#include "anv_private.h"
#include "vk_format_info.h"
+#include "vk_util.h"
/** \file anv_cmd_buffer.c
*
.front = 0u,
.back = 0u,
},
+ .line_stipple = {
+ .factor = 0u,
+ .pattern = 0u,
+ },
};
-void
+/**
+ * Copy the dynamic state from src to dest based on the copy_mask.
+ *
+ * Avoid copying states that have not changed, except for VIEWPORT, SCISSOR and
+ * BLEND_CONSTANTS (always copy them if they are in the copy_mask).
+ *
+ * Returns a mask of the states which changed.
+ */
+anv_cmd_dirty_mask_t
anv_dynamic_state_copy(struct anv_dynamic_state *dest,
const struct anv_dynamic_state *src,
- uint32_t copy_mask)
+ anv_cmd_dirty_mask_t copy_mask)
{
- if (copy_mask & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
+ anv_cmd_dirty_mask_t changed = 0;
+
+ if (copy_mask & ANV_CMD_DIRTY_DYNAMIC_VIEWPORT) {
dest->viewport.count = src->viewport.count;
typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
src->viewport.count);
+ changed |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
}
- if (copy_mask & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
+ if (copy_mask & ANV_CMD_DIRTY_DYNAMIC_SCISSOR) {
dest->scissor.count = src->scissor.count;
typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
src->scissor.count);
+ changed |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
}
- if (copy_mask & (1 << VK_DYNAMIC_STATE_LINE_WIDTH))
- dest->line_width = src->line_width;
+ if (copy_mask & ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS) {
+ typed_memcpy(dest->blend_constants, src->blend_constants, 4);
+ changed |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
+ }
+
+#define ANV_CMP_COPY(field, flag) \
+ if (copy_mask & flag) { \
+ if (dest->field != src->field) { \
+ dest->field = src->field; \
+ changed |= flag; \
+ } \
+ }
- if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS))
- dest->depth_bias = src->depth_bias;
+ ANV_CMP_COPY(line_width, ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH);
- if (copy_mask & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS))
- typed_memcpy(dest->blend_constants, src->blend_constants, 4);
+ ANV_CMP_COPY(depth_bias.bias, ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS);
+ ANV_CMP_COPY(depth_bias.clamp, ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS);
+ ANV_CMP_COPY(depth_bias.slope, ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS);
+
+ ANV_CMP_COPY(depth_bounds.min, ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS);
+ ANV_CMP_COPY(depth_bounds.max, ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS);
+
+ ANV_CMP_COPY(stencil_compare_mask.front, ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK);
+ ANV_CMP_COPY(stencil_compare_mask.back, ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK);
+
+ ANV_CMP_COPY(stencil_write_mask.front, ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK);
+ ANV_CMP_COPY(stencil_write_mask.back, ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK);
- if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS))
- dest->depth_bounds = src->depth_bounds;
+ ANV_CMP_COPY(stencil_reference.front, ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE);
+ ANV_CMP_COPY(stencil_reference.back, ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE);
- if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK))
- dest->stencil_compare_mask = src->stencil_compare_mask;
+ ANV_CMP_COPY(line_stipple.factor, ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE);
+ ANV_CMP_COPY(line_stipple.pattern, ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE);
- if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK))
- dest->stencil_write_mask = src->stencil_write_mask;
+#undef ANV_CMP_COPY
- if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE))
- dest->stencil_reference = src->stencil_reference;
+ return changed;
}
static void
anv_cmd_pipeline_state_finish(struct anv_cmd_buffer *cmd_buffer,
struct anv_cmd_pipeline_state *pipe_state)
{
- for (uint32_t i = 0; i < ARRAY_SIZE(pipe_state->push_descriptors); i++)
- vk_free(&cmd_buffer->pool->alloc, pipe_state->push_descriptors[i]);
+ for (uint32_t i = 0; i < ARRAY_SIZE(pipe_state->push_descriptors); i++) {
+ if (pipe_state->push_descriptors[i]) {
+ anv_descriptor_set_layout_unref(cmd_buffer->device,
+ pipe_state->push_descriptors[i]->set.layout);
+ vk_free(&cmd_buffer->pool->alloc, pipe_state->push_descriptors[i]);
+ }
+ }
}
static void
anv_cmd_pipeline_state_finish(cmd_buffer, &state->gfx.base);
anv_cmd_pipeline_state_finish(cmd_buffer, &state->compute.base);
- for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++)
- vk_free(&cmd_buffer->pool->alloc, state->push_constants[i]);
-
vk_free(&cmd_buffer->pool->alloc, state->attachments);
}
anv_cmd_state_init(cmd_buffer);
}
-/**
- * This function updates the size of the push constant buffer we need to emit.
