fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-22 15:40:11 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

anv: rework gfx state emission (again)

Browse source 
Previously we had 2 stages :

  runtime -> precomputed values -> packing/emission

With this change we use 3 stages :

   runtime -> precomputed values -> packing -> emission

Now blorp & other changes to the pipeline should not retrigger
repacking of instructions.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36711>
This commit is contained in:
Lionel Landwerlin 2025-03-31 13:09:39 +03:00 • committed by Marge Bot
parent 5a2fb0da32
commit 44aaf88425
11 changed files with 1162 additions and 850 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
src/intel/vulkan/anv_cmd_buffer.c
View file

@ -57,9 +57,8 @@ anv_cmd_state_init(struct anv_cmd_buffer *cmd_buffer)
state->compute.z_pass_async_compute_thread_limit = UINT8_MAX; state->compute.z_pass_async_compute_thread_limit = UINT8_MAX;
state->compute.np_z_async_throttle_settings = UINT8_MAX; state->compute.np_z_async_throttle_settings = UINT8_MAX;
memcpy(state->gfx.dyn_state.dirty, BITSET_COPY(state->gfx.dyn_state.pack_dirty,
cmd_buffer->device->gfx_dirty_state, cmd_buffer->device->gfx_dirty_state);
sizeof(state->gfx.dyn_state.dirty));
} }
static void static void
@ -516,25 +515,25 @@ anv_cmd_buffer_flush_pipeline_hw_state(struct anv_cmd_buffer *cmd_buffer,
assert(old_pipeline == NULL || \ assert(old_pipeline == NULL || \
old_pipeline->name.len == new_pipeline->name.len); \ old_pipeline->name.len == new_pipeline->name.len); \
/* Don't bother memcmp if the state is already dirty */ \ /* Don't bother memcmp if the state is already dirty */ \
if (!BITSET_TEST(hw_state->dirty, ANV_GFX_STATE_##bit) && \ if (!BITSET_TEST(hw_state->pack_dirty, ANV_GFX_STATE_##bit) && \
(old_pipeline == NULL || \ (old_pipeline == NULL || \
memcmp(&old_pipeline->batch_data[old_pipeline->name.offset], \ memcmp(&old_pipeline->batch_data[old_pipeline->name.offset], \
&new_pipeline->batch_data[new_pipeline->name.offset], \ &new_pipeline->batch_data[new_pipeline->name.offset], \
4 * new_pipeline->name.len) != 0)) \ 4 * new_pipeline->name.len) != 0)) \
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_##bit); \ BITSET_SET(hw_state->pack_dirty, ANV_GFX_STATE_##bit); \
} while (0) } while (0)
#define diff_var_state(bit, name) \ #define diff_var_state(bit, name) \
do { \ do { \
/* Don't bother memcmp if the state is already dirty */ \ /* Don't bother memcmp if the state is already dirty */ \
/* Also if the new state is empty, avoid marking dirty */ \ /* Also if the new state is empty, avoid marking dirty */ \
if (!BITSET_TEST(hw_state->dirty, ANV_GFX_STATE_##bit) && \ if (!BITSET_TEST(hw_state->pack_dirty, ANV_GFX_STATE_##bit) && \
new_pipeline->name.len != 0 && \ new_pipeline->name.len != 0 && \
(old_pipeline == NULL || \ (old_pipeline == NULL || \
old_pipeline->name.len != new_pipeline->name.len || \ old_pipeline->name.len != new_pipeline->name.len || \
memcmp(&old_pipeline->batch_data[old_pipeline->name.offset], \ memcmp(&old_pipeline->batch_data[old_pipeline->name.offset], \
&new_pipeline->batch_data[new_pipeline->name.offset], \ &new_pipeline->batch_data[new_pipeline->name.offset], \
4 * new_pipeline->name.len) != 0)) \ 4 * new_pipeline->name.len) != 0)) \
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_##bit); \ BITSET_SET(hw_state->pack_dirty, ANV_GFX_STATE_##bit); \
} while (0) } while (0)
#define assert_identical(bit, name) \ #define assert_identical(bit, name) \
do { \ do { \

