diff --git a/src/vulkan/common.h b/src/vulkan/common.h
index daeba2019e894b366948247376d0d9888178c787..ead439c4999bbadee823d8410bf0205218f7c88d 100644
--- a/src/vulkan/common.h
+++ b/src/vulkan/common.h
@@ -46,6 +46,11 @@
#define PL_VK_MAX_QUEUED_CMDS 1024
#define PL_VK_MAX_PENDING_CMDS 1024
+// Shitty compatibility alias for very old vulkan.h versions
+#ifndef VK_API_VERSION_1_2
+#define VK_API_VERSION_1_2 VK_MAKE_VERSION(1, 2, 0)
+#endif
+
// Shared struct used to hold vulkan context information
struct vk_ctx {
void *ta; // allocations bound to the lifetime of this vk_ctx
@@ -205,6 +210,7 @@ struct vk_ctx {
VK_FUN(QueueSubmit);
VK_FUN(ResetEvent);
VK_FUN(ResetFences);
+ VK_FUN(ResetQueryPoolEXT);
VK_FUN(SetDebugUtilsObjectNameEXT);
VK_FUN(SetHdrMetadataEXT);
VK_FUN(UpdateDescriptorSets);
diff --git a/src/vulkan/context.c b/src/vulkan/context.c
index 89bef52118d482b37fe1c1c8ef145e117c9016c9..fac07eb818b0a9a8901f6cd17dd5c988764f6b2b 100644
--- a/src/vulkan/context.c
+++ b/src/vulkan/context.c
@@ -24,6 +24,7 @@ const struct pl_vk_inst_params pl_vk_inst_default_params = {0};
struct vk_fun {
const char *name;
+ const char *alias;
size_t offset;
bool device_level;
};
@@ -45,6 +46,13 @@ struct vk_ext {
.device_level = true, \
}
+#define VK_DEV_FUN_ALIAS(N, ALIAS) \
+ { .name = "vk" #N, \
+ .alias = #ALIAS, \
+ .offset = offsetof(struct vk_ctx, N), \
+ .device_level = true, \
+ }
+
// Table of optional vulkan instance extensions
static const char *vk_instance_extensions[] = {
VK_KHR_SURFACE_EXTENSION_NAME,
@@ -156,6 +164,13 @@ static const struct vk_ext vk_device_extensions[] = {
VK_DEV_FUN(SetHdrMetadataEXT),
{0},
},
+ }, {
+ .name = VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
+ .core_ver = VK_API_VERSION_1_2,
+ .funs = (struct vk_fun[]) {
+ VK_DEV_FUN_ALIAS(ResetQueryPoolEXT, vkResetQueryPool),
+ {0},
+ },
},
};
@@ -173,13 +188,21 @@ const char * const pl_vulkan_recommended_extensions[] = {
#endif
VK_EXT_PCI_BUS_INFO_EXTENSION_NAME,
VK_EXT_HDR_METADATA_EXTENSION_NAME,
+ VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
};
const int pl_vulkan_num_recommended_extensions =
PL_ARRAY_SIZE(pl_vulkan_recommended_extensions);
+// pNext chain of features we want enabled
+static const VkPhysicalDeviceHostQueryResetFeaturesEXT host_query_reset = {
+ .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT,
+ .hostQueryReset = true,
+};
+
const VkPhysicalDeviceFeatures2KHR pl_vulkan_recommended_features = {
.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR,
+ .pNext = (void *) &host_query_reset,
.features = {
.shaderImageGatherExtended = true,
@@ -1161,6 +1184,8 @@ static bool device_init(struct vk_ctx *vk, const struct pl_vulkan_params *params
PFN_vkVoidFunction *pfn = (void *) ((uintptr_t) vk + (ptrdiff_t) fun->offset);
if (fun->device_level) {
*pfn = vk->GetDeviceProcAddr(vk->dev, fun->name);
+ if (fun->alias && !*pfn)
+ *pfn = vk->GetDeviceProcAddr(vk->dev, fun->alias);
} else {
*pfn = vk->GetInstanceProcAddr(vk->inst, fun->name);
};
diff --git a/src/vulkan/gpu.c b/src/vulkan/gpu.c
index 8af8798d659ace8348a0da409e61f339b851112d..2e7140e94498b4b77d08056091dac675d15d9ff6 100644
