fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-26 14:38:13 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

iris: Emit a EXECUTE_INDIRECT_DISPATCH when available

Browse source 
On newer platforms (Arrowlake and above) we can issue a
EXECUTE_INDIRECT_DISPATCH that allows us to:

  * Skip issuing mi load/store instructions for indirect parameters

Signed-off-by: Rohan Garg <rohan.garg@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26421>
This commit is contained in:
Rohan Garg 2022-06-27 14:49:30 +02:00 • committed by Marge Bot
parent 580728564e
commit d161e3c2e2
1 changed files with 81 additions and 31 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

80
src/gallium/drivers/iris/iris_state.c
View file

@ -8332,8 +8332,52 @@ iris_load_indirect_location(struct iris_context *ice,
mi_store(&b, mi_reg32(GPGPU_DISPATCHDIMZ), size_z); mi_store(&b, mi_reg32(GPGPU_DISPATCHDIMZ), size_z);
} }
static bool iris_emit_indirect_dispatch_supported(const struct intel_device_info *devinfo)
{
// TODO: Swizzling X and Y workgroup sizes is not supported in execute indirect dispatch
return devinfo->has_indirect_unroll;
}
#if GFX_VERx10 >= 125 #if GFX_VERx10 >= 125
static void iris_emit_execute_indirect_dispatch(struct iris_context *ice,
struct iris_batch *batch,
const struct pipe_grid_info *grid,
const struct GENX(INTERFACE_DESCRIPTOR_DATA) idd)
{
const struct iris_screen *screen = batch->screen;
const struct intel_device_info *devinfo = screen->devinfo;
struct iris_compiled_shader *shader =
ice->shaders.prog[MESA_SHADER_COMPUTE];
struct brw_stage_prog_data *prog_data = shader->prog_data;
struct brw_cs_prog_data *cs_prog_data = (void *) prog_data;
const struct brw_cs_dispatch_info dispatch =
brw_cs_get_dispatch_info(devinfo, cs_prog_data, grid->block);
struct iris_bo *indirect = iris_resource_bo(grid->indirect);
const int dispatch_size = dispatch.simd_size / 16;
struct GENX(COMPUTE_WALKER_BODY) body = {};
body.SIMDSize = dispatch_size;
body.MessageSIMD = dispatch_size;
body.LocalXMaximum = grid->block[0] - 1;
body.LocalYMaximum = grid->block[1] - 1;
body.LocalZMaximum = grid->block[2] - 1;
body.ExecutionMask = dispatch.right_mask;
body.PostSync.MOCS = iris_mocs(NULL, &screen->isl_dev, 0);
body.InterfaceDescriptor = idd;
struct iris_address indirect_bo = ro_bo(indirect, grid->indirect_offset);
iris_emit_cmd(batch, GENX(EXECUTE_INDIRECT_DISPATCH), ind) {
ind.PredicateEnable =
ice->state.predicate == IRIS_PREDICATE_STATE_USE_BIT;
ind.MaxCount = 1;
ind.COMPUTE_WALKER_BODY = body;
ind.ArgumentBufferStartAddress = indirect_bo;
ind.MOCS =
iris_mocs(indirect_bo.bo, &screen->isl_dev, 0);
}
}
static void static void
iris_upload_compute_walker(struct iris_context *ice, iris_upload_compute_walker(struct iris_context *ice,
struct iris_batch *batch, struct iris_batch *batch,
@ -8363,6 +8407,25 @@ iris_upload_compute_walker(struct iris_context *ice,
} }
} }
struct GENX(INTERFACE_DESCRIPTOR_DATA) idd = {};
idd.KernelStartPointer = KSP(shader);
idd.NumberofThreadsinGPGPUThreadGroup = dispatch.threads;
idd.SharedLocalMemorySize =
encode_slm_size(GFX_VER, prog_data->total_shared);
idd.SamplerStatePointer = shs->sampler_table.offset;
idd.SamplerCount = encode_sampler_count(shader),
idd.BindingTablePointer = binder->bt_offset[MESA_SHADER_COMPUTE];
/* Typically set to 0 to avoid prefetching on every thread dispatch. */
idd.BindingTableEntryCount = devinfo->verx10 == 125 ?
0 : MIN2(shader->bt.size_bytes / 4, 31);
idd.PreferredSLMAllocationSize = preferred_slm_allocation_size(devinfo);
idd.NumberOfBarriers = cs_prog_data->uses_barrier;
iris_measure_snapshot(ice, batch, INTEL_SNAPSHOT_COMPUTE, NULL, NULL, NULL);
if (iris_emit_indirect_dispatch_supported(devinfo) && grid->indirect) {
iris_emit_execute_indirect_dispatch(ice, batch, grid, idd);
} else {
if (grid->indirect) if (grid->indirect)
iris_load_indirect_location(ice, batch, grid); iris_load_indirect_location(ice, batch, grid);
@ -8382,24 +8445,11 @@ iris_upload_compute_walker(struct iris_context *ice,
cw.ThreadGroupIDZDimension = grid->grid[2]; cw.ThreadGroupIDZDimension = grid->grid[2];
cw.ExecutionMask = dispatch.right_mask; cw.ExecutionMask = dispatch.right_mask;
cw.PostSync.MOCS = iris_mocs(NULL, &screen->isl_dev, 0); cw.PostSync.MOCS = iris_mocs(NULL, &screen->isl_dev, 0);
cw.InterfaceDescriptor = idd;
cw.InterfaceDescriptor = (struct GENX(INTERFACE_DESCRIPTOR_DATA)) {
.KernelStartPointer = KSP(shader),
.NumberofThreadsinGPGPUThreadGroup = dispatch.threads,
.SharedLocalMemorySize =
encode_slm_size(GFX_VER, prog_data->total_shared),
.PreferredSLMAllocationSize = preferred_slm_allocation_size(devinfo),
.NumberOfBarriers = cs_prog_data->uses_barrier,
.SamplerStatePointer = shs->sampler_table.offset,
.SamplerCount = encode_sampler_count(shader),
.BindingTablePointer = binder->bt_offset[MESA_SHADER_COMPUTE],
/* Typically set to 0 to avoid prefetching on every thread dispatch. */
.BindingTableEntryCount = devinfo->verx10 == 125 ?
0 : MIN2(shader->bt.size_bytes / 4, 31),
};
assert(brw_cs_push_const_total_size(cs_prog_data, dispatch.threads) == 0); assert(brw_cs_push_const_total_size(cs_prog_data, dispatch.threads) == 0);
} }
}
trace_intel_end_compute(&batch->trace, grid->grid[0], grid->grid[1], grid->grid[2]); trace_intel_end_compute(&batch->trace, grid->grid[0], grid->grid[1], grid->grid[2]);
} }