fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-06 15:58:05 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

panfrost: Remove progress_increment from all CS builders

Browse source 
Progression logic is deprecated since v11 and we don't plan to use it.
Let's get ride of all increment logic on all instructions.

Signed-off-by: Mary Guillemard <mary.guillemard@collabora.com>
Acked-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34947>
This commit is contained in:
Mary Guillemard 2025-05-13 12:56:25 +02:00 • committed by Marge Bot
parent 5ba982f166
commit 53f780ec91
11 changed files with 164 additions and 182 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

52
src/gallium/drivers/panfrost/pan_csf.c
View file

@ -148,7 +148,7 @@ csf_oom_handler_init(struct panfrost_context *ctx)
/* Use different framebuffer descriptor depending on whether incremental
* rendering has already been triggered */
cs_load32_to(&b, counter, tiler_oom_ctx, FIELD_OFFSET(counter));
cs_wait_slot(&b, 0, false);
cs_wait_slot(&b, 0);
cs_if(&b, MALI_CS_CONDITION_GREATER, counter) {
cs_load64_to(&b, cs_sr_reg64(&b, FRAGMENT, FBD_POINTER), tiler_oom_ctx,
FBD_OFFSET(MIDDLE));
@ -164,12 +164,12 @@ csf_oom_handler_init(struct panfrost_context *ctx)
FIELD_OFFSET(bbox_max));
cs_move64_to(&b, cs_sr_reg64(&b, FRAGMENT, TEM_POINTER), 0);
cs_move32_to(&b, cs_sr_reg32(&b, FRAGMENT, TEM_ROW_STRIDE), 0);
cs_wait_slot(&b, 0, false);
cs_wait_slot(&b, 0);
/* Run the fragment job and wait */
cs_select_sb_entries_for_async_ops(&b, 3);
cs_run_fragment(&b, false, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
cs_wait_slot(&b, 3, false);
cs_run_fragment(&b, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
cs_wait_slot(&b, 3);
/* Increment counter */
cs_add32(&b, counter, counter, 1);
@ -177,9 +177,9 @@ csf_oom_handler_init(struct panfrost_context *ctx)
/* Load completed chunks */
cs_load64_to(&b, tiler_ctx, tiler_oom_ctx, FIELD_OFFSET(tiler_desc));
cs_wait_slot(&b, 0, false);
cs_wait_slot(&b, 0);
cs_load_to(&b, completed_chunks, tiler_ctx, BITFIELD_MASK(4), 10 * 4);
cs_wait_slot(&b, 0, false);
cs_wait_slot(&b, 0);
cs_finish_fragment(&b, false, completed_top, completed_bottom, cs_now());
@ -195,7 +195,7 @@ csf_oom_handler_init(struct panfrost_context *ctx)
MALI_CS_OTHER_FLUSH_MODE_INVALIDATE, flush_id,
cs_defer(0, 0));
cs_wait_slot(&b, 0, false);
cs_wait_slot(&b, 0);
cs_select_sb_entries_for_async_ops(&b, 2);
}
@ -347,7 +347,7 @@ csf_emit_batch_end(struct panfrost_batch *batch)
struct cs_builder *b = batch->csf.cs.builder;
/* Barrier to let everything finish */
cs_wait_slots(b, BITFIELD_MASK(8), false);
cs_wait_slots(b, BITFIELD_MASK(8));
if (dev->debug & PAN_DBG_SYNC) {
/* Get the CS state */
@ -367,7 +367,7 @@ csf_emit_batch_end(struct panfrost_batch *batch)
cs_flush_caches(b, MALI_CS_FLUSH_MODE_CLEAN, MALI_CS_FLUSH_MODE_CLEAN,
MALI_CS_OTHER_FLUSH_MODE_INVALIDATE, flush_id,
cs_defer(0, 0));
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
/* Finish the command stream */
if (!cs_is_valid(batch->csf.cs.builder))
@ -821,8 +821,8 @@ GENX(csf_emit_fragment_job)(struct panfrost_batch *batch,
if (batch->draw_count > 0) {
/* Finish tiling and wait for IDVS and tiling */
cs_finish_tiling(b, false);
cs_wait_slot(b, 2, false);
cs_finish_tiling(b);
cs_wait_slot(b, 2);
cs_vt_end(b, cs_now());
}
@ -841,7 +841,7 @@ GENX(csf_emit_fragment_job)(struct panfrost_batch *batch,
if (batch->draw_count > 0) {
struct cs_index counter = cs_reg32(b, 78);
cs_load32_to(b, counter, cs_reg64(b, TILER_OOM_CTX_REG), 0);
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
cs_if(b, MALI_CS_CONDITION_GREATER, counter) {
cs_move64_to(b, cs_sr_reg64(b, FRAGMENT, FBD_POINTER),
GET_FBD(oom_ctx, LAST).gpu);
@ -849,8 +849,8 @@ GENX(csf_emit_fragment_job)(struct panfrost_batch *batch,
}
/* Run the fragment job and wait */
cs_run_fragment(b, false, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
cs_wait_slot(b, 2, false);
cs_run_fragment(b, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
cs_wait_slot(b, 2);
/* Gather freed heap chunks and add them to the heap context free list
* so they can be re-used next time the tiler heap runs out of chunks.
@ -862,7 +862,7 @@ GENX(csf_emit_fragment_job)(struct panfrost_batch *batch,
cs_move64_to(b, cs_reg64(b, 90), batch->tiler_ctx.valhall.desc);
cs_load_to(b, cs_reg_tuple(b, 86, 4), cs_reg64(b, 90), BITFIELD_MASK(4),
40);
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
cs_finish_fragment(b, true, cs_reg64(b, 86), cs_reg64(b, 88), cs_now());
}
}
@ -950,7 +950,7 @@ GENX(csf_launch_grid)(struct panfrost_batch *batch,
cs_load_to(b, grid_xyz, address, BITFIELD_MASK(3), 0);
/* Wait for the load */
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
/* Copy to FAU */
for (unsigned i = 0; i < 3; ++i) {
@ -962,7 +962,7 @@ GENX(csf_launch_grid)(struct panfrost_batch *batch,
}
/* Wait for the stores */
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
/* Use run_compute with a set task axis instead of run_compute_indirect as
* run_compute_indirect has been found to cause intermittent hangs. This
@ -973,7 +973,7 @@ GENX(csf_launch_grid)(struct panfrost_batch *batch,
* this is somewhat offset by run_compute being a native instruction. */
unsigned task_axis = MALI_TASK_AXIS_X;
cs_run_compute(b, DIV_ROUND_UP(max_thread_cnt, threads_per_wg), task_axis,
false, cs_shader_res_sel(0, 0, 0, 0));
cs_shader_res_sel(0, 0, 0, 0));
} else {
/* Set size in workgroups per dimension immediately */
cs_move32_to(b, cs_sr_reg32(b, COMPUTE, JOB_SIZE_X), info->grid[0]);
@ -1008,7 +1008,7 @@ GENX(csf_launch_grid)(struct panfrost_batch *batch,
assert(task_axis <= MALI_TASK_AXIS_Z);
assert(task_increment > 0);
cs_run_compute(b, task_increment, task_axis, false,
cs_run_compute(b, task_increment, task_axis,
cs_shader_res_sel(0, 0, 0, 0));
}
}
@ -1050,10 +1050,10 @@ GENX(csf_launch_xfb)(struct panfrost_batch *batch,
csf_emit_shader_regs(batch, PIPE_SHADER_VERTEX,
batch->rsd[PIPE_SHADER_VERTEX]);
/* force a barrier to avoid read/write sync issues with buffers */
cs_wait_slot(b, 2, false);
cs_wait_slot(b, 2);
/* XXX: Choose correctly */
cs_run_compute(b, 1, MALI_TASK_AXIS_Z, false, cs_shader_res_sel(0, 0, 0, 0));
cs_run_compute(b, 1, MALI_TASK_AXIS_Z, cs_shader_res_sel(0, 0, 0, 0));
}
static void
@ -1364,10 +1364,10 @@ GENX(csf_launch_draw)(struct panfrost_batch *batch,
}
#if PAN_ARCH >= 12
cs_run_idvs2(b, flags_override, false, true, drawid,
cs_run_idvs2(b, flags_override, true, drawid,
MALI_IDVS_SHADING_MODE_EARLY);
#else
cs_run_idvs(b, flags_override, false, true, cs_shader_res_sel(0, 0, 1, 0),
cs_run_idvs(b, flags_override, true, cs_shader_res_sel(0, 0, 1, 0),
cs_shader_res_sel(2, 2, 2, 0), drawid);
#endif
}
@ -1409,12 +1409,12 @@ GENX(csf_launch_draw_indirect)(struct panfrost_batch *batch,
cs_move32_to(b, cs_sr_reg32(b, IDVS, INDEX_BUFFER_SIZE), 0);
}
cs_wait_slot(b, 0, false);
cs_wait_slot(b, 0);
#if PAN_ARCH >= 12
cs_run_idvs2(b, flags_override, false, true, drawid,
cs_run_idvs2(b, flags_override, true, drawid,
MALI_IDVS_SHADING_MODE_EARLY);
#else
cs_run_idvs(b, flags_override, false, true, cs_shader_res_sel(0, 0, 1, 0),
cs_run_idvs(b, flags_override, true, cs_shader_res_sel(0, 0, 1, 0),
cs_shader_res_sel(2, 2, 2, 0), drawid);
#endif