- * This is called in various parts of the driver to ensure that different
- * pieces of push constant data get emitted as needed. However, it is important
- * that we never shrink the size of the buffer. For example, a compute shader
- * dispatch will always call this for the base group id, which has an
- * offset in the push constant buffer that is smaller than the offset for
- * storage image data. If the compute shader has storage images, we will call
- * this again with a larger size during binding table emission. However,
- * if we dispatch the compute shader again without dirtying our descriptors,
- * we would still call this function with a smaller size for the base group
- * id, and not for the images, which would incorrectly shrink the size of the
- * push constant data we emit with that dispatch, making us drop the image data.
- */
-VkResult
-anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
- gl_shader_stage stage, uint32_t size)
-{
- struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
-
- if (*ptr == NULL) {
- *ptr = vk_alloc(&cmd_buffer->pool->alloc, size, 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
- if (*ptr == NULL) {
- anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- }
- (*ptr)->size = size;
- } else if ((*ptr)->size < size) {
- *ptr = vk_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
- if (*ptr == NULL) {
- anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- }
- (*ptr)->size = size;
- }
-
- return VK_SUCCESS;
-}
-
static VkResult anv_create_cmd_buffer(
struct anv_device * device,
struct anv_cmd_pool * pool,
cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
/* Apply the dynamic state from the pipeline */
- cmd_buffer->state.gfx.dirty |= pipeline->dynamic_state_mask;
- anv_dynamic_state_copy(&cmd_buffer->state.gfx.dynamic,
- &pipeline->dynamic_state,
- pipeline->dynamic_state_mask);
+ cmd_buffer->state.gfx.dirty |=
+ anv_dynamic_state_copy(&cmd_buffer->state.gfx.dynamic,
+ &pipeline->dynamic_state,
+ pipeline->dynamic_state_mask);
break;
default:
cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
}
+void anv_CmdSetLineStippleEXT(
+ VkCommandBuffer commandBuffer,
+ uint32_t lineStippleFactor,
+ uint16_t lineStipplePattern)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ cmd_buffer->state.gfx.dynamic.line_stipple.factor = lineStippleFactor;
+ cmd_buffer->state.gfx.dynamic.line_stipple.pattern = lineStipplePattern;
+
+ cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE;
+}
+
static void
anv_cmd_buffer_bind_descriptor_set(struct anv_cmd_buffer *cmd_buffer,
VkPipelineBindPoint bind_point,
*dynamic_offsets += set_layout->dynamic_offset_count;
*dynamic_offset_count -= set_layout->dynamic_offset_count;
+
+ if (bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) {
+ cmd_buffer->state.push_constants_dirty |=
+ VK_SHADER_STAGE_COMPUTE_BIT;
+ } else {
+ cmd_buffer->state.push_constants_dirty |=
+ VK_SHADER_STAGE_ALL_GRAPHICS;
+ }
}
}
}
static uint32_t
-anv_push_constant_value(struct anv_push_constants *data, uint32_t param)
+anv_push_constant_value(const struct anv_cmd_pipeline_state *state,
+ const struct anv_push_constants *data, uint32_t param)
{
if (BRW_PARAM_IS_BUILTIN(param)) {
switch (param) {
default:
unreachable("Invalid param builtin");
}
- } else {
+ } else if (ANV_PARAM_IS_PUSH(param)) {
uint32_t offset = ANV_PARAM_PUSH_OFFSET(param);
assert(offset % sizeof(uint32_t) == 0);
- if (offset < data->size)
+ if (offset < sizeof(data->client_data))
return *(uint32_t *)((uint8_t *)data + offset);
else
return 0;
+ } else if (ANV_PARAM_IS_DYN_OFFSET(param)) {
+ unsigned idx = ANV_PARAM_DYN_OFFSET_IDX(param);
+ assert(idx < MAX_DYNAMIC_BUFFERS);
+ return state->dynamic_offsets[idx];
}
+
+ assert(!"Invalid param");
+ return 0;
}
struct anv_state
anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
gl_shader_stage stage)
{
+ struct anv_cmd_pipeline_state *pipeline_state = &cmd_buffer->state.gfx.base;
struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
/* If we don't have this stage, bail. */
return (struct anv_state) { .offset = 0 };
struct anv_push_constants *data =
- cmd_buffer->state.push_constants[stage];
+ &cmd_buffer->state.push_constants[stage];