4
src/intel/vulkan/anv_genX.h
View file

@ -131,7 +131,7 @@ genX(cmd_buffer_ensure_wa_14018283232)(struct anv_cmd_buffer *cmd_buffer,
if (intel_needs_workaround(cmd_buffer->device->info, 14018283232) && if (intel_needs_workaround(cmd_buffer->device->info, 14018283232) &&
hw_state->wa_14018283232_toggle != toggle) { hw_state->wa_14018283232_toggle != toggle) {
hw_state->wa_14018283232_toggle = toggle; hw_state->wa_14018283232_toggle = toggle;
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WA_14018283232); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WA_14018283232);
genX(batch_emit_wa_14018283232)(&cmd_buffer->batch); genX(batch_emit_wa_14018283232)(&cmd_buffer->batch);
} }
} }
@ -164,7 +164,7 @@ genX(cmd_buffer_set_coarse_pixel_active)(struct anv_cmd_buffer *cmd_buffer,
if (intel_needs_workaround(cmd_buffer->device->info, 18038825448) && if (intel_needs_workaround(cmd_buffer->device->info, 18038825448) &&
gfx->dyn_state.coarse_state != state) { gfx->dyn_state.coarse_state != state) {
gfx->dyn_state.coarse_state = state; gfx->dyn_state.coarse_state = state;
BITSET_SET(gfx->dyn_state.dirty, ANV_GFX_STATE_COARSE_STATE); BITSET_SET(gfx->dyn_state.emit_dirty, ANV_GFX_STATE_PS_EXTRA);
return true; return true;
} }
return false; return false;

30
src/intel/vulkan/anv_measure.c
View file

@ -154,16 +154,13 @@ anv_measure_start_snapshot(struct anv_cmd_buffer *cmd_buffer,
snapshot->renderpass = (type == INTEL_SNAPSHOT_COMPUTE) ? 0 snapshot->renderpass = (type == INTEL_SNAPSHOT_COMPUTE) ? 0
: measure->base.renderpass; : measure->base.renderpass;
if (type == INTEL_SNAPSHOT_COMPUTE && cmd_buffer->state.compute.base.pipeline) { if (type == INTEL_SNAPSHOT_COMPUTE && cmd_buffer->state.compute.shader) {
const struct anv_compute_pipeline *pipeline = snapshot->cs = cmd_buffer->state.compute.shader->prog_data->source_hash;
anv_pipeline_to_compute(cmd_buffer->state.compute.base.pipeline); } else if (type == INTEL_SNAPSHOT_DRAW) {
snapshot->cs = pipeline->cs->prog_data->source_hash; const struct anv_cmd_graphics_state *gfx = &cmd_buffer->state.gfx;
} else if (type == INTEL_SNAPSHOT_DRAW && cmd_buffer->state.gfx.base.pipeline) { #define SHADER_SOURCE_HASH(_name) \
const struct anv_graphics_pipeline *pipeline = (gfx->shaders[MESA_SHADER_##_name] ? \
anv_pipeline_to_graphics(cmd_buffer->state.gfx.base.pipeline); gfx->shaders[MESA_SHADER_##_name]->prog_data->source_hash : 0)
#define SHADER_SOURCE_HASH(_name) \
(pipeline->base.shaders[MESA_SHADER_##VERTEX] ? \
pipeline->base.shaders[MESA_SHADER_##VERTEX]->prog_data->source_hash : 0)
snapshot->vs = SHADER_SOURCE_HASH(VERTEX); snapshot->vs = SHADER_SOURCE_HASH(VERTEX);
snapshot->tcs = SHADER_SOURCE_HASH(TESS_CTRL); snapshot->tcs = SHADER_SOURCE_HASH(TESS_CTRL);
snapshot->tes = SHADER_SOURCE_HASH(TESS_EVAL); snapshot->tes = SHADER_SOURCE_HASH(TESS_EVAL);
@ -220,17 +217,12 @@ state_changed(struct anv_cmd_buffer *cmd_buffer,
return false; return false;
if (type == INTEL_SNAPSHOT_COMPUTE) { if (type == INTEL_SNAPSHOT_COMPUTE) {
const struct anv_compute_pipeline *cs_pipe = cs = cmd_buffer->state.compute.shader->prog_data->source_hash;
anv_pipeline_to_compute(cmd_buffer->state.compute.base.pipeline);
assert(cs_pipe);
cs = cs_pipe->cs->prog_data->source_hash;
} else if (type == INTEL_SNAPSHOT_DRAW) { } else if (type == INTEL_SNAPSHOT_DRAW) {
const struct anv_graphics_pipeline *gfx = const struct anv_cmd_graphics_state *gfx = &cmd_buffer->state.gfx;
anv_pipeline_to_graphics(cmd_buffer->state.gfx.base.pipeline);
assert(gfx);
#define SHADER_SOURCE_HASH(_name) \ #define SHADER_SOURCE_HASH(_name) \
(gfx->base.shaders[MESA_SHADER_##VERTEX] ? \ (gfx->shaders[MESA_SHADER_##_name] ? \
gfx->base.shaders[MESA_SHADER_##VERTEX]->prog_data->source_hash : 0) gfx->shaders[MESA_SHADER_##_name]->prog_data->source_hash : 0)
vs = SHADER_SOURCE_HASH(VERTEX); vs = SHADER_SOURCE_HASH(VERTEX);
tcs = SHADER_SOURCE_HASH(TESS_CTRL); tcs = SHADER_SOURCE_HASH(TESS_CTRL);
tes = SHADER_SOURCE_HASH(TESS_EVAL); tes = SHADER_SOURCE_HASH(TESS_EVAL);