3
src/gallium/drivers/panfrost/pan_precomp.c
View file

@ -386,7 +386,6 @@ GENX(panfrost_launch_precomp)(struct panfrost_batch *batch,
assert(task_axis <= MALI_TASK_AXIS_Z);
assert(task_increment > 0);
cs_run_compute(b, task_increment, task_axis, false,
cs_shader_res_sel(0, 0, 0, 0));
cs_run_compute(b, task_increment, task_axis, cs_shader_res_sel(0, 0, 0, 0));
#endif
}

107
src/panfrost/genxml/cs_builder.h
View file

@ -598,7 +598,7 @@ cs_flush_block_instrs(struct cs_builder *b)
sizeof(uint64_t));
/* Drop the prev_load_ip_target value and replace it by the final
* IP. */
* IP. */
*instr &= ~BITFIELD64_MASK(32);
*instr |= ip;
@ -1110,13 +1110,12 @@ cs_move64_to(struct cs_builder *b, struct cs_index dest, uint64_t imm)
}
static inline void
cs_wait_slots(struct cs_builder *b, unsigned wait_mask, bool progress_inc)
cs_wait_slots(struct cs_builder *b, unsigned wait_mask)
{
struct cs_load_store_tracker *ls_tracker = b->conf.ls_tracker;
cs_emit(b, WAIT, I) {
I.wait_mask = wait_mask;
I.progress_increment = progress_inc;
}
/* We don't do advanced tracking of cs_defer(), and assume that
@ -1130,11 +1129,11 @@ cs_wait_slots(struct cs_builder *b, unsigned wait_mask, bool progress_inc)
}
static inline void
cs_wait_slot(struct cs_builder *b, unsigned slot, bool progress_inc)
cs_wait_slot(struct cs_builder *b, unsigned slot)
{
assert(slot < 8 && "invalid slot");
cs_wait_slots(b, BITFIELD_BIT(slot), progress_inc);
cs_wait_slots(b, BITFIELD_BIT(slot));
}
struct cs_shader_res_sel {
@ -1154,13 +1153,11 @@ cs_shader_res_sel(unsigned srt, unsigned fau, unsigned spd, unsigned tsd)
static inline void
cs_run_compute(struct cs_builder *b, unsigned task_increment,
enum mali_task_axis task_axis, bool progress_inc,
struct cs_shader_res_sel res_sel)
enum mali_task_axis task_axis, struct cs_shader_res_sel res_sel)
{
cs_emit(b, RUN_COMPUTE, I) {
I.task_increment = task_increment;
I.task_axis = task_axis;
I.progress_increment = progress_inc;
I.srt_select = res_sel.srt;
I.spd_select = res_sel.spd;
I.tsd_select = res_sel.tsd;
@ -1170,12 +1167,11 @@ cs_run_compute(struct cs_builder *b, unsigned task_increment,
#if PAN_ARCH == 10
static inline void
cs_run_tiling(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
cs_run_tiling(struct cs_builder *b, uint32_t flags_override,
struct cs_shader_res_sel res_sel)
{
cs_emit(b, RUN_TILING, I) {
I.flags_override = flags_override;
I.progress_increment = progress_inc;
I.srt_select = res_sel.srt;
I.spd_select = res_sel.spd;
I.tsd_select = res_sel.tsd;
@ -1186,13 +1182,12 @@ cs_run_tiling(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
#if PAN_ARCH >= 12
static inline void
cs_run_idvs2(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
bool malloc_enable, struct cs_index draw_id,
cs_run_idvs2(struct cs_builder *b, uint32_t flags_override, bool malloc_enable,
struct cs_index draw_id,
enum mali_idvs_shading_mode vertex_shading_mode)
{
cs_emit(b, RUN_IDVS2, I) {
I.flags_override = flags_override;
I.progress_increment = progress_inc;
I.malloc_enable = malloc_enable;
I.vertex_shading_mode = vertex_shading_mode;
@ -1206,13 +1201,12 @@ cs_run_idvs2(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
}
#else
static inline void
cs_run_idvs(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
bool malloc_enable, struct cs_shader_res_sel varying_sel,
cs_run_idvs(struct cs_builder *b, uint32_t flags_override, bool malloc_enable,
struct cs_shader_res_sel varying_sel,
struct cs_shader_res_sel frag_sel, struct cs_index draw_id)
{
cs_emit(b, RUN_IDVS, I) {
I.flags_override = flags_override;
I.progress_increment = progress_inc;
I.malloc_enable = malloc_enable;
if (draw_id.type == CS_INDEX_UNDEF) {
@ -1242,31 +1236,29 @@ cs_run_idvs(struct cs_builder *b, uint32_t flags_override, bool progress_inc,
static inline void
cs_run_fragment(struct cs_builder *b, bool enable_tem,
enum mali_tile_render_order tile_order, bool progress_inc)
enum mali_tile_render_order tile_order)
{
cs_emit(b, RUN_FRAGMENT, I) {
I.enable_tem = enable_tem;
I.tile_order = tile_order;
I.progress_increment = progress_inc;
}
}
static inline void
cs_run_fullscreen(struct cs_builder *b, uint32_t flags_override,
bool progress_inc, struct cs_index dcd)
struct cs_index dcd)
{
cs_emit(b, RUN_FULLSCREEN, I) {
I.flags_override = flags_override;
I.progress_increment = progress_inc;
I.dcd = cs_src64(b, dcd);
}
}
static inline void
cs_finish_tiling(struct cs_builder *b, bool progress_inc)
cs_finish_tiling(struct cs_builder *b)
{
cs_emit(b, FINISH_TILING, I)
I.progress_increment = progress_inc;
;
}
static inline void
@ -1596,11 +1588,10 @@ cs_progress_load(struct cs_builder *b, struct cs_index dst)
static inline void
cs_run_compute_indirect(struct cs_builder *b, unsigned wg_per_task,
bool progress_inc, struct cs_shader_res_sel res_sel)
struct cs_shader_res_sel res_sel)
{
cs_emit(b, RUN_COMPUTE_INDIRECT, I) {
I.workgroups_per_task = wg_per_task;
I.progress_increment = progress_inc;
I.srt_select = res_sel.srt;
I.spd_select = res_sel.spd;
I.tsd_select = res_sel.tsd;
@ -1924,7 +1915,7 @@ cs_exception_handler_end(struct cs_builder *b,
cs_load64_to(b, addr_reg, handler->ctx.ctx_reg,
handler->ctx.dump_addr_offset);
cs_wait_slot(b, handler->ctx.ls_sb_slot, false);
cs_wait_slot(b, handler->ctx.ls_sb_slot);
for (unsigned i = 0; i < num_ranges; ++i) {
unsigned reg_count = util_bitcount(masks[i]);
@ -1933,7 +1924,7 @@ cs_exception_handler_end(struct cs_builder *b,
offset += reg_count * 4;
}
cs_wait_slot(b, handler->ctx.ls_sb_slot, false);
cs_wait_slot(b, handler->ctx.ls_sb_slot);
}
/* Now that the preamble is emitted, we can flush the instructions we have in
@ -1946,7 +1937,7 @@ cs_exception_handler_end(struct cs_builder *b,
cs_load64_to(b, addr_reg, handler->ctx.ctx_reg,
handler->ctx.dump_addr_offset);
cs_wait_slot(b, handler->ctx.ls_sb_slot, false);
cs_wait_slot(b, handler->ctx.ls_sb_slot);
for (unsigned i = 0; i < num_ranges; ++i) {
unsigned reg_count = util_bitcount(masks[i]);
@ -1955,7 +1946,7 @@ cs_exception_handler_end(struct cs_builder *b,
offset += reg_count * 4;
}
cs_wait_slot(b, handler->ctx.ls_sb_slot, false);
cs_wait_slot(b, handler->ctx.ls_sb_slot);
}
/* Fill the rest of the buffer with NOPs. */
@ -1992,10 +1983,10 @@ cs_trace_preamble(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
* access. Use cs_trace_field_offset() to get an offset taking this
* pre-increment into account. */
cs_load64_to(b, tracebuf_addr, ctx->ctx_reg, ctx->tracebuf_addr_offset);
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
cs_add64(b, tracebuf_addr, tracebuf_addr, trace_size);
cs_store64(b, tracebuf_addr, ctx->ctx_reg, ctx->tracebuf_addr_offset);
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}
#define cs_trace_field_offset(__type, __field) \
@ -2010,10 +2001,10 @@ struct cs_run_fragment_trace {
