panfrost: Emit texture/sampler points for compute
[mesa.git] / src / gallium / drivers / panfrost / pan_compute.c
index 5ef26f3e35d6ad588001bffead2aa4a282fdd27b..3628a38dbf388c59444f367a2952314f37340323 100644 (file)
@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2019 Collabora, Ltd.
+ * Copyright (C) 2019 Red Hat Inc.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
  */
 
 #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"
+
+/* Compute CSOs are tracked like graphics shader CSOs, but are
+ * considerably simpler. We do not implement multiple
+ * variants/keying. So the CSO create function just goes ahead and
+ * compiles the thing. */
 
 static void *
 panfrost_create_compute_state(
         struct pipe_context *pctx,
         const struct pipe_compute_state *cso)
 {
-        return mem_dup(cso, sizeof(*cso));
+        struct panfrost_context *ctx = pan_context(pctx);
+
+        struct panfrost_shader_variants *so = CALLOC_STRUCT(panfrost_shader_variants);
+        so->cbase = *cso;
+        so->is_compute = true;
+
+        struct panfrost_shader_state *v = calloc(1, sizeof(*v));
+        so->variants = v;
+
+        so->variant_count = 1;
+        so->active_variant = 0;
+
+        if (cso->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
+                struct blob_reader reader;
+                const struct pipe_binary_program_header *hdr = cso->prog;
+
+                blob_reader_init(&reader, hdr->blob, hdr->num_bytes);
+                so->cbase.prog = nir_deserialize(NULL, &midgard_nir_options, &reader);
+                so->cbase.ir_type = PIPE_SHADER_IR_NIR;
+        }
+
+        panfrost_shader_compile(ctx, so->cbase.ir_type, so->cbase.prog,
+                                MESA_SHADER_COMPUTE, v, NULL);
+
+        return so;
 }
 
 static void
 panfrost_bind_compute_state(struct pipe_context *pipe, void *cso)
 {
-        struct pipe_compute_state *state = (struct pipe_compute_state *) cso;
+        struct panfrost_context *ctx = pan_context(pipe);
 
-        printf("Binding compute %p\n", state);
-        /* Stub */
+        struct panfrost_shader_variants *variants =
+                (struct panfrost_shader_variants *) cso;
+
+        ctx->shader[PIPE_SHADER_COMPUTE] = variants;
 }
 
 static void
@@ -51,15 +87,98 @@ panfrost_delete_compute_state(struct pipe_context *pipe, void *cso)
         free(cso);
 }
 
+/* Launch grid is the compute equivalent of draw_vbo, so in this routine, we
+ * construct the COMPUTE job and some of its payload.
+ */
+
 static void
 panfrost_launch_grid(struct pipe_context *pipe,
                 const struct pipe_grid_info *info)
 {
-        printf("Launch grid %dx%dx%d ... %dx%dx%d\n",
-                        info->block[0], info->block[1], info->block[2],
-                        info->grid[0], info->grid[1], info->grid[2]);
+        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;
+
+        /* TODO: Stub */
+        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 */
+
+        struct pipe_constant_buffer ubuf = {
+                .buffer = NULL,
+                .buffer_offset = 0,
+                .buffer_size = ctx->shader[PIPE_SHADER_COMPUTE]->cbase.req_input_mem,
+                .user_buffer = info->input
+        };
+
+        if (info->input)
+                pipe->set_constant_buffer(pipe, PIPE_SHADER_COMPUTE, 0, &ubuf);
+
+        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);
+        }
+
+        unsigned magic =
+                util_logbase2_ceil(info->block[0] + 1) +
+                util_logbase2_ceil(info->block[1] + 1) +
+                util_logbase2_ceil(info->block[2] + 1);
 
-        /* Stub */
+        payload.prefix.primitive.opaque[0] = (magic) << 26; /* XXX */
+
+        memcpy(&payload.postfix, &postfix, sizeof(postfix));
+
+        /* Invoke according to the grid info */
+
+        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
+panfrost_set_compute_resources(struct pipe_context *pctx,
+                         unsigned start, unsigned count,
+                         struct pipe_surface **resources)
+{
+        /* TODO */
+}
+
+static void
+panfrost_set_global_binding(struct pipe_context *pctx,
+                      unsigned first, unsigned count,
+                      struct pipe_resource **resources,
+                      uint32_t **handles)
+{
+        /* TODO */
+}
+
+static void
+panfrost_memory_barrier(struct pipe_context *pctx, unsigned flags)
+{
+        /* TODO */
 }
 
 void
@@ -70,6 +189,9 @@ panfrost_compute_context_init(struct pipe_context *pctx)
         pctx->delete_compute_state = panfrost_delete_compute_state;
 
         pctx->launch_grid = panfrost_launch_grid;
-}
 
+        pctx->set_compute_resources = panfrost_set_compute_resources;
+        pctx->set_global_binding = panfrost_set_global_binding;
 
+        pctx->memory_barrier = panfrost_memory_barrier;
+}