87
src/intel/vulkan/anv_private.h
View file

@ -1534,11 +1534,9 @@ enum anv_gfx_state_bits {
ANV_GFX_STATE_TASK_SHADER, ANV_GFX_STATE_TASK_SHADER,
ANV_GFX_STATE_TASK_REDISTRIB, ANV_GFX_STATE_TASK_REDISTRIB,
/* Dynamic states */ /* Dynamic states */
ANV_GFX_STATE_BLEND_STATE, /* Just the dynamic state structure */ ANV_GFX_STATE_BLEND_STATE,
ANV_GFX_STATE_BLEND_STATE_PTR, /* The pointer to the dynamic state */
ANV_GFX_STATE_CLIP, ANV_GFX_STATE_CLIP,
ANV_GFX_STATE_CC_STATE, ANV_GFX_STATE_CC_STATE,
ANV_GFX_STATE_CC_STATE_PTR,
ANV_GFX_STATE_COARSE_PIXEL, ANV_GFX_STATE_COARSE_PIXEL,
ANV_GFX_STATE_CPS, ANV_GFX_STATE_CPS,
ANV_GFX_STATE_DEPTH_BOUNDS, ANV_GFX_STATE_DEPTH_BOUNDS,
@ -1558,7 +1556,6 @@ enum anv_gfx_state_bits {
ANV_GFX_STATE_VF_TOPOLOGY, ANV_GFX_STATE_VF_TOPOLOGY,
ANV_GFX_STATE_VFG, ANV_GFX_STATE_VFG,
ANV_GFX_STATE_VIEWPORT_CC, ANV_GFX_STATE_VIEWPORT_CC,
ANV_GFX_STATE_VIEWPORT_CC_PTR,
ANV_GFX_STATE_VIEWPORT_SF_CLIP, ANV_GFX_STATE_VIEWPORT_SF_CLIP,
ANV_GFX_STATE_WM, ANV_GFX_STATE_WM,
ANV_GFX_STATE_WM_DEPTH_STENCIL, ANV_GFX_STATE_WM_DEPTH_STENCIL,
@ -1569,7 +1566,6 @@ enum anv_gfx_state_bits {
ANV_GFX_STATE_TBIMR_TILE_PASS_INFO, ANV_GFX_STATE_TBIMR_TILE_PASS_INFO,
ANV_GFX_STATE_FS_MSAA_FLAGS, ANV_GFX_STATE_FS_MSAA_FLAGS,
ANV_GFX_STATE_TCS_INPUT_VERTICES, ANV_GFX_STATE_TCS_INPUT_VERTICES,
ANV_GFX_STATE_COARSE_STATE,
ANV_GFX_STATE_MESH_PROVOKING_VERTEX, ANV_GFX_STATE_MESH_PROVOKING_VERTEX,
ANV_GFX_STATE_MAX, ANV_GFX_STATE_MAX,
@ -1995,7 +1991,64 @@ struct anv_gfx_dynamic_state {
*/ */
enum anv_coarse_pixel_state coarse_state; enum anv_coarse_pixel_state coarse_state;
BITSET_DECLARE(dirty, ANV_GFX_STATE_MAX); /** Dirty bits of what needs to be repacked */
BITSET_DECLARE(pack_dirty, ANV_GFX_STATE_MAX);
struct {
uint32_t vf[2];
uint32_t vft[2];
uint32_t vfs[1];
uint32_t vfg[4];
uint32_t vf_sgvs[2];
uint32_t vf_sgvs_2[3];
uint32_t vf_sgvs_instancing[6];
uint32_t vf_sgvs_instancing_len;
uint32_t vf_component_packing[5];
uint32_t ib[5];
uint32_t so[5];
uint32_t so_decl_list[3 + 2 * 128];
uint32_t so_decl_list_len;
uint32_t clip[4];
uint32_t clip_mesh[2];
uint32_t sf_clip[2];
uint32_t cc_viewport[2];
uint32_t scissor[2];
uint32_t mesh_control[3];
uint32_t task_control[3];
uint32_t mesh_shader[8];
uint32_t task_shader[7];
uint32_t mesh_distrib[2];
uint32_t task_redistrib[2];
uint32_t vs[9];
uint32_t hs[9];
uint32_t te[5];
uint32_t ds[11];
uint32_t gs[10];
uint32_t sf[4];
uint32_t ms[2];
uint32_t sm[2];
uint32_t sp[9];
uint32_t raster[5];
uint32_t cps[9];
uint32_t ls[3];
uint32_t db[4];
uint32_t wm_ds[4];
uint32_t wm[2];
uint32_t pr[6];
uint32_t sbe[6];
uint32_t sbe_swiz[11];
uint32_t sbe_mesh[2];
uint32_t ps[12];
uint32_t ps_extra[2];
uint32_t ps_extra_dep[2];
uint32_t ps_blend[2];
uint32_t blend_state[2];
uint32_t cc_state[2];
uint32_t tbimr[4];
} packed;
/** Dirty bits of what needs to be reemitted */
BITSET_DECLARE(emit_dirty, ANV_GFX_STATE_MAX);
}; };
enum anv_internal_kernel_name { enum anv_internal_kernel_name {
@ -2768,28 +2821,6 @@ _anv_combine_address(struct anv_batch *batch, void *location,
_dst = NULL; \ _dst = NULL; \
})) }))
#define anv_batch_emit_merge_protected(batch, cmd, pipeline, state, \
name, protected) \
for (struct cmd name = { 0 }, \
*_dst = anv_batch_emit_dwords(batch, __anv_cmd_length(cmd)); \
__builtin_expect(_dst != NULL, 1); \
({ struct anv_gfx_state_ptr *_cmd_state = protected ? \
&(pipeline)->state##_protected : \
&(pipeline)->state; \
uint32_t _partial[__anv_cmd_length(cmd)]; \
assert(_cmd_state->len == __anv_cmd_length(cmd)); \
__anv_cmd_pack(cmd)(batch, _partial, &name); \
for (uint32_t i = 0; i < __anv_cmd_length(cmd); i++) { \
assert((_partial[i] & \
(pipeline)->batch_data[ \
(pipeline)->state.offset + i]) == 0); \
((uint32_t *)_dst)[i] = _partial[i] | \
(pipeline)->batch_data[_cmd_state->offset + i]; \
} \
VG(VALGRIND_CHECK_MEM_IS_DEFINED(_dst, __anv_cmd_length(cmd) * 4)); \
_dst = NULL; \
}))
#define anv_batch_emit(batch, cmd, name) \ #define anv_batch_emit(batch, cmd, name) \
for (struct cmd name = { __anv_cmd_header(cmd) }, \ for (struct cmd name = { __anv_cmd_header(cmd) }, \
*_dst = anv_batch_emit_dwords(batch, __anv_cmd_length(cmd)); \ *_dst = anv_batch_emit_dwords(batch, __anv_cmd_length(cmd)); \