static inline void
cs_trace_run_fragment(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
struct cs_index scratch_regs, bool enable_tem,
enum mali_tile_render_order tile_order, bool progress_inc)
enum mali_tile_render_order tile_order)
{
if (likely(!ctx->enabled)) {
cs_run_fragment(b, enable_tem, tile_order, progress_inc);
cs_run_fragment(b, enable_tem, tile_order);
return;
}
@ -2026,12 +2017,12 @@ cs_trace_run_fragment(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
/* cs_run_xx() must immediately follow cs_load_ip_to() otherwise the IP
* won't point to the right instruction. */
cs_load_ip_to(b, data);
cs_run_fragment(b, enable_tem, tile_order, progress_inc);
cs_run_fragment(b, enable_tem, tile_order);
cs_store64(b, data, tracebuf_addr, cs_trace_field_offset(run_fragment, ip));
cs_store(b, cs_reg_tuple(b, 40, 7), tracebuf_addr, BITFIELD_MASK(7),
cs_trace_field_offset(run_fragment, sr));
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}
#if PAN_ARCH >= 12
@ -2045,12 +2036,11 @@ struct cs_run_idvs2_trace {
static inline void
cs_trace_run_idvs2(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
struct cs_index scratch_regs, uint32_t flags_override,
bool progress_inc, bool malloc_enable,
struct cs_index draw_id,
bool malloc_enable, struct cs_index draw_id,
enum mali_idvs_shading_mode vertex_shading_mode)
{
if (likely(!ctx->enabled)) {
cs_run_idvs2(b, flags_override, progress_inc, malloc_enable, draw_id,
cs_run_idvs2(b, flags_override, malloc_enable, draw_id,
vertex_shading_mode);
return;
}
@ -2063,8 +2053,7 @@ cs_trace_run_idvs2(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
/* cs_run_xx() must immediately follow cs_load_ip_to() otherwise the IP
* won't point to the right instruction. */
cs_load_ip_to(b, data);
cs_run_idvs2(b, flags_override, progress_inc, malloc_enable, draw_id,
vertex_shading_mode);
cs_run_idvs2(b, flags_override, malloc_enable, draw_id, vertex_shading_mode);
cs_store64(b, data, tracebuf_addr, cs_trace_field_offset(run_idvs2, ip));
if (draw_id.type != CS_INDEX_UNDEF)
@ -2076,7 +2065,7 @@ cs_trace_run_idvs2(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
cs_trace_field_offset(run_idvs2, sr[i]));
cs_store(b, cs_reg_tuple(b, 64, 2), tracebuf_addr, BITFIELD_MASK(2),
cs_trace_field_offset(run_idvs2, sr[64]));
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}
#else
struct cs_run_idvs_trace {
@ -2089,27 +2078,25 @@ struct cs_run_idvs_trace {
static inline void
cs_trace_run_idvs(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
struct cs_index scratch_regs, uint32_t flags_override,
bool progress_inc, bool malloc_enable,
struct cs_shader_res_sel varying_sel,
bool malloc_enable, struct cs_shader_res_sel varying_sel,
struct cs_shader_res_sel frag_sel, struct cs_index draw_id)
{
if (likely(!ctx->enabled)) {
cs_run_idvs(b, flags_override, progress_inc, malloc_enable, varying_sel,
frag_sel, draw_id);
cs_run_idvs(b, flags_override, malloc_enable, varying_sel, frag_sel,
draw_id);
return;
}
struct cs_index tracebuf_addr = cs_reg64(b, scratch_regs.reg);
struct cs_index data = cs_reg64(b, scratch_regs.reg + 2);
cs_trace_preamble(b, ctx, scratch_regs,
sizeof(struct cs_run_idvs_trace));
cs_trace_preamble(b, ctx, scratch_regs, sizeof(struct cs_run_idvs_trace));
/* cs_run_xx() must immediately follow cs_load_ip_to() otherwise the IP
* won't point to the right instruction. */
cs_load_ip_to(b, data);
cs_run_idvs(b, flags_override, progress_inc, malloc_enable, varying_sel,
frag_sel, draw_id);
cs_run_idvs(b, flags_override, malloc_enable, varying_sel, frag_sel,
draw_id);
cs_store64(b, data, tracebuf_addr, cs_trace_field_offset(run_idvs, ip));
if (draw_id.type != CS_INDEX_UNDEF)
@ -2121,7 +2108,7 @@ cs_trace_run_idvs(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
cs_trace_field_offset(run_idvs, sr[i]));
cs_store(b, cs_reg_tuple(b, 48, 13), tracebuf_addr, BITFIELD_MASK(13),
cs_trace_field_offset(run_idvs, sr[48]));
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}
#endif
@ -2133,24 +2120,23 @@ struct cs_run_compute_trace {
static inline void
cs_trace_run_compute(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
struct cs_index scratch_regs, unsigned task_increment,
enum mali_task_axis task_axis, bool progress_inc,
enum mali_task_axis task_axis,
struct cs_shader_res_sel res_sel)
{
if (likely(!ctx->enabled)) {
cs_run_compute(b, task_increment, task_axis, progress_inc, res_sel);
cs_run_compute(b, task_increment, task_axis, res_sel);
return;
}
struct cs_index tracebuf_addr = cs_reg64(b, scratch_regs.reg);
struct cs_index data = cs_reg64(b, scratch_regs.reg + 2);
cs_trace_preamble(b, ctx, scratch_regs,
sizeof(struct cs_run_compute_trace));
cs_trace_preamble(b, ctx, scratch_regs, sizeof(struct cs_run_compute_trace));
/* cs_run_xx() must immediately follow cs_load_ip_to() otherwise the IP
* won't point to the right instruction. */
cs_load_ip_to(b, data);
cs_run_compute(b, task_increment, task_axis, progress_inc, res_sel);
cs_run_compute(b, task_increment, task_axis, res_sel);
cs_store64(b, data, tracebuf_addr, cs_trace_field_offset(run_compute, ip));
for (unsigned i = 0; i < 32; i += 16)
@ -2158,31 +2144,30 @@ cs_trace_run_compute(struct cs_builder *b, const struct cs_tracing_ctx *ctx,
cs_trace_field_offset(run_compute, sr[i]));
cs_store(b, cs_reg_tuple(b, 32, 8), tracebuf_addr, BITFIELD_MASK(8),
cs_trace_field_offset(run_compute, sr[32]));
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}
static inline void
cs_trace_run_compute_indirect(struct cs_builder *b,
const struct cs_tracing_ctx *ctx,
struct cs_index scratch_regs,
unsigned wg_per_task, bool progress_inc,
unsigned wg_per_task,
struct cs_shader_res_sel res_sel)
{
if (likely(!ctx->enabled)) {
cs_run_compute_indirect(b, wg_per_task, progress_inc, res_sel);
cs_run_compute_indirect(b, wg_per_task, res_sel);
return;
}
struct cs_index tracebuf_addr = cs_reg64(b, scratch_regs.reg);
struct cs_index data = cs_reg64(b, scratch_regs.reg + 2);
cs_trace_preamble(b, ctx, scratch_regs,
sizeof(struct cs_run_compute_trace));
cs_trace_preamble(b, ctx, scratch_regs, sizeof(struct cs_run_compute_trace));
/* cs_run_xx() must immediately follow cs_load_ip_to() otherwise the IP
* won't point to the right instruction. */
cs_load_ip_to(b, data);
cs_run_compute_indirect(b, wg_per_task, progress_inc, res_sel);
cs_run_compute_indirect(b, wg_per_task, res_sel);
cs_store64(b, data, tracebuf_addr, cs_trace_field_offset(run_compute, ip));
for (unsigned i = 0; i < 32; i += 16)
@ -2190,5 +2175,5 @@ cs_trace_run_compute_indirect(struct cs_builder *b,
cs_trace_field_offset(run_compute, sr[i]));
cs_store(b, cs_reg_tuple(b, 32, 8), tracebuf_addr, BITFIELD_MASK(8),
cs_trace_field_offset(run_compute, sr[32]));
cs_wait_slot(b, ctx->ls_sb_slot, false);
cs_wait_slot(b, ctx->ls_sb_slot);
}

