streamout: WIP getting there slowlyso-wip

4 files changed, 80 insertions, 43 deletions

diff --git a/src/virgl_protocol.h b/src/virgl_protocol.h
index c6704e0..33d43e1 100644
--- a/src/virgl_protocol.h
+++ b/src/virgl_protocol.h
@@ -246,7 +246,7 @@ enum virgl_context_cmd {
 #define VIRGL_SET_UNIFORM_BUFFER_RES_HANDLE 5
 
 /* draw VBO */
-#define VIRGL_DRAW_VBO_SIZE 11
+#define VIRGL_DRAW_VBO_SIZE 12
 #define VIRGL_DRAW_VBO_START 1
 #define VIRGL_DRAW_VBO_COUNT 2
 #define VIRGL_DRAW_VBO_MODE 3
@@ -258,6 +258,7 @@ enum virgl_context_cmd {
 #define VIRGL_DRAW_VBO_RESTART_INDEX 9
 #define VIRGL_DRAW_VBO_MIN_INDEX 10
 #define VIRGL_DRAW_VBO_MAX_INDEX 11
+#define VIRGL_DRAW_VBO_COUNT_FROM_SO 12
 
 /* create surface */
 #define VIRGL_OBJ_SURFACE_SIZE 5
@@ -416,7 +417,8 @@ enum virgl_context_cmd {
 
 /* set streamout targets */
 #define VIRGL_SET_STREAMOUT_TARGETS_APPEND_BITMASK 1
-#define VIRGL_SET_STREAMOUT_TARGETS_H0 2
+#define VIRGL_SET_STREAMOUT_TARGETS_FLAGS 2
+#define VIRGL_SET_STREAMOUT_TARGETS_H0 3
 
 /* set sample mask */
 #define VIRGL_SET_SAMPLE_MASK_SIZE 1
diff --git a/src/vrend_decode.c b/src/vrend_decode.c
index 1da942c..26b4a30 100644
--- a/src/vrend_decode.c
+++ b/src/vrend_decode.c
@@ -343,7 +343,7 @@ static int vrend_decode_resource_inline_write(struct vrend_decode_ctx *ctx, uint
 static int vrend_decode_draw_vbo(struct vrend_decode_ctx *ctx, int length)
 {
    struct pipe_draw_info info;
-
+   uint32_t cso;
    if (length != VIRGL_DRAW_VBO_SIZE)
       return EINVAL;
    memset(&info, 0, sizeof(struct pipe_draw_info));
@@ -359,7 +359,10 @@ static int vrend_decode_draw_vbo(struct vrend_decode_ctx *ctx, int length)
    info.restart_index = get_buf_entry(ctx, VIRGL_DRAW_VBO_RESTART_INDEX);
    info.min_index = get_buf_entry(ctx, VIRGL_DRAW_VBO_MIN_INDEX);
    info.max_index = get_buf_entry(ctx, VIRGL_DRAW_VBO_MAX_INDEX);
-   vrend_draw_vbo(ctx->grctx, &info);
+
+   cso = get_buf_entry(ctx, VIRGL_DRAW_VBO_COUNT_FROM_SO);
+
+   vrend_draw_vbo(ctx->grctx, &info, cso);
    return 0;
 }
 
@@ -994,16 +997,17 @@ static int vrend_decode_set_streamout_targets(struct vrend_decode_ctx *ctx,
                                               uint16_t length)
 {
    uint32_t handles[16];
-   uint32_t num_handles = length - 1;
-   uint32_t append_bitmask;
+   uint32_t num_handles = length - 2;
+   uint32_t append_bitmask, flags;
    int i;
 
    if (length < 1)
       return EINVAL;
    append_bitmask = get_buf_entry(ctx, VIRGL_SET_STREAMOUT_TARGETS_APPEND_BITMASK);
+   flags = get_buf_entry(ctx, VIRGL_SET_STREAMOUT_TARGETS_FLAGS);
    for (i = 0; i < num_handles; i++)
       handles[i] = get_buf_entry(ctx, VIRGL_SET_STREAMOUT_TARGETS_H0 + i);
-   vrend_set_streamout_targets(ctx->grctx, append_bitmask, num_handles, handles);
+   vrend_set_streamout_targets(ctx->grctx, append_bitmask, flags,num_handles, handles);
    return 0;
 }
 
diff --git a/src/vrend_renderer.c b/src/vrend_renderer.c
index b27db9b..654f198 100644
--- a/src/vrend_renderer.c
+++ b/src/vrend_renderer.c
@@ -246,6 +246,15 @@ struct vrend_viewport {
    GLclampd near_val, far_val;
 };
 