4
src/intel/vulkan/anv_util.c
View file

@ -159,10 +159,8 @@ anv_gfx_state_bit_to_str(enum anv_gfx_state_bits state)
NAME(TASK_CONTROL); NAME(TASK_CONTROL);
NAME(TASK_SHADER); NAME(TASK_SHADER);
NAME(TASK_REDISTRIB); NAME(TASK_REDISTRIB);
NAME(BLEND_STATE_PTR);
NAME(CLIP); NAME(CLIP);
NAME(CC_STATE); NAME(CC_STATE);
NAME(CC_STATE_PTR);
NAME(CPS); NAME(CPS);
NAME(DEPTH_BOUNDS); NAME(DEPTH_BOUNDS);
NAME(INDEX_BUFFER); NAME(INDEX_BUFFER);
@ -180,7 +178,6 @@ anv_gfx_state_bit_to_str(enum anv_gfx_state_bits state)
NAME(VF_TOPOLOGY); NAME(VF_TOPOLOGY);
NAME(VFG); NAME(VFG);
NAME(VIEWPORT_CC); NAME(VIEWPORT_CC);
NAME(VIEWPORT_CC_PTR);
NAME(VIEWPORT_SF_CLIP); NAME(VIEWPORT_SF_CLIP);
NAME(WM); NAME(WM);
NAME(WM_DEPTH_STENCIL); NAME(WM_DEPTH_STENCIL);
@ -190,7 +187,6 @@ anv_gfx_state_bit_to_str(enum anv_gfx_state_bits state)
NAME(TBIMR_TILE_PASS_INFO); NAME(TBIMR_TILE_PASS_INFO);
NAME(FS_MSAA_FLAGS); NAME(FS_MSAA_FLAGS);
NAME(TCS_INPUT_VERTICES); NAME(TCS_INPUT_VERTICES);
NAME(COARSE_STATE);
NAME(MESH_PROVOKING_VERTEX); NAME(MESH_PROVOKING_VERTEX);
default: UNREACHABLE("invalid state"); default: UNREACHABLE("invalid state");
} }