22
src/panfrost/vulkan/csf/panvk_vX_cmd_buffer.c
View file

@ -128,11 +128,11 @@ finish_cs(struct panvk_cmd_buffer *cmdbuf, uint32_t subqueue)
struct cs_index flush_id = cs_scratch_reg32(b, 0);
cs_move32_to(b, flush_id, 0);
cs_wait_slots(b, SB_ALL_MASK, false);
cs_wait_slots(b, SB_ALL_MASK);
cs_flush_caches(b, MALI_CS_FLUSH_MODE_CLEAN, MALI_CS_FLUSH_MODE_CLEAN,
MALI_CS_OTHER_FLUSH_MODE_NONE, flush_id,
cs_defer(SB_IMM_MASK, SB_ID(IMM_FLUSH)));
cs_wait_slot(b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(b, SB_ID(IMM_FLUSH));
/* If we're in sync/trace more, we signal the debug object. */
if (instance->debug_flags & (PANVK_DEBUG_SYNC | PANVK_DEBUG_TRACE)) {
@ -144,12 +144,12 @@ finish_cs(struct panvk_cmd_buffer *cmdbuf, uint32_t subqueue)
cs_move32_to(b, one, 1);
cs_load64_to(b, debug_sync_addr, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, debug.syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, debug_sync_addr, debug_sync_addr,
sizeof(struct panvk_cs_sync32) * subqueue);
cs_load32_to(b, error, debug_sync_addr,
offsetof(struct panvk_cs_sync32, error));
cs_wait_slots(b, SB_ALL_MASK, false);
cs_wait_slots(b, SB_ALL_MASK);
if (cmdbuf->vk.level == VK_COMMAND_BUFFER_LEVEL_PRIMARY)
cs_sync32_add(b, true, MALI_CS_SYNC_SCOPE_CSG, one,
debug_sync_addr, cs_now());
@ -162,7 +162,7 @@ finish_cs(struct panvk_cmd_buffer *cmdbuf, uint32_t subqueue)
/* Overwrite the sync error with the first error we encountered. */
cs_store32(b, error, debug_sync_addr,
offsetof(struct panvk_cs_sync32, error));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
}
}
@ -566,7 +566,7 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
struct panvk_cs_state *cs_state = &cmdbuf->state.cs[i];
if (deps.src[i].wait_sb_mask)
cs_wait_slots(b, deps.src[i].wait_sb_mask, false);
cs_wait_slots(b, deps.src[i].wait_sb_mask);
struct panvk_cache_flush_info cache_flush = deps.src[i].cache_flush;
if (cache_flush.l2 != MALI_CS_FLUSH_MODE_NONE ||
@ -577,7 +577,7 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
cs_move32_to(b, flush_id, 0);
cs_flush_caches(b, cache_flush.l2, cache_flush.lsc, cache_flush.others,
flush_id, cs_defer(SB_IMM_MASK, SB_ID(IMM_FLUSH)));
cs_wait_slot(b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(b, SB_ID(IMM_FLUSH));
}
/* If no one waits on us, there's no point signaling the sync object. */
@ -589,7 +589,7 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
cs_load64_to(b, sync_addr, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, sync_addr, sync_addr, sizeof(struct panvk_cs_sync64) * i);
cs_move64_to(b, add_val, 1);
cs_sync64_add(b, false, MALI_CS_SYNC_SCOPE_CSG, add_val, sync_addr,
@ -607,7 +607,7 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
cs_load64_to(b, sync_addr, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, sync_addr, sync_addr, sizeof(struct panvk_cs_sync64) * j);
cs_add64(b, wait_val, cs_progress_seqno_reg(b, j),
@ -628,12 +628,12 @@ panvk_per_arch(cs_pick_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
cs_load32_to(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_match(b, iter_sb, cmp_scratch) {
#define CASE(x) \
cs_case(b, x) { \
cs_wait_slot(b, SB_ITER(x), false); \
cs_wait_slot(b, SB_ITER(x)); \
cs_select_sb_entries_for_async_ops(b, SB_ITER(x)); \
}

16
src/panfrost/vulkan/csf/panvk_vX_cmd_dispatch.c
View file

@ -227,10 +227,10 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
if (shader->info.tls_size) {
cs_move64_to(b, cs_scratch_reg64(b, 0), cmdbuf->state.tls.desc.gpu);
cs_load64_to(b, cs_scratch_reg64(b, 2), cs_scratch_reg64(b, 0), 8);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_move64_to(b, cs_scratch_reg64(b, 0), tsd);
cs_store64(b, cs_scratch_reg64(b, 2), cs_scratch_reg64(b, 0), 8);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
cs_update_compute_ctx(b) {
@ -279,7 +279,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
cs_scratch_reg64(b, 0), BITFIELD_MASK(3), 0);
cs_move64_to(b, cs_scratch_reg64(b, 0),
cmdbuf->state.compute.push_uniforms);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
if (shader_uses_sysval(shader, compute, num_work_groups.x)) {
cs_store32(b, cs_sr_reg32(b, COMPUTE, JOB_SIZE_X),
@ -302,7 +302,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
shader, sysval_offset(compute, num_work_groups.z)));
}
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
} else {
cs_move32_to(b, cs_sr_reg32(b, COMPUTE, JOB_SIZE_X),
info->direct.wg_count.x);
@ -326,7 +326,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
* this is somewhat offset by run_compute being a native instruction. */
unsigned task_axis = MALI_TASK_AXIS_X;
cs_trace_run_compute(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
wg_per_task, task_axis, false,
wg_per_task, task_axis,
cs_shader_res_sel(0, 0, 0, 0));
} else {
unsigned task_axis = MALI_TASK_AXIS_X;
@ -334,7 +334,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
panvk_per_arch(calculate_task_axis_and_increment)(
shader, phys_dev, &task_axis, &task_increment);
cs_trace_run_compute(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
task_increment, task_axis, false,
task_increment, task_axis,
cs_shader_res_sel(0, 0, 0, 0));
}
cs_req_res(b, 0);
@ -347,7 +347,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
cs_load_to(b, cs_scratch_reg_tuple(b, 0, 3), cs_subqueue_ctx_reg(b),
BITFIELD_MASK(3),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, sync_addr, sync_addr,
PANVK_SUBQUEUE_COMPUTE * sizeof(struct panvk_cs_sync64));
@ -371,7 +371,7 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
++cmdbuf->state.cs[PANVK_SUBQUEUE_COMPUTE].relative_sync_point;
clear_dirty_after_dispatch(cmdbuf);