const struct brw_stage_prog_data *prog_data =
pipeline->shaders[stage]->prog_data;
/* If we don't actually have any push constants, bail. */
- if (data == NULL || prog_data == NULL || prog_data->nr_params == 0)
+ if (prog_data == NULL || prog_data->nr_params == 0)
return (struct anv_state) { .offset = 0 };
struct anv_state state =
/* Walk through the param array and fill the buffer with data */
uint32_t *u32_map = state.map;
- for (unsigned i = 0; i < prog_data->nr_params; i++)
- u32_map[i] = anv_push_constant_value(data, prog_data->param[i]);
+ for (unsigned i = 0; i < prog_data->nr_params; i++) {
+ u32_map[i] = anv_push_constant_value(pipeline_state, data,
+ prog_data->param[i]);
+ }
return state;
}
struct anv_state
anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
{
+ struct anv_cmd_pipeline_state *pipeline_state = &cmd_buffer->state.compute.base;
struct anv_push_constants *data =
- cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
+ &cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
i < cs_prog_data->push.cross_thread.dwords;
i++) {
assert(prog_data->param[i] != BRW_PARAM_BUILTIN_SUBGROUP_ID);
- u32_map[i] = anv_push_constant_value(data, prog_data->param[i]);
+ u32_map[i] = anv_push_constant_value(pipeline_state, data,
+ prog_data->param[i]);
}
}
if (prog_data->param[src] == BRW_PARAM_BUILTIN_SUBGROUP_ID) {
u32_map[dst] = t;
} else {
- u32_map[dst] =
- anv_push_constant_value(data, prog_data->param[src]);
+ u32_map[dst] = anv_push_constant_value(pipeline_state, data,
+ prog_data->param[src]);
}
}
}
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
anv_foreach_stage(stage, stageFlags) {
- VkResult result =
- anv_cmd_buffer_ensure_push_constant_field(cmd_buffer,
- stage, client_data);
- if (result != VK_SUCCESS)
- return;
-
- memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
+ memcpy(cmd_buffer->state.push_constants[stage].client_data + offset,
pValues, size);
}
anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
{
const struct anv_subpass *subpass = cmd_buffer->state.subpass;
- const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
if (subpass->depth_stencil_attachment == NULL)
return NULL;
const struct anv_image_view *iview =
- fb->attachments[subpass->depth_stencil_attachment->attachment];
+ cmd_buffer->state.attachments[subpass->depth_stencil_attachment->attachment].image_view;
assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
VK_IMAGE_ASPECT_STENCIL_BIT));
return iview;
}
-static struct anv_push_descriptor_set *
-anv_cmd_buffer_get_push_descriptor_set(struct anv_cmd_buffer *cmd_buffer,
- VkPipelineBindPoint bind_point,
- uint32_t set)
+static struct anv_descriptor_set *
+anv_cmd_buffer_push_descriptor_set(struct anv_cmd_buffer *cmd_buffer,
+ VkPipelineBindPoint bind_point,
+ struct anv_descriptor_set_layout *layout,
+ uint32_t _set)
{
struct anv_cmd_pipeline_state *pipe_state;
if (bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) {
}
struct anv_push_descriptor_set **push_set =
- &pipe_state->push_descriptors[set];
+ &pipe_state->push_descriptors[_set];
if (*push_set == NULL) {
- *push_set = vk_alloc(&cmd_buffer->pool->alloc,
- sizeof(struct anv_push_descriptor_set), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+ *push_set = vk_zalloc(&cmd_buffer->pool->alloc,
+ sizeof(struct anv_push_descriptor_set), 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (*push_set == NULL) {
anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
return NULL;
}
}
- return *push_set;
+ struct anv_descriptor_set *set = &(*push_set)->set;
+
+ if (set->layout != layout) {
+ if (set->layout)
+ anv_descriptor_set_layout_unref(cmd_buffer->device, set->layout);
+ anv_descriptor_set_layout_ref(layout);
+ set->layout = layout;
+ }
+ set->size = anv_descriptor_set_layout_size(layout);
+ set->buffer_view_count = layout->buffer_view_count;
+ set->buffer_views = (*push_set)->buffer_views;
+
+ if (layout->descriptor_buffer_size &&
+ ((*push_set)->set_used_on_gpu ||
+ set->desc_mem.alloc_size < layout->descriptor_buffer_size)) {
+ /* The previous buffer is either actively used by some GPU command (so
+ * we can't modify it) or is too small. Allocate a new one.
