*/
#include "pan_context.h"
+#include "pan_cmdstream.h"
+#include "panfrost-quirks.h"
+#include "pan_bo.h"
#include "util/u_memory.h"
#include "nir_serialize.h"
so->variant_count = 1;
so->active_variant = 0;
- /* calloc, instead of malloc - to zero unused fields */
- v->tripipe = CALLOC_STRUCT(mali_shader_meta);
-
if (cso->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
struct blob_reader reader;
const struct pipe_binary_program_header *hdr = cso->prog;
so->cbase.ir_type = PIPE_SHADER_IR_NIR;
}
- panfrost_shader_compile(ctx, v->tripipe,
- so->cbase.ir_type, so->cbase.prog,
- MESA_SHADER_COMPUTE, v, NULL);
+ panfrost_shader_compile(ctx, so->cbase.ir_type, so->cbase.prog,
+ MESA_SHADER_COMPUTE, v, NULL);
return so;
}
const struct pipe_grid_info *info)
{
struct panfrost_context *ctx = pan_context(pipe);
+ struct panfrost_device *dev = pan_device(pipe->screen);
/* TODO: Do we want a special compute-only batch? */
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
ctx->compute_grid = info;
- struct mali_job_descriptor_header job = {
- .job_type = JOB_TYPE_COMPUTE,
- .job_descriptor_size = 1,
- .job_barrier = 1
- };
-
/* TODO: Stub */
- struct midgard_payload_vertex_tiler *payload = &ctx->payloads[PIPE_SHADER_COMPUTE];
+ struct midgard_payload_vertex_tiler payload = { 0 };
+ struct mali_invocation_packed invocation;
+ struct mali_draw_packed postfix;
/* We implement OpenCL inputs as uniforms (or a UBO -- same thing), so
* reuse the graphics path for this by lowering to Gallium */
if (info->input)
pipe->set_constant_buffer(pipe, PIPE_SHADER_COMPUTE, 0, &ubuf);
- panfrost_emit_for_draw(ctx, false);
-
- struct mali_shared_memory shared = {
- .shared_workgroup_count = ~0
- };
-
- payload->postfix.shared_memory =
- panfrost_upload_transient(batch, &shared, sizeof(shared));
-
- /* Invoke according to the grid info */
+ pan_pack(&postfix, DRAW, cfg) {
+ cfg.unknown_1 = (dev->quirks & IS_BIFROST) ? 0x2 : 0x6;
+ cfg.state = panfrost_emit_compute_shader_meta(batch, PIPE_SHADER_COMPUTE);
+ cfg.shared = panfrost_emit_shared_memory(batch, info);
+ cfg.uniform_buffers = panfrost_emit_const_buf(batch,
+ PIPE_SHADER_COMPUTE, &cfg.push_uniforms);
+ cfg.textures = panfrost_emit_texture_descriptors(batch,
+ PIPE_SHADER_COMPUTE);
+ cfg.samplers = panfrost_emit_sampler_descriptors(batch,
+ PIPE_SHADER_COMPUTE);
+ }
- panfrost_pack_work_groups_compute(&payload->prefix,
- info->grid[0], info->grid[1], info->grid[2],
- info->block[0], info->block[1], info->block[2], false);
+ unsigned magic =
+ util_logbase2_ceil(info->block[0] + 1) +
+ util_logbase2_ceil(info->block[1] + 1) +
+ util_logbase2_ceil(info->block[2] + 1);
- /* Upload the payload */
+ payload.prefix.primitive.opaque[0] = (magic) << 26; /* XXX */
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch, sizeof(job) + sizeof(*payload));
- memcpy(transfer.cpu, &job, sizeof(job));
- memcpy(transfer.cpu + sizeof(job), payload, sizeof(*payload));
+ memcpy(&payload.postfix, &postfix, sizeof(postfix));
- /* Queue the job */
- panfrost_scoreboard_queue_compute_job(batch, transfer);
+ /* Invoke according to the grid info */
- panfrost_flush_all_batches(ctx, true);
+ panfrost_pack_work_groups_compute(&invocation,
+ info->grid[0], info->grid[1],
+ info->grid[2],
+ info->block[0], info->block[1],
+ info->block[2],
+ false);
+ payload.prefix.invocation = invocation;
+
+ panfrost_new_job(&batch->pool, &batch->scoreboard,
+ MALI_JOB_TYPE_COMPUTE, true, 0, &payload,
+ sizeof(payload), false);
+ panfrost_flush_all_batches(ctx, 0);
}
static void