93
src/panfrost/vulkan/csf/panvk_vX_cmd_draw.c
View file

@ -769,7 +769,7 @@ cs_render_desc_ringbuf_reserve(struct cs_builder *b, uint32_t size)
cs_load64_to(
b, ringbuf_sync, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, render.desc_ringbuf.syncobj));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Wait for the other end to release memory. */
cs_move32_to(b, sz_reg, size - 1);
@ -793,7 +793,7 @@ cs_render_desc_ringbuf_move_ptr(struct cs_builder *b, uint32_t size,
b, cs_scratch_reg_tuple(b, 2, 3), cs_subqueue_ctx_reg(b),
BITFIELD_MASK(3),
offsetof(struct panvk_cs_subqueue_context, render.desc_ringbuf.ptr));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Update the relative position and absolute address. */
cs_add32(b, ptr_lo, ptr_lo, size);
@ -813,7 +813,7 @@ cs_render_desc_ringbuf_move_ptr(struct cs_builder *b, uint32_t size,
b, cs_scratch_reg_tuple(b, 2, 3), cs_subqueue_ctx_reg(b),
BITFIELD_MASK(3),
offsetof(struct panvk_cs_subqueue_context, render.desc_ringbuf.ptr));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
static VkResult
@ -927,7 +927,7 @@ get_tiler_desc(struct panvk_cmd_buffer *cmdbuf)
cs_move64_to(b, cs_scratch_reg64(b, 12), 0);
cs_move64_to(b, cs_scratch_reg64(b, 14), 0);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Take care of the tiler desc with layer_offset=0 outside of the loop. */
cs_move32_to(b, cs_scratch_reg32(b, 4),
@ -942,7 +942,7 @@ get_tiler_desc(struct panvk_cmd_buffer *cmdbuf)
cs_store(b, cs_scratch_reg_tuple(b, 0, 16), tiler_ctx_addr,
BITFIELD_RANGE(0, 2) | BITFIELD_RANGE(10, 6), 96);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
uint32_t remaining_layers =
td_count > 1
@ -970,7 +970,7 @@ get_tiler_desc(struct panvk_cmd_buffer *cmdbuf)
BITFIELD_RANGE(0, 2) | BITFIELD_RANGE(10, 6), 64);
cs_store(b, cs_scratch_reg_tuple(b, 0, 16), tiler_ctx_addr,
BITFIELD_RANGE(0, 2) | BITFIELD_RANGE(10, 6), 96);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_update_vt_ctx(b)
cs_add64(b, tiler_ctx_addr, tiler_ctx_addr,
@ -1006,7 +1006,7 @@ get_tiler_desc(struct panvk_cmd_buffer *cmdbuf)
cs_store(b, cs_scratch_reg_tuple(b, 0, 16), tiler_ctx_addr,
BITFIELD_RANGE(0, 2) | BITFIELD_RANGE(10, 6), 96);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add32(b, cs_scratch_reg32(b, 4), cs_scratch_reg32(b, 4),
MAX_LAYERS_PER_TILER_DESC << 8);
@ -1230,7 +1230,7 @@ get_fb_descs(struct panvk_cmd_buffer *cmdbuf)
cs_load64_to(b, cur_tiler, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context,
render.desc_ringbuf.ptr));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, dst_fbd_ptr, cur_tiler,
pan_size(TILER_CONTEXT) * td_count);
}
@ -1258,10 +1258,10 @@ get_fb_descs(struct panvk_cmd_buffer *cmdbuf)
cs_load_to(b, cs_scratch_reg_tuple(b, 0, 16),
pass_src_fbd_ptr, BITFIELD_MASK(16), fbd_off);
}
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_store(b, cs_scratch_reg_tuple(b, 0, 16), pass_dst_fbd_ptr,
BITFIELD_MASK(16), fbd_off);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
cs_add64(b, pass_src_fbd_ptr, pass_src_fbd_ptr, fbd_ir_pass_offset);
cs_add64(b, pass_dst_fbd_ptr, pass_dst_fbd_ptr, fbd_ir_pass_offset);
@ -1608,15 +1608,15 @@ wrap_prev_oq(struct panvk_cmd_buffer *cmdbuf)
cs_load64_to(
b, prev_oq_node_reg, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, render.oq_chain));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_store64(b, prev_oq_node_reg, oq_node_reg,
offsetof(struct panvk_cs_occlusion_query, next));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
cs_store64(b, oq_node_reg, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, render.oq_chain));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
return VK_SUCCESS;
}
@ -2084,11 +2084,11 @@ panvk_cmd_draw(struct panvk_cmd_buffer *cmdbuf, struct panvk_draw_info *draw)
cs_while(b, MALI_CS_CONDITION_GREATER, counter_reg) {
#if PAN_ARCH >= 12
cs_trace_run_idvs2(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
flags_override.opaque[0], false, true, cs_undef(),
flags_override.opaque[0], true, cs_undef(),
MALI_IDVS_SHADING_MODE_EARLY);
#else
cs_trace_run_idvs(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
flags_override.opaque[0], false, true,
flags_override.opaque[0], true,
cs_shader_res_sel(0, 0, 1, 0),
cs_shader_res_sel(2, 2, 2, 0), cs_undef());
#endif
@ -2107,11 +2107,11 @@ panvk_cmd_draw(struct panvk_cmd_buffer *cmdbuf, struct panvk_draw_info *draw)
} else {
#if PAN_ARCH >= 12
cs_trace_run_idvs2(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
flags_override.opaque[0], false, true, cs_undef(),
MALI_IDVS_SHADING_MODE_EARLY);
flags_override.opaque[0], true, cs_undef(),
MALI_IDVS_SHADING_MODE_EARLY);
#else
cs_trace_run_idvs(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
flags_override.opaque[0], false, true,
flags_override.opaque[0], true,
cs_shader_res_sel(0, 0, 1, 0),
cs_shader_res_sel(2, 2, 2, 0), cs_undef());
#endif
@ -2264,7 +2264,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
cs_load32_to(b, draw_count, draw_params_addr, 0);
/* wait for draw_count to load from buffer */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_umin32(b, draw_count, draw_count, max_draw_count);
} else {
cs_move32_to(b, draw_count, draw->indirect.draw_count);
@ -2288,7 +2288,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
}
/* Wait for the SR33-37 indirect buffer load. */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
if (patch_faus) {
if (shader_uses_sysval(vs, graphics, vs.first_vertex)) {
@ -2305,7 +2305,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
/* Wait for the store using SR-37 as src to finish, so we can
* overwrite it. */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
if (patch_attribs != 0) {
@ -2322,7 +2322,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
cs_load32_to(b, attrib_offset, vs_drv_set,
pan_size(ATTRIBUTE) * i + (2 * sizeof(uint32_t)));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Emulated immediate multiply: we walk the bits in
