num_chunks,