--- a/src/vulkan/gpu.c
+++ b/src/vulkan/gpu.c
@@ -40,9 +40,10 @@ struct pl_vk {
struct vk_malloc *alloc;
struct spirv_compiler *spirv;
- // Some additional cached device limits
+ // Some additional cached device limits and features checks
uint32_t max_push_descriptors;
size_t min_texel_alignment;
+ bool host_query_reset;
// This is a pl_dispatch used (on ourselves!) for the purposes of
// dispatching compute shaders for performing various emulation tasks
@@ -119,7 +120,7 @@ static inline bool supports_marks(struct vk_cmd *cmd) {
} while (0)
#define MAKE_LAZY_DESTRUCTOR(fun, argtype) \
- static void fun##_lazy(const struct pl_gpu *gpu, const argtype *arg) { \
+ static void fun##_lazy(const struct pl_gpu *gpu, argtype *arg) { \
struct pl_vk *p = TA_PRIV(gpu); \
struct vk_ctx *vk = p->vk; \
if (p->cmd) { \
@@ -419,6 +420,15 @@ const struct pl_gpu *pl_gpu_create_vk(struct vk_ctx *vk)
p->max_push_descriptors = pushd.maxPushDescriptors;
}
+ if (vk->ResetQueryPoolEXT) {
+ const VkPhysicalDeviceHostQueryResetFeaturesEXT *host_query_reset;
+ host_query_reset = vk_find_struct(&vk->features,
+ VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT);
+
+ if (host_query_reset)
+ p->host_query_reset = host_query_reset->hostQueryReset;
+ }
+
// We ostensibly support this, although it can still fail on buffer
// creation (for certain combinations of buffers)
gpu->caps |= PL_GPU_CAP_MAPPED_BUFFERS;
@@ -678,7 +688,7 @@ static void vk_tex_destroy(const struct pl_gpu *gpu, struct pl_tex *tex)
talloc_free(tex);
}
-MAKE_LAZY_DESTRUCTOR(vk_tex_destroy, struct pl_tex)
+MAKE_LAZY_DESTRUCTOR(vk_tex_destroy, const struct pl_tex)
static const VkFilter filters[] = {
[PL_TEX_SAMPLE_NEAREST] = VK_FILTER_NEAREST,
@@ -1487,7 +1497,7 @@ static void buf_flush(const struct pl_gpu *gpu, struct vk_cmd *cmd,
}
#define vk_buf_destroy vk_buf_deref
-MAKE_LAZY_DESTRUCTOR(vk_buf_destroy, struct pl_buf)
+MAKE_LAZY_DESTRUCTOR(vk_buf_destroy, const struct pl_buf)
static void vk_buf_write(const struct pl_gpu *gpu, const struct pl_buf *buf,
size_t offset, const void *data, size_t size)
@@ -2084,7 +2094,7 @@ static void vk_pass_destroy(const struct pl_gpu *gpu, struct pl_pass *pass)
talloc_free(pass);
}
-MAKE_LAZY_DESTRUCTOR(vk_pass_destroy, struct pl_pass)
+MAKE_LAZY_DESTRUCTOR(vk_pass_destroy, const struct pl_pass)
static const VkDescriptorType dsType[] = {
[PL_DESC_SAMPLED_TEX] = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
@@ -3079,27 +3089,29 @@ error:
#define VK_QUERY_POOL_SIZE 16
struct pl_timer {
- int refcount;
bool recording; // true between vk_cmd_timer_begin() and vk_cmd_timer_end()
VkQueryPool qpool; // even=start, odd=stop
int index_write; // next index to write to
int index_read; // next index to read from
+ uint64_t pending; // bitmask of queries that are still running
};
+static inline uint64_t timer_bit(int index)
+{
+ return 1llu << (index / 2);
+}
+
static void vk_timer_destroy(const struct pl_gpu *gpu, struct pl_timer *timer)
{
struct pl_vk *p = TA_PRIV(gpu);
struct vk_ctx *vk = p->vk;
+ pl_assert(!timer->pending);
vk->DestroyQueryPool(vk->dev, timer->qpool, VK_ALLOC);
talloc_free(timer);
}