66
src/intel/vulkan/genX_blorp_exec.c
View file

@ -328,7 +328,7 @@ blorp_exec_on_render(struct blorp_batch *batch,
* will trigger a PIPE_CONTROL too. * will trigger a PIPE_CONTROL too.
*/ */
hw_state->ds_write_state = blorp_ds_state; hw_state->ds_write_state = blorp_ds_state;
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WA_18019816803); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WA_18019816803);
/* Add the stall that will flush prior to the blorp operation by /* Add the stall that will flush prior to the blorp operation by
* genX(cmd_buffer_apply_pipe_flushes) * genX(cmd_buffer_apply_pipe_flushes)
@ -388,45 +388,45 @@ blorp_exec_on_render(struct blorp_batch *batch,
#endif #endif
/* Flag all the instructions emitted by BLORP. */ /* Flag all the instructions emitted by BLORP. */
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_URB); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_URB);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_STATISTICS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_STATISTICS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_TOPOLOGY); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_TOPOLOGY);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VERTEX_INPUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VERTEX_INPUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS);
#if GFX_VER >= 11 #if GFX_VER >= 11
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS_2); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS_2);
#endif #endif
#if GFX_VER >= 12 #if GFX_VER >= 12
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION);
#endif #endif
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VIEWPORT_CC_PTR); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VIEWPORT_CC);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_STREAMOUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_STREAMOUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_RASTER); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_RASTER);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_CLIP); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_CLIP);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SAMPLE_MASK); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SAMPLE_MASK);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MULTISAMPLE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MULTISAMPLE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SBE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SBE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SBE_SWIZ); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SBE_SWIZ);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_DEPTH_BOUNDS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_DEPTH_BOUNDS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WM); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WM);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WM_DEPTH_STENCIL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WM_DEPTH_STENCIL);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_HS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_HS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_DS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_DS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_GS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_GS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS_EXTRA); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS_EXTRA);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_BLEND_STATE_PTR); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_BLEND_STATE);
if (batch->blorp->config.use_mesh_shading) { if (batch->blorp->config.use_mesh_shading) {
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MESH_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MESH_CONTROL);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TASK_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TASK_CONTROL);
} }
if (params->wm_prog_data) { if (params->wm_prog_data) {
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_CC_STATE_PTR); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_CC_STATE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS_BLEND); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS_BLEND);
} }
anv_cmd_dirty_mask_t dirty = ~(ANV_CMD_DIRTY_INDEX_BUFFER | anv_cmd_dirty_mask_t dirty = ~(ANV_CMD_DIRTY_INDEX_BUFFER |

6
src/intel/vulkan/genX_cmd_buffer.c
View file

@ -3688,9 +3688,9 @@ genX(CmdExecuteCommands)(
memset(&container->state.gfx.urb_cfg, 0, sizeof(struct intel_urb_config)); memset(&container->state.gfx.urb_cfg, 0, sizeof(struct intel_urb_config));
/* Reemit all GFX instructions in container */ /* Reemit all GFX instructions in container */
memcpy(container->state.gfx.dyn_state.dirty, BITSET_OR(container->state.gfx.dyn_state.emit_dirty,
device->gfx_dirty_state, container->state.gfx.dyn_state.emit_dirty,
sizeof(container->state.gfx.dyn_state.dirty)); device->gfx_dirty_state);
if (container->device->vk.enabled_extensions.KHR_fragment_shading_rate) { if (container->device->vk.enabled_extensions.KHR_fragment_shading_rate) {
/* Also recompute the CPS_STATE offset */ /* Also recompute the CPS_STATE offset */
struct vk_dynamic_graphics_state *dyn = struct vk_dynamic_graphics_state *dyn =

