freedreno: add global_bindings state

Signed-off-by: Rob Clark <robdclark@gmail.com>

src/gallium/drivers/freedreno/a5xx/fd5_compute.c

@@ -42,6 +42,18 @@ fd5_create_compute_state(struct pipe_context *pctx,
 		const struct pipe_compute_state *cso)
 {
 	struct fd_context *ctx = fd_context(pctx);
+
+	/* req_input_mem will only be non-zero for cl kernels (ie. clover).
+	 * This isn't a perfect test because I guess it is possible (but
+	 * uncommon) for none for the kernel parameters to be a global,
+	 * but ctx->set_global_bindings() can't fail, so this is the next
+	 * best place to fail if we need a newer version of kernel driver:
+	 */
+	if ((cso->req_input_mem > 0) &&
+			fd_device_version(ctx->dev) < FD_VERSION_BO_IOVA) {
+		return NULL;
+	}
+
 	struct ir3_compiler *compiler = ctx->screen->compiler;
 	struct fd5_compute_stateobj *so = CALLOC_STRUCT(fd5_compute_stateobj);
 	so->shader = ir3_shader_create_compute(compiler, cso, &ctx->debug);
@@ -156,6 +168,7 @@ fd5_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info)
 	struct ir3_shader_key key = {0};
 	struct ir3_shader_variant *v;
 	struct fd_ringbuffer *ring = ctx->batch->draw;
+	unsigned i, nglobal = 0;
 
 	emit_setup(ctx);
 
@@ -167,6 +180,23 @@ fd5_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info)
 	fd5_emit_cs_state(ctx, ring, v);
 	ir3_emit_cs_consts(v, ring, ctx, info);
 
+	foreach_bit(i, ctx->global_bindings.enabled_mask)
+		nglobal++;
+
+	if (nglobal > 0) {
+		/* global resources don't otherwise get an OUT_RELOC(), since
+		 * the raw ptr address is emitted ir ir3_emit_cs_consts().
+		 * So to make the kernel aware that these buffers are referenced
+		 * by the batch, emit dummy reloc's as part of a no-op packet
+		 * payload:
+		 */
+		OUT_PKT7(ring, CP_NOP, 2 * nglobal);
+		foreach_bit(i, ctx->global_bindings.enabled_mask) {
+			struct pipe_resource *prsc = ctx->global_bindings.buf[i];
+			OUT_RELOCW(ring, fd_resource(prsc)->bo, 0, 0, 0);
+		}
+	}
+
 	const unsigned *local_size = info->block; // v->shader->nir->info->cs.local_size;
 	const unsigned *num_groups = info->grid;
 	/* for some reason, mesa/st doesn't set info->work_dim, so just assume 3: */

src/gallium/drivers/freedreno/freedreno_context.h

@@ -110,6 +110,12 @@ struct fd_streamout_stateobj {
 	unsigned offsets[PIPE_MAX_SO_BUFFERS];
 };
 
+#define MAX_GLOBAL_BUFFERS 16
+struct fd_global_bindings_stateobj {
+	struct pipe_resource *buf[MAX_GLOBAL_BUFFERS];
+	uint32_t enabled_mask;
+};
+
 /* group together the vertex and vertexbuf state.. for ease of passing
  * around, and because various internal operations (gmem<->mem, etc)
  * need their own vertex state:
@@ -282,6 +288,7 @@ struct fd_context {
 	struct fd_shaderbuf_stateobj shaderbuf[PIPE_SHADER_TYPES];
 	struct fd_shaderimg_stateobj shaderimg[PIPE_SHADER_TYPES];
 	struct fd_streamout_stateobj streamout;
+	struct fd_global_bindings_stateobj global_bindings;
 	struct pipe_clip_state ucp;
 
 	struct pipe_query *cond_query;

src/gallium/drivers/freedreno/freedreno_draw.c

@@ -488,6 +488,12 @@ fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 	foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 		resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 
+	/* For global buffers, we don't really know if read or written, so assume
+	 * the worst:
+	 */
+	foreach_bit(i, ctx->global_bindings.enabled_mask)
+		resource_written(batch, ctx->global_bindings.buf[i]);
+
 	if (info->indirect)
 		resource_read(batch, info->indirect);
 

src/gallium/drivers/freedreno/freedreno_state.c

@@ -495,15 +495,53 @@ fd_set_compute_resources(struct pipe_context *pctx,
 	// TODO
 }
 
+/* used by clover to bind global objects, returning the bo address
+ * via handles[n]
+ */
 static void
 fd_set_global_binding(struct pipe_context *pctx,
 		unsigned first, unsigned count, struct pipe_resource **prscs,
 		uint32_t **handles)
 {
-	/* TODO only used by clover.. seems to need us to return the actual
-	 * gpuaddr of the buffer.. which isn't really exposed to mesa atm.
-	 * How is this used?
-	 */
+	struct fd_context *ctx = fd_context(pctx);
+	struct fd_global_bindings_stateobj *so = &ctx->global_bindings;
+	unsigned mask = 0;
+
+	if (prscs) {
+		for (unsigned i = 0; i < count; i++) {
+			unsigned n = i + first;
+
+			mask |= BIT(n);
+
+			pipe_resource_reference(&so->buf[n], prscs[i]);
+
+			if (so->buf[n]) {
+				struct fd_resource *rsc = fd_resource(so->buf[n]);
+				uint64_t iova = fd_bo_get_iova(rsc->bo);
+				// TODO need to scream if iova > 32b or fix gallium API..
+				*handles[i] += iova;
+			}
+
+			if (prscs[i])
+				so->enabled_mask |= BIT(n);
+			else
+				so->enabled_mask &= ~BIT(n);
+		}
+	} else {
+		mask = (BIT(count) - 1) << first;
+
+		for (unsigned i = 0; i < count; i++) {
+			unsigned n = i + first;
+			if (so->buf[n]) {
+				struct fd_resource *rsc = fd_resource(so->buf[n]);
+				fd_bo_put_iova(rsc->bo);
+			}
+			pipe_resource_reference(&so->buf[n], NULL);
+		}
+
+		so->enabled_mask &= ~mask;
+	}
+
 }
 
 void