-static void vk_timer_deref(const struct pl_gpu *gpu, struct pl_timer *timer)
-{
- if (--timer->refcount == 0)
- vk_timer_destroy(gpu, timer);
-}
+MAKE_LAZY_DESTRUCTOR(vk_timer_destroy, struct pl_timer)
static struct pl_timer *vk_timer_create(const struct pl_gpu *gpu)
{
@@ -3107,9 +3119,7 @@ static struct pl_timer *vk_timer_create(const struct pl_gpu *gpu)
struct vk_ctx *vk = p->vk;
struct pl_timer *timer = talloc_ptrtype(NULL, timer);
- *timer = (struct pl_timer) {
- .refcount = 1,
- };
+ *timer = (struct pl_timer) {0};
struct VkQueryPoolCreateInfo qinfo = {
.sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
@@ -3167,19 +3177,36 @@ static void vk_cmd_timer_begin(const struct pl_gpu *gpu, struct vk_cmd *cmd,
return;
}
+ vk_poll_commands(vk, 0);
+ if (timer->pending & timer_bit(timer->index_write))
+ return; // next query is still running, skip this timer
+
VkQueueFlags reset_flags = VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT;
- if (!(cmd->pool->props.queueFlags & reset_flags)) {
- PL_TRACE(gpu, "QF %d does not support query pool resets", cmd->pool->qf);
+ if (cmd->pool->props.queueFlags & reset_flags) {
+ // Use direct command buffer resets
+ vk->CmdResetQueryPool(cmd->buf, timer->qpool, timer->index_write, 2);
+ } else if (p->host_query_reset) {
+ // Use host query resets
+ vk->ResetQueryPoolEXT(vk->dev, timer->qpool, timer->index_write, 2);
+ } else {
+ PL_TRACE(gpu, "QF %d supports no mechanism for resetting queries",
+ cmd->pool->qf);
return;
}
- vk->CmdResetQueryPool(cmd->buf, timer->qpool, timer->index_write, 2);
vk->CmdWriteTimestamp(cmd->buf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
timer->qpool, timer->index_write);
timer->recording = true;
}
+static void vk_timer_cb(void *ptimer, void *pindex)
+{
+ struct pl_timer *timer = ptimer;
+ int index = (uintptr_t) pindex;
+ timer->pending &= ~timer_bit(index);
+}
+
static void vk_cmd_timer_end(const struct pl_gpu *gpu, struct vk_cmd *cmd,
struct pl_timer *timer)
{
@@ -3192,18 +3219,18 @@ static void vk_cmd_timer_end(const struct pl_gpu *gpu, struct vk_cmd *cmd,
vk->CmdWriteTimestamp(cmd->buf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
timer->qpool, timer->index_write + 1);
+ timer->recording = false;
+ timer->pending |= timer_bit(timer->index_write);
+ vk_cmd_callback(cmd, (vk_cb) vk_timer_cb, timer,
+ (void *) (uintptr_t) timer->index_write);
+
timer->index_write = (timer->index_write + 2) % VK_QUERY_POOL_SIZE;
if (timer->index_write == timer->index_read) {
// forcibly drop the least recent result to make space
timer->index_read = (timer->index_read + 2) % VK_QUERY_POOL_SIZE;
}
-
- timer->recording = false;
- timer->refcount++;
- vk_cmd_callback(cmd, (vk_cb) vk_timer_deref, gpu, timer);
}
-
static void vk_gpu_flush(const struct pl_gpu *gpu)
{
struct pl_vk *p = TA_PRIV(gpu);
@@ -3252,7 +3279,7 @@ static const struct pl_gpu_fns pl_fns_vk = {
.sync_destroy = vk_sync_deref,
.tex_export = vk_tex_export,
.timer_create = vk_timer_create,
- .timer_destroy = vk_timer_deref,
+ .timer_destroy = vk_timer_destroy_lazy,
.timer_query = vk_timer_query,
.gpu_flush = vk_gpu_flush,
.gpu_finish = vk_gpu_finish,