* base_instance, and accumulate (stride << bit_pos) if the bit
@ -2353,7 +2353,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
cs_store32(b, attrib_offset, vs_drv_set,
pan_size(ATTRIBUTE) * i + (2 * sizeof(uint32_t)));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
}
}
@ -2368,13 +2368,12 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
#if PAN_ARCH >= 12
cs_trace_run_idvs2(b, tracing_ctx, tracing_scratch_regs,
flags_override.opaque[0], false, true, draw_id,
MALI_IDVS_SHADING_MODE_EARLY);
flags_override.opaque[0], true, draw_id,
MALI_IDVS_SHADING_MODE_EARLY);
#else
cs_trace_run_idvs(b, tracing_ctx, tracing_scratch_regs,
flags_override.opaque[0], false, true,
cs_shader_res_sel(0, 0, 1, 0),
cs_shader_res_sel(2, 2, 2, 0), draw_id);
cs_trace_run_idvs(
b, tracing_ctx, tracing_scratch_regs, flags_override.opaque[0], true,
cs_shader_res_sel(0, 0, 1, 0), cs_shader_res_sel(2, 2, 2, 0), draw_id);
#endif
cs_add32(b, draw_count, draw_count, -1);
@ -2627,7 +2626,7 @@ flush_tiling(struct panvk_cmd_buffer *cmdbuf)
if (cmdbuf->state.gfx.render.tiler || inherits_render_ctx(cmdbuf)) {
/* Flush the tiling operations and signal the internal sync object. */
cs_req_res(b, CS_TILER_RES);
cs_finish_tiling(b, false);
cs_finish_tiling(b);
cs_req_res(b, 0);
struct cs_index sync_addr = cs_scratch_reg64(b, 0);
@ -2638,7 +2637,7 @@ flush_tiling(struct panvk_cmd_buffer *cmdbuf)
cs_load_to(b, cs_scratch_reg_tuple(b, 0, 3), cs_subqueue_ctx_reg(b),
BITFIELD_MASK(3),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* We're relying on PANVK_SUBQUEUE_VERTEX_TILER being the first queue to
* skip an ADD operation on the syncobjs pointer. */
@ -2669,14 +2668,14 @@ flush_tiling(struct panvk_cmd_buffer *cmdbuf)
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Update the vertex seqno. */
++cmdbuf->state.cs[PANVK_SUBQUEUE_VERTEX_TILER].relative_sync_point;
} else {
cs_load64_to(b, render_ctx, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, render));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
}
@ -2691,7 +2690,7 @@ wait_finish_tiling(struct panvk_cmd_buffer *cmdbuf)
cs_load64_to(b, vt_sync_addr, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, vt_sync_point,
cs_progress_seqno_reg(b, PANVK_SUBQUEUE_VERTEX_TILER),
@ -2750,7 +2749,7 @@ setup_tiler_oom_ctx(struct panvk_cmd_buffer *cmdbuf)
cs_store32(b, layer_count, cs_subqueue_ctx_reg(b),
TILER_OOM_CTX_FIELD_OFFSET(layer_count));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
static VkResult
@ -2837,7 +2836,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_load32_to(
b, counter, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, tiler_oom_ctx.counter));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_if(b, MALI_CS_CONDITION_GREATER, counter)
cs_update_frag_ctx(b)
cs_add64(b, cs_sr_reg64(b, FRAGMENT, FBD_POINTER),
@ -2854,7 +2853,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_flush_caches(b, MALI_CS_FLUSH_MODE_NONE, MALI_CS_FLUSH_MODE_NONE,
MALI_CS_OTHER_FLUSH_MODE_INVALIDATE, length_reg,
cs_defer(0x0, SB_ID(IMM_FLUSH)));
cs_wait_slot(b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(b, SB_ID(IMM_FLUSH));
}
cs_req_res(b, CS_FRAG_RES);
@ -2864,7 +2863,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_move32_to(b, layer_count, calc_enabled_layer_count(cmdbuf));
cs_while(b, MALI_CS_CONDITION_GREATER, layer_count) {
cs_trace_run_fragment(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
false, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
false, MALI_TILE_RENDER_ORDER_Z_ORDER);
cs_add32(b, layer_count, layer_count, -1);
cs_update_frag_ctx(b)
@ -2873,7 +2872,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
}
} else {
cs_trace_run_fragment(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
false, MALI_TILE_RENDER_ORDER_Z_ORDER, false);
false, MALI_TILE_RENDER_ORDER_Z_ORDER);
}
cs_req_res(b, 0);
@ -2907,7 +2906,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
render.desc_ringbuf.syncobj));
}
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, sync_addr, sync_addr,
PANVK_SUBQUEUE_FRAGMENT * sizeof(struct panvk_cs_sync64));
@ -2920,12 +2919,12 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_defer(SB_WAIT_ITER(x), SB_ID(DEFERRED_SYNC)); \
if (td_count == 1) { \
cs_load_to(b, completed, cur_tiler, BITFIELD_MASK(4), 40); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
cs_finish_fragment(b, true, completed_top, completed_bottom, async); \
} else if (td_count > 1) { \
cs_while(b, MALI_CS_CONDITION_GREATER, tiler_count) { \
cs_load_to(b, completed, cur_tiler, BITFIELD_MASK(4), 40); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
cs_finish_fragment(b, false, completed_top, completed_bottom, \
async); \
cs_update_frag_ctx(b) \
@ -2948,20 +2947,20 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_load64_to( \
b, oq_chain, cs_subqueue_ctx_reg(b), \
offsetof(struct panvk_cs_subqueue_context, render.oq_chain)); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
/* We use oq_syncobj as a placeholder to reset the oq_chain. */ \
cs_move64_to(b, oq_syncobj, 0); \
cs_store64( \
b, oq_syncobj, cs_subqueue_ctx_reg(b), \
offsetof(struct panvk_cs_subqueue_context, render.oq_chain)); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
cs_while(b, MALI_CS_CONDITION_ALWAYS, cs_undef()) { \
cs_load64_to(b, oq_syncobj, oq_chain, \
offsetof(struct panvk_cs_occlusion_query, syncobj)); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
cs_load64_to(b, oq_chain, oq_chain, \
offsetof(struct panvk_cs_occlusion_query, next)); \
cs_wait_slot(b, SB_ID(LS), false); \
cs_wait_slot(b, SB_ID(LS)); \
cs_sync32_set( \
b, true, MALI_CS_SYNC_SCOPE_CSG, add_val_lo, oq_syncobj, \
cs_defer(SB_MASK(DEFERRED_FLUSH), SB_ID(DEFERRED_SYNC))); \
@ -2987,7 +2986,7 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* Update the ring buffer position. */
if (free_render_descs) {