2
src/intel/vulkan/genX_cmd_draw.c
View file

@ -874,7 +874,7 @@ cmd_buffer_flush_gfx_state(struct anv_cmd_buffer *cmd_buffer)
genX(cmd_buffer_flush_gfx_runtime_state)(cmd_buffer); genX(cmd_buffer_flush_gfx_runtime_state)(cmd_buffer);
/* Flush the HW state into the commmand buffer */ /* Flush the HW state into the commmand buffer */
if (!BITSET_IS_EMPTY(cmd_buffer->state.gfx.dyn_state.dirty)) if (!BITSET_IS_EMPTY(cmd_buffer->state.gfx.dyn_state.emit_dirty))
genX(cmd_buffer_flush_gfx_hw_state)(cmd_buffer); genX(cmd_buffer_flush_gfx_hw_state)(cmd_buffer);
/* If the pipeline changed, we may need to re-allocate push constant space /* If the pipeline changed, we may need to re-allocate push constant space

1698
src/intel/vulkan/genX_gfx_state.c
View file

File diff suppressed because it is too large Load diff

44
src/intel/vulkan/genX_gpu_memcpy.c
View file

@ -333,33 +333,33 @@ genX(emit_so_memcpy_fini)(struct anv_memcpy_state *state)
genX(cmd_buffer_ensure_wa_14018283232)(state->cmd_buffer, false); genX(cmd_buffer_ensure_wa_14018283232)(state->cmd_buffer, false);
#endif #endif
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_URB); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_URB);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_STATISTICS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_STATISTICS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_TOPOLOGY); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_TOPOLOGY);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VERTEX_INPUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VERTEX_INPUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS);
#if GFX_VER >= 11 #if GFX_VER >= 11
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS_2); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS_2);
#endif #endif
#if GFX_VER >= 12 #if GFX_VER >= 12
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION);
#endif #endif
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SO_DECL_LIST); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SO_DECL_LIST);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_STREAMOUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_STREAMOUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SAMPLE_MASK); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SAMPLE_MASK);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MULTISAMPLE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MULTISAMPLE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SBE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SBE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_HS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_HS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_DS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_DS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_GS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_GS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS);
if (state->cmd_buffer->device->vk.enabled_extensions.EXT_mesh_shader) { if (state->cmd_buffer->device->vk.enabled_extensions.EXT_mesh_shader) {
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MESH_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MESH_CONTROL);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TASK_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TASK_CONTROL);
} }
state->cmd_buffer->state.gfx.dirty |= state->cmd_buffer->state.gfx.dirty |=

58
src/intel/vulkan/genX_simple_shader.c
View file

@ -346,40 +346,40 @@ genX(emit_simpler_shader_init_fragment)(struct anv_simple_shader *state)
struct anv_gfx_dynamic_state *hw_state = struct anv_gfx_dynamic_state *hw_state =
&state->cmd_buffer->state.gfx.dyn_state; &state->cmd_buffer->state.gfx.dyn_state;
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_URB); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_URB);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_STATISTICS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_STATISTICS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_TOPOLOGY); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_TOPOLOGY);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VERTEX_INPUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VERTEX_INPUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS);
#if GFX_VER >= 11 #if GFX_VER >= 11
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VF_SGVS_2); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VF_SGVS_2);
#endif #endif
#if GFX_VER >= 12 #if GFX_VER >= 12
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PRIMITIVE_REPLICATION);
#endif #endif
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_STREAMOUT); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_STREAMOUT);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VIEWPORT_CC); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VIEWPORT_CC);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_CLIP); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_CLIP);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_RASTER); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_RASTER);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SAMPLE_MASK); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SAMPLE_MASK);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MULTISAMPLE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MULTISAMPLE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_DEPTH_BOUNDS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_DEPTH_BOUNDS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WM); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WM);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_WM_DEPTH_STENCIL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_WM_DEPTH_STENCIL);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SF); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SF);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_SBE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_SBE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_VS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_VS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_HS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_HS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_DS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_DS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TE); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TE);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_GS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_GS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS_EXTRA); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS_EXTRA);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_PS_BLEND); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS_BLEND);
if (device->vk.enabled_extensions.EXT_mesh_shader) { if (device->vk.enabled_extensions.EXT_mesh_shader) {
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_MESH_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_MESH_CONTROL);
BITSET_SET(hw_state->dirty, ANV_GFX_STATE_TASK_CONTROL); BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TASK_CONTROL);
} }
/* Update urb config after simple shader. */ /* Update urb config after simple shader. */