+/* create a streamout object to support pause/resume */
+struct vrend_streamout_object {
+   GLuint id;
+   uint32_t num_targets;
+   uint32_t handles[16];
+   struct list_head head;
+   struct vrend_so_target *so_targets[16];
+};
+
 #define XFB_STATE_OFF 0
 #define XFB_STATE_STARTED_NEED_BEGIN 1
 #define XFB_STATE_STARTED 2
@@ -313,8 +322,11 @@ struct vrend_sub_context {
    struct pipe_depth_stencil_alpha_state dsa_state;
    struct pipe_rasterizer_state rs_state;
 
+   /* number in use - 0 for paused */
    int num_so_targets;
-   struct vrend_so_target *so_targets[16];
+   /* number of bounds */
+   int bound_so_targets;
+
 
    uint8_t stencil_refs[2];
 
@@ -334,6 +346,8 @@ struct vrend_sub_context {
    struct pipe_blend_state hw_blend_state;
 
    int xfb_state;
+
+   struct list_head streamout_list;
 };
 
 struct vrend_context {
@@ -2091,7 +2105,8 @@ static GLenum get_xfb_mode(GLenum mode)
 }
 
 void vrend_draw_vbo(struct vrend_context *ctx,
-                    const struct pipe_draw_info *info)
+                    const struct pipe_draw_info *info,
+		    uint32_t cso)
 {
    int i;
    int sampler_id;
@@ -2400,12 +2415,14 @@ void vrend_draw_vbo(struct vrend_context *ctx,
    /* set the vertex state up now on a delay */
    if (!info->indexed) {
       GLenum mode = info->mode;
+      int count = cso ? cso : info->count;
+      int start = cso ? 0 : info->start;
       if (info->instance_count <= 1)
-         glDrawArrays(mode, info->start, info->count);
+         glDrawArrays(mode, start, count);
       else if (info->start_instance)
-         glDrawArraysInstancedBaseInstance(mode, info->start, info->count, info->instance_count, info->start_instance);
+         glDrawArraysInstancedBaseInstance(mode, start, count, info->instance_count, info->start_instance);
       else
-         glDrawArraysInstancedARB(mode, info->start, info->count, info->instance_count);
+         glDrawArraysInstancedARB(mode, start, count, info->instance_count);
    } else {
       GLenum elsz;
       GLenum mode = info->mode;
@@ -3289,7 +3306,7 @@ bool vrend_destroy_context(struct vrend_context *ctx)
    vrend_set_num_sampler_views(ctx, PIPE_SHADER_FRAGMENT, 0, 0);
    vrend_set_num_sampler_views(ctx, PIPE_SHADER_GEOMETRY, 0, 0);
 
-   vrend_set_streamout_targets(ctx, 0, 0, NULL);
+   vrend_set_streamout_targets(ctx, 0, 0, 0, NULL);
    vrend_set_num_vbo(ctx, 0);
 
    vrend_set_index_buffer(ctx, 0, 0, 0);
@@ -4424,54 +4441,64 @@ void vrend_set_sample_mask(struct vrend_context *ctx, unsigned sample_mask)
 }
 
 
-static void vrend_hw_emit_streamout_targets(struct vrend_context *ctx)
+static void vrend_hw_emit_streamout_targets(struct vrend_context *ctx, struct vrend_streamout_object *so_obj)
 {
    int i;
 
-   for (i = 0; i < ctx->sub->num_so_targets; i++) {
-      if (ctx->sub->so_targets[i]->buffer_offset || ctx->sub->so_targets[i]->buffer_size < ctx->sub->so_targets[i]->buffer->base.width0)
-         glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, i, ctx->sub->so_targets[i]->buffer->id, ctx->sub->so_targets[i]->buffer_offset, ctx->sub->so_targets[i]->buffer_size);
+   for (i = 0; i < so_obj->num_targets; i++) {
+      if (so_obj->so_targets[i]->buffer_offset || so_obj->so_targets[i]->buffer_size < so_obj->so_targets[i]->buffer->base.width0)
+         glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, i, so_obj->so_targets[i]->buffer->id, so_obj->so_targets[i]->buffer_offset, so_obj->so_targets[i]->buffer_size);
       else
-         glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, i, ctx->sub->so_targets[i]->buffer->id);
+         glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, i, so_obj->so_targets[i]->buffer->id);
    }
 }
 