4
src/panfrost/vulkan/csf/panvk_vX_cmd_event.c
View file

@ -41,7 +41,7 @@ panvk_per_arch(CmdResetEvent2)(VkCommandBuffer commandBuffer, VkEvent _event,
(i * sizeof(struct panvk_cs_sync32)));
cs_load32_to(b, seqno, sync_addr,
offsetof(struct panvk_cs_sync32, seqno));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_match(b, seqno, cmp_scratch) {
cs_case(b, 0) {
@ -83,7 +83,7 @@ panvk_per_arch(CmdSetEvent2)(VkCommandBuffer commandBuffer, VkEvent _event,
(i * sizeof(struct panvk_cs_sync32)));
cs_load32_to(b, seqno, sync_addr,
offsetof(struct panvk_cs_sync32, seqno));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_match(b, seqno, cmp_scratch) {
cs_case(b, 0) {

10
src/panfrost/vulkan/csf/panvk_vX_cmd_precomp.c
View file

@ -95,10 +95,10 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
if (shader->info.tls_size) {
cs_move64_to(b, cs_scratch_reg64(b, 0), cmdbuf->state.tls.desc.gpu);
cs_load64_to(b, cs_scratch_reg64(b, 2), cs_scratch_reg64(b, 0), 8);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_move64_to(b, cs_scratch_reg64(b, 0), tsd);
cs_store64(b, cs_scratch_reg64(b, 2), cs_scratch_reg64(b, 0), 8);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
cs_update_compute_ctx(b) {
@ -146,7 +146,7 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
panvk_per_arch(calculate_task_axis_and_increment)(
shader, phys_dev, &task_axis, &task_increment);
cs_trace_run_compute(b, tracing_ctx, cs_scratch_reg_tuple(b, 0, 4),
task_increment, task_axis, false,
task_increment, task_axis,
cs_shader_res_sel(0, 0, 0, 0));
cs_req_res(b, 0);
@ -158,7 +158,7 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
cs_load_to(b, cs_scratch_reg_tuple(b, 0, 3), cs_subqueue_ctx_reg(b),
BITFIELD_MASK(3),
offsetof(struct panvk_cs_subqueue_context, syncobjs));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_add64(b, sync_addr, sync_addr,
PANVK_SUBQUEUE_COMPUTE * sizeof(struct panvk_cs_sync64));
@ -182,7 +182,7 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
++cmdbuf->state.cs[PANVK_SUBQUEUE_COMPUTE].relative_sync_point;

