radv: Update CTS version.
[mesa.git] / src / amd / vulkan / radv_device.c
index dda535d553822d4468f9bb6c1dda5f55e2c2c24e..5139cf6fcbebb4cec341f6e9f34b36219170b10f 100644 (file)
@@ -380,7 +380,7 @@ radv_physical_device_try_create(struct radv_instance *instance,
        disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
        device->disk_cache = disk_cache_create(device->name, buf, shader_env_flags);
 
-       if (device->rad_info.chip_class < GFX8 || !device->use_llvm)
+       if (device->rad_info.chip_class < GFX8)
                fprintf(stderr, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
 
        radv_get_driver_uuid(&device->driver_uuid);
@@ -1645,7 +1645,7 @@ radv_get_physical_device_properties_1_2(struct radv_physical_device *pdevice,
        p->conformanceVersion = (VkConformanceVersion) {
                .major = 1,
                .minor = 2,
-               .subminor = 0,
+               .subminor = 3,
                .patch = 0,
        };
 
@@ -2590,6 +2590,25 @@ static void radv_device_finish_border_color(struct radv_device *device)
        }
 }
 
+VkResult
+_radv_device_set_lost(struct radv_device *device,
+                     const char *file, int line,
+                     const char *msg, ...)
+{
+       VkResult err;
+       va_list ap;
+
+       p_atomic_inc(&device->lost);
+
+       va_start(ap, msg);
+       err = __vk_errorv(device->physical_device->instance, device,
+                         VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
+                         VK_ERROR_DEVICE_LOST, file, line, msg, ap);
+       va_end(ap);
+
+       return err;
+}
+
 VkResult radv_CreateDevice(
        VkPhysicalDevice                            physicalDevice,
        const VkDeviceCreateInfo*                   pCreateInfo,
@@ -2798,6 +2817,19 @@ VkResult radv_CreateDevice(
                        goto fail;
        }
 
+       if (getenv("RADV_TRAP_HANDLER")) {
+               /* TODO: Add support for more hardware. */
+               assert(device->physical_device->rad_info.chip_class == GFX8);
+
+               /* To get the disassembly of the faulty shaders, we have to
+                * keep some shader info around.
+                */
+               keep_shader_info = true;
+
+               if (!radv_trap_handler_init(device))
+                       goto fail;
+       }
+
        device->keep_shader_info = keep_shader_info;
        result = radv_device_init_meta(device);
        if (result != VK_SUCCESS)
@@ -2874,6 +2906,8 @@ fail:
 
        radv_thread_trace_finish(device);
 
+       radv_trap_handler_finish(device);
+
        if (device->trace_bo)
                device->ws->buffer_destroy(device->trace_bo);
 
@@ -2923,6 +2957,8 @@ void radv_DestroyDevice(
        VkPipelineCache pc = radv_pipeline_cache_to_handle(device->mem_cache);
        radv_DestroyPipelineCache(radv_device_to_handle(device), pc, NULL);
 
+       radv_trap_handler_finish(device);
+
        radv_destroy_shader_slabs(device);
 
        pthread_cond_destroy(&device->timeline_cond);
@@ -3401,6 +3437,50 @@ radv_emit_global_shader_pointers(struct radv_queue *queue,
        }
 }
 
+static void
+radv_emit_trap_handler(struct radv_queue *queue,
+                      struct radeon_cmdbuf *cs,
+                      struct radeon_winsys_bo *tma_bo)
+{
+       struct radv_device *device = queue->device;
+       struct radeon_winsys_bo *tba_bo;
+       uint64_t tba_va, tma_va;
+
+       if (!device->trap_handler_shader || !tma_bo)
+               return;
+
+       tba_bo = device->trap_handler_shader->bo;
+
+       tba_va = radv_buffer_get_va(tba_bo) + device->trap_handler_shader->bo_offset;
+       tma_va = radv_buffer_get_va(tma_bo);
+
+       radv_cs_add_buffer(queue->device->ws, cs, tba_bo);
+       radv_cs_add_buffer(queue->device->ws, cs, tma_bo);
+
+       if (queue->queue_family_index == RADV_QUEUE_GENERAL) {
+               uint32_t regs[] = {R_00B000_SPI_SHADER_TBA_LO_PS,
+                                  R_00B100_SPI_SHADER_TBA_LO_VS,
+                                  R_00B200_SPI_SHADER_TBA_LO_GS,
+                                  R_00B300_SPI_SHADER_TBA_LO_ES,
+                                  R_00B400_SPI_SHADER_TBA_LO_HS,
+                                  R_00B500_SPI_SHADER_TBA_LO_LS};
+
+               for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
+                       radeon_set_sh_reg_seq(cs, regs[i], 4);
+                       radeon_emit(cs, tba_va >> 8);
+                       radeon_emit(cs, tba_va >> 40);
+                       radeon_emit(cs, tma_va >> 8);
+                       radeon_emit(cs, tma_va >> 40);
+               }
+       } else {
+               radeon_set_sh_reg_seq(cs, R_00B838_COMPUTE_TBA_LO, 4);
+               radeon_emit(cs, tba_va >> 8);
+               radeon_emit(cs, tba_va >> 40);
+               radeon_emit(cs, tma_va >> 8);
+               radeon_emit(cs, tma_va >> 40);
+       }
+}
+
 static void
 radv_init_graphics_state(struct radeon_cmdbuf *cs, struct radv_queue *queue)
 {
@@ -3705,6 +3785,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
                                          compute_scratch_waves, compute_scratch_bo);
                radv_emit_graphics_scratch(queue, cs, scratch_size_per_wave,
                                           scratch_waves, scratch_bo);
+               radv_emit_trap_handler(queue, cs, queue->device->tma_bo);
 
