freedreno/a6xx: pre-calculate userconst stateobj size

The AnTuTu "garden" benchmark overflows the fixed size constbuffer
stateobject, so lets be more clever and calculate (a potentially
slightly pessimistic) actual size.

Signed-off-by: Rob Clark <robdclark@chromium.org>

src/freedreno/ir3/ir3_shader.h

@@ -527,6 +527,7 @@ struct ir3_ubo_analysis_state
 	struct ir3_ubo_range range[IR3_MAX_CONSTANT_BUFFERS];
 	uint32_t size;
 	uint32_t lower_count;
+	uint32_t cmdstream_size; /* for per-gen backend to stash required cmdstream size */
 };
 
 

src/gallium/drivers/freedreno/a6xx/fd6_emit.c

@@ -928,7 +928,8 @@ fd6_emit_state(struct fd_ringbuffer *ring, struct fd6_emit *emit)
 
 	if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & DIRTY_CONST) {
 		struct fd_ringbuffer *vsconstobj = fd_submit_new_ringbuffer(
-				ctx->batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
+				ctx->batch->submit, vp->shader->ubo_state.cmdstream_size,
+				FD_RINGBUFFER_STREAMING);
 
 		ir3_emit_user_consts(ctx->screen, vp, vsconstobj,
 				&ctx->constbuf[PIPE_SHADER_VERTEX]);
@@ -940,7 +941,8 @@ fd6_emit_state(struct fd_ringbuffer *ring, struct fd6_emit *emit)
 
 	if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & DIRTY_CONST) {
 		struct fd_ringbuffer *fsconstobj = fd_submit_new_ringbuffer(
-				ctx->batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
+				ctx->batch->submit, fp->shader->ubo_state.cmdstream_size,
+				FD_RINGBUFFER_STREAMING);
 
 		ir3_emit_user_consts(ctx->screen, fp, fsconstobj,
 				&ctx->constbuf[PIPE_SHADER_FRAGMENT]);

src/gallium/drivers/freedreno/a6xx/fd6_program.c

@@ -45,7 +45,21 @@ create_shader_stateobj(struct pipe_context *pctx, const struct pipe_shader_state
 {
 	struct fd_context *ctx = fd_context(pctx);
 	struct ir3_compiler *compiler = ctx->screen->compiler;
-	return ir3_shader_create(compiler, cso, type, &ctx->debug, pctx->screen);
+	struct ir3_shader *shader =
+		ir3_shader_create(compiler, cso, type, &ctx->debug, pctx->screen);
+	unsigned packets, size;
+
+	/* pre-calculate size required for userconst stateobj: */
+	ir3_user_consts_size(&shader->ubo_state, &packets, &size);
+
+	/* also account for UBO addresses: */
+	packets += 1;
+	size += 2 * shader->const_state.num_ubos;
+
+	unsigned sizedwords = (4 * packets) + size;
+	shader->ubo_state.cmdstream_size = sizedwords * 4;
+
+	return shader;
 }
 
 static void *

src/gallium/drivers/freedreno/ir3/ir3_gallium.c

@@ -211,6 +211,29 @@ emit_const(struct fd_screen *screen, struct fd_ringbuffer *ring,
 			offset, size, user_buffer, buffer);
 }
 
+/**
+ * Indirectly calculates size of cmdstream needed for ir3_emit_user_consts().
+ * Returns number of packets, and total size of all the payload.
+ *
+ * The value can be a worst-case, ie. some shader variants may not read all
+ * consts, etc.
+ *
+ * Returns size in dwords.
+ */
+void
+ir3_user_consts_size(struct ir3_ubo_analysis_state *state,
+		unsigned *packets, unsigned *size)
+{
+	*packets = *size = 0;
+
+	for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
+		if (state->range[i].start < state->range[i].end) {
+			*size += state->range[i].end - state->range[i].start;
+			(*packets)++;
+		}
+	}
+}
+
 void
 ir3_emit_user_consts(struct fd_screen *screen, const struct ir3_shader_variant *v,
 		struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)

src/gallium/drivers/freedreno/ir3/ir3_gallium.h

@@ -51,6 +51,8 @@ struct fd_constbuf_stateobj;
 struct fd_shaderbuf_stateobj;
 struct fd_shaderimg_stateobj;
 
+void ir3_user_consts_size(struct ir3_ubo_analysis_state *state,
+		unsigned *packets, unsigned *size);
 void ir3_emit_user_consts(struct fd_screen *screen, const struct ir3_shader_variant *v,
 		struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf);
 void ir3_emit_ubos(struct fd_screen *screen, const struct ir3_shader_variant *v,