14
src/panfrost/vulkan/csf/panvk_vX_cmd_query.c
View file

@ -93,7 +93,7 @@ panvk_cmd_reset_occlusion_queries(struct panvk_cmd_buffer *cmd,
/* Wait on deferred sync to ensure all prior query operations have
* completed
*/
cs_wait_slot(b, SB_ID(DEFERRED_SYNC), false);
cs_wait_slot(b, SB_ID(DEFERRED_SYNC));
struct cs_index addr = cs_scratch_reg64(b, 16);
struct cs_index zero_regs = cs_scratch_reg_tuple(b, 0, 16);
@ -113,7 +113,7 @@ panvk_cmd_reset_occlusion_queries(struct panvk_cmd_buffer *cmd,
/* reset_oq_batch() only does the stores, we need to flush those explicitly
* here. */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
/* We flush the caches to make the new value visible to the CPU. */
struct cs_index flush_id = cs_scratch_reg32(b, 0);
@ -121,7 +121,7 @@ panvk_cmd_reset_occlusion_queries(struct panvk_cmd_buffer *cmd,
cs_flush_caches(b, MALI_CS_FLUSH_MODE_CLEAN, MALI_CS_FLUSH_MODE_CLEAN,
MALI_CS_OTHER_FLUSH_MODE_NONE, flush_id,
cs_defer(SB_IMM_MASK, SB_ID(IMM_FLUSH)));
cs_wait_slot(b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(b, SB_ID(IMM_FLUSH));
}
static void
@ -152,7 +152,7 @@ panvk_cmd_begin_occlusion_query(struct panvk_cmd_buffer *cmd,
cs_move64_to(b, report_addr_gpu, report_addr);
cs_move64_to(b, clear_value, 0);
cs_store64(b, clear_value, report_addr_gpu, 0);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
static void
@ -219,7 +219,7 @@ copy_oq_result_batch(struct cs_builder *b,
}
/* Flush the loads. */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
for (uint32_t i = 0; i < query_count; i++) {
struct cs_index store_src =
@ -230,7 +230,7 @@ copy_oq_result_batch(struct cs_builder *b,
}
/* Flush the stores. */
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
static void
@ -245,7 +245,7 @@ panvk_copy_occlusion_query_results(struct panvk_cmd_buffer *cmd,
/* Wait for occlusion query syncobjs to be signalled. */
if (flags & VK_QUERY_RESULT_WAIT_BIT)
cs_wait_slot(b, SB_ID(DEFERRED_SYNC), false);
cs_wait_slot(b, SB_ID(DEFERRED_SYNC));
uint32_t res_size = (flags & VK_QUERY_RESULT_64_BIT) ? 2 : 1;
uint32_t regs_per_copy =

17
src/panfrost/vulkan/csf/panvk_vX_exception_handler.c
View file

@ -76,7 +76,7 @@ generate_tiler_oom_handler(struct panvk_device *dev,
* rendering has already been triggered */
cs_load32_to(&b, counter, subqueue_ctx,
TILER_OOM_CTX_FIELD_OFFSET(counter));
cs_wait_slot(&b, SB_ID(LS), false);
cs_wait_slot(&b, SB_ID(LS));
cs_if(&b, MALI_CS_CONDITION_GREATER, counter)
cs_load64_to(&b, fbd_ptr, subqueue_ctx,
@ -87,19 +87,18 @@ generate_tiler_oom_handler(struct panvk_device *dev,
cs_load32_to(&b, layer_count, subqueue_ctx,
TILER_OOM_CTX_FIELD_OFFSET(layer_count));
cs_wait_slot(&b, SB_ID(LS), false);
cs_wait_slot(&b, SB_ID(LS));
cs_req_res(&b, CS_FRAG_RES);
cs_while(&b, MALI_CS_CONDITION_GREATER, layer_count) {
cs_trace_run_fragment(&b, &tracing_ctx,
cs_scratch_reg_tuple(&b, 8, 4), false,
MALI_TILE_RENDER_ORDER_Z_ORDER, false);
cs_trace_run_fragment(&b, &tracing_ctx, cs_scratch_reg_tuple(&b, 8, 4),
false, MALI_TILE_RENDER_ORDER_Z_ORDER);
cs_add32(&b, layer_count, layer_count, -1);
cs_add64(&b, fbd_ptr, fbd_ptr, fbd_size);
}
cs_req_res(&b, 0);
/* Wait for all iter scoreboards for simplicity. */
cs_wait_slots(&b, SB_ALL_ITERS_MASK, false);
cs_wait_slots(&b, SB_ALL_ITERS_MASK);
/* Increment counter */
cs_add32(&b, counter, counter, 1);
@ -111,12 +110,12 @@ generate_tiler_oom_handler(struct panvk_device *dev,
cs_load32_to(&b, td_count, subqueue_ctx,
TILER_OOM_CTX_FIELD_OFFSET(td_count));
cs_move64_to(&b, zero, 0);
cs_wait_slot(&b, SB_ID(LS), false);
cs_wait_slot(&b, SB_ID(LS));
cs_while(&b, MALI_CS_CONDITION_GREATER, td_count) {
/* Load completed chunks */
cs_load_to(&b, completed_chunks, tiler_ptr, BITFIELD_MASK(4), 10 * 4);
cs_wait_slot(&b, SB_ID(LS), false);
cs_wait_slot(&b, SB_ID(LS));
cs_finish_fragment(&b, false, completed_top, completed_bottom,
cs_now());
@ -136,7 +135,7 @@ generate_tiler_oom_handler(struct panvk_device *dev,
MALI_CS_OTHER_FLUSH_MODE_INVALIDATE, flush_id,
cs_defer(SB_IMM_MASK, SB_ID(IMM_FLUSH)));
cs_wait_slot(&b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(&b, SB_ID(IMM_FLUSH));
}
assert(cs_is_valid(&b));

8
src/panfrost/vulkan/csf/panvk_vX_utrace.c
View file

@ -31,7 +31,7 @@ cmd_copy_data(struct cs_builder *b, uint64_t dst_addr, uint64_t src_addr,
assert((dst_addr | src_addr | size) % sizeof(uint32_t) == 0);
/* wait for timestamp writes */
cs_wait_slot(b, SB_ID(DEFERRED_SYNC), false);
cs_wait_slot(b, SB_ID(DEFERRED_SYNC));
/* Depending on where this is called from, we could potentially use SR
* registers or copy with a compute job.
@ -52,7 +52,7 @@ cmd_copy_data(struct cs_builder *b, uint64_t dst_addr, uint64_t src_addr,
const struct cs_index reg = cs_scratch_reg_tuple(b, 4, count);
cs_load_to(b, reg, src_addr_reg, BITFIELD_MASK(count), offset);
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
cs_store(b, reg, dst_addr_reg, BITFIELD_MASK(count), offset);
copy_count -= count;
@ -64,7 +64,7 @@ cmd_copy_data(struct cs_builder *b, uint64_t dst_addr, uint64_t src_addr,
size -= offset;
}
cs_wait_slot(b, SB_ID(LS), false);
cs_wait_slot(b, SB_ID(LS));
}
static struct cs_builder *
@ -170,7 +170,7 @@ panvk_per_arch(utrace_clone_finish_builder)(struct cs_builder *b)
cs_flush_caches(b, MALI_CS_FLUSH_MODE_CLEAN, MALI_CS_FLUSH_MODE_NONE,
MALI_CS_OTHER_FLUSH_MODE_NONE, flush_id,
cs_defer(SB_IMM_MASK, SB_ID(IMM_FLUSH)));
cs_wait_slot(b, SB_ID(IMM_FLUSH), false);
cs_wait_slot(b, SB_ID(IMM_FLUSH));
cs_finish(b);
}