+ */
+ struct anv_state desc_mem =
+ anv_state_stream_alloc(&cmd_buffer->dynamic_state_stream,
+ layout->descriptor_buffer_size, 32);
+ if (set->desc_mem.alloc_size) {
+ /* TODO: Do we really need to copy all the time? */
+ memcpy(desc_mem.map, set->desc_mem.map,
+ MIN2(desc_mem.alloc_size, set->desc_mem.alloc_size));
+ }
+ set->desc_mem = desc_mem;
+
+ struct anv_address addr = {
+ .bo = cmd_buffer->dynamic_state_stream.state_pool->block_pool.bo,
+ .offset = set->desc_mem.offset,
+ };
+
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+ set->desc_surface_state =
+ anv_state_stream_alloc(&cmd_buffer->surface_state_stream,
+ isl_dev->ss.size, isl_dev->ss.align);
+ anv_fill_buffer_surface_state(cmd_buffer->device,
+ set->desc_surface_state,
+ ISL_FORMAT_R32G32B32A32_FLOAT,
+ addr, layout->descriptor_buffer_size, 1);
+ }
+
+ return set;
}
void anv_CmdPushDescriptorSetKHR(
struct anv_descriptor_set_layout *set_layout = layout->set[_set].layout;
- struct anv_push_descriptor_set *push_set =
- anv_cmd_buffer_get_push_descriptor_set(cmd_buffer,
- pipelineBindPoint, _set);
- if (!push_set)
+ struct anv_descriptor_set *set =
+ anv_cmd_buffer_push_descriptor_set(cmd_buffer, pipelineBindPoint,
+ set_layout, _set);
+ if (!set)
return;
- struct anv_descriptor_set *set = &push_set->set;
-
- set->layout = set_layout;
- set->size = anv_descriptor_set_layout_size(set_layout);
- set->buffer_count = set_layout->buffer_count;
- set->buffer_views = push_set->buffer_views;
-
/* Go through the user supplied descriptors. */
for (uint32_t i = 0; i < descriptorWriteCount; i++) {
const VkWriteDescriptorSet *write = &pDescriptorWrites[i];
struct anv_descriptor_set_layout *set_layout = layout->set[_set].layout;
- struct anv_push_descriptor_set *push_set =
- anv_cmd_buffer_get_push_descriptor_set(cmd_buffer,
- template->bind_point, _set);
- if (!push_set)
+ struct anv_descriptor_set *set =
+ anv_cmd_buffer_push_descriptor_set(cmd_buffer, template->bind_point,
+ set_layout, _set);
+ if (!set)
return;
- struct anv_descriptor_set *set = &push_set->set;
-
- set->layout = set_layout;
- set->size = anv_descriptor_set_layout_size(set_layout);
- set->buffer_count = set_layout->buffer_count;
- set->buffer_views = push_set->buffer_views;
-
anv_descriptor_set_write_template(cmd_buffer->device, set,
&cmd_buffer->surface_state_stream,
template,