                if (gds_bo)
                        radv_cs_add_buffer(queue->device->ws, cs, gds_bo);
@@ -4470,6 +4551,10 @@ radv_queue_submit_deferred(struct radv_deferred_queue_submission *submission,
                        if (queue->device->trace_bo) {
                                radv_check_gpu_hangs(queue, cs_array[j]);
                        }
+
+                       if (queue->device->tma_bo) {
+                               radv_check_trap_handler(queue);
+                       }
                }
 
                free(cs_array);
@@ -4503,7 +4588,7 @@ fail:
                 * VK_ERROR_DEVICE_LOST to ensure the clients do not attempt
                 * to submit the same job again to this device.
                 */
-               result = VK_ERROR_DEVICE_LOST;
+               result = radv_device_set_lost(queue->device, "vkQueueSubmit() failed");
        }
 
        radv_free_temp_syncobjs(queue->device,
@@ -4724,6 +4809,9 @@ VkResult radv_QueueSubmit(
        uint32_t fence_idx = 0;
        bool flushed_caches = false;
 
+       if (radv_device_is_lost(queue->device))
+               return VK_ERROR_DEVICE_LOST;
+
        if (fence != VK_NULL_HANDLE) {
                for (uint32_t i = 0; i < submitCount; ++i)
                        if (radv_submit_has_effects(pSubmits + i))
@@ -4793,6 +4881,9 @@ VkResult radv_QueueWaitIdle(
 {
        RADV_FROM_HANDLE(radv_queue, queue, _queue);
 
+       if (radv_device_is_lost(queue->device))
+               return VK_ERROR_DEVICE_LOST;
+
        pthread_mutex_lock(&queue->pending_mutex);
        while (!list_is_empty(&queue->pending_submissions)) {
                pthread_cond_wait(&queue->device->timeline_cond, &queue->pending_mutex);
@@ -4802,9 +4893,10 @@ VkResult radv_QueueWaitIdle(
        if (!queue->device->ws->ctx_wait_idle(queue->hw_ctx,
                                              radv_queue_family_to_ring(queue->queue_family_index),
                                              queue->queue_idx)) {
-               return vk_errorf(queue->device->instance, VK_ERROR_DEVICE_LOST,
-                                "Failed to wait for a '%s' queue to be idle. "
-                                "GPU hang ?", radv_get_queue_family_name(queue));
+               return radv_device_set_lost(queue->device,
+                                           "Failed to wait for a '%s' queue "
+                                           "to be idle. GPU hang ?",
+                                           radv_get_queue_family_name(queue));
        }
 
        return VK_SUCCESS;
@@ -5471,6 +5563,9 @@ static bool radv_sparse_bind_has_effects(const VkBindSparseInfo *info)
        VkResult result;
        uint32_t fence_idx = 0;
 
+       if (radv_device_is_lost(queue->device))
+               return VK_ERROR_DEVICE_LOST;
+
        if (fence != VK_NULL_HANDLE) {
                for (uint32_t i = 0; i < bindInfoCount; ++i)
                        if (radv_sparse_bind_has_effects(pBindInfo + i))
@@ -5653,6 +5748,10 @@ VkResult radv_WaitForFences(
        uint64_t                                    timeout)
 {
        RADV_FROM_HANDLE(radv_device, device, _device);
+
+       if (radv_device_is_lost(device))
+               return VK_ERROR_DEVICE_LOST;
+
        timeout = radv_get_absolute_timeout(timeout);
 
        if (device->always_use_syncobj &&
@@ -5809,6 +5908,9 @@ VkResult radv_GetFenceStatus(VkDevice _device, VkFence _fence)
                fence->temporary.kind != RADV_FENCE_NONE ?
                &fence->temporary : &fence->permanent;
 
+       if (radv_device_is_lost(device))
+               return VK_ERROR_DEVICE_LOST;
+
        switch (part->kind) {
        case RADV_FENCE_NONE:
                break;
@@ -6134,6 +6236,9 @@ radv_GetSemaphoreCounterValue(VkDevice _device,
        RADV_FROM_HANDLE(radv_device, device, _device);
        RADV_FROM_HANDLE(radv_semaphore, semaphore, _semaphore);
 
+       if (radv_device_is_lost(device))
+               return VK_ERROR_DEVICE_LOST;
+
        struct radv_semaphore_part *part =
                semaphore->temporary.kind != RADV_SEMAPHORE_NONE ? &semaphore->temporary : &semaphore->permanent;
 
@@ -6191,6 +6296,10 @@ radv_WaitSemaphores(VkDevice _device,
                    uint64_t timeout)
 {
        RADV_FROM_HANDLE(radv_device, device, _device);
+
+       if (radv_device_is_lost(device))
+               return VK_ERROR_DEVICE_LOST;
+
        uint64_t abs_timeout = radv_get_absolute_timeout(timeout);
 
        if (radv_semaphore_from_handle(pWaitInfo->pSemaphores[0])->permanent.kind == RADV_SEMAPHORE_TIMELINE)
@@ -6327,8 +6436,12 @@ VkResult radv_GetEventStatus(
        VkDevice                                    _device,
        VkEvent                                     _event)
 {
+       RADV_FROM_HANDLE(radv_device, device, _device);
        RADV_FROM_HANDLE(radv_event, event, _event);
 
+       if (radv_device_is_lost(device))
+               return VK_ERROR_DEVICE_LOST;
+
        if (*event->map == 1)
                return VK_EVENT_SET;
        return VK_EVENT_RESET;