 void vrend_set_streamout_targets(struct vrend_context *ctx,
                                  uint32_t append_bitmask,
+				 uint32_t flags,
                                  uint32_t num_targets,
                                  uint32_t *handles)
 {
    struct vrend_so_target *target;
    int i;
-   int old_num = ctx->sub->num_so_targets;
-
-   ctx->sub->num_so_targets = num_targets;
-
-   if (num_targets == 0 && (ctx->sub->xfb_state == XFB_STATE_STARTED || ctx->sub->xfb_state == XFB_STATE_PAUSED)) {
-      glEndTransformFeedback();
-      ctx->sub->xfb_state = XFB_STATE_OFF;
-   }
-
-   for (i = 0; i < num_targets; i++) {
-      target = vrend_object_lookup(ctx->sub->object_hash, handles[i], VIRGL_OBJECT_STREAMOUT_TARGET);
-      if (!target) {
-         report_context_error(ctx, VIRGL_ERROR_CTX_ILLEGAL_HANDLE, handles[i]);
-         return;
+   int old_num = ctx->sub->bound_so_targets;
+   bool new_buffers = false;
+
+   if (num_targets) {
+      bool found = false;
+      struct vrend_streamout_object *obj;
+      LIST_FOR_EACH_ENTRY(obj, &ctx->sub->streamout_list, head) {
+	 if (obj->num_targets == num_targets) {
+	    if (!memcmp(handles, obj->handles, num_targets * 4)) {
+	       found = true;
+	       break;
+	    }
+	 }
+      }
+      if (found) {
+	 glBindTransformFeedback(GL_TRANSFORM_FEEDBACK, obj->id);
+	 return;
       }
-      vrend_so_target_reference(&ctx->sub->so_targets[i], target);
-   }
 
-   for (i = num_targets; i < old_num; i++) {
-      glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, i, 0);
-      vrend_so_target_reference(&ctx->sub->so_targets[i], NULL);
+      obj = CALLOC_STRUCT(vrend_streamout_object);
+      glGenTransformFeedbacks(1, &obj->id);
+      glBindTransformFeedback(GL_TRANSFORM_FEEDBACK, obj->id);
+      obj->num_targets = num_targets;
+      for (i = 0; i < num_targets; i++) {
+	 obj->handles[i] = handles[i];
+	 target = vrend_object_lookup(ctx->sub->object_hash, handles[i], VIRGL_OBJECT_STREAMOUT_TARGET);
+	 if (!target) {
+	    report_context_error(ctx, VIRGL_ERROR_CTX_ILLEGAL_HANDLE, handles[i]);
+	    return;
+	 }
+	 vrend_so_target_reference(&obj->so_targets[i], target);
+      }
+      vrend_hw_emit_streamout_targets(ctx, obj);
+      list_addtail(&obj->head, &ctx->sub->streamout_list);
    }
 
-   vrend_hw_emit_streamout_targets(ctx);
-
-   if (ctx->sub->num_so_targets) {
-      if (ctx->sub->xfb_state == XFB_STATE_OFF)
+   if (ctx->sub->xfb_state == XFB_STATE_OFF && num_targets)
         ctx->sub->xfb_state = XFB_STATE_STARTED_NEED_BEGIN;
-   }
 
 }
 
@@ -5566,6 +5593,8 @@ void vrend_renderer_create_sub_ctx(struct vrend_context *ctx, int sub_ctx_id)
    glGenFramebuffers(2, sub->blit_fb_ids);
 
    list_inithead(&sub->programs);
+   list_inithead(&sub->streamout_list);
+
    sub->object_hash = vrend_object_init_ctx_table();
    vrend_bind_va(sub->vaoid);
 
diff --git a/src/vrend_renderer.h b/src/vrend_renderer.h
index a08aa8e..8eff9fa 100644
--- a/src/vrend_renderer.h
+++ b/src/vrend_renderer.h
@@ -124,7 +124,8 @@ void vrend_clear(struct vrend_context *ctx,
                  double depth, unsigned stencil);
 
 void vrend_draw_vbo(struct vrend_context *ctx,
-                    const struct pipe_draw_info *info);
+                    const struct pipe_draw_info *info,
+		    uint32_t cso);
 
 void vrend_set_framebuffer_state(struct vrend_context *ctx,
                                  uint32_t nr_cbufs, uint32_t surf_handle[8],
@@ -191,6 +192,7 @@ int vrend_create_so_target(struct vrend_context *ctx,
 
 void vrend_set_streamout_targets(struct vrend_context *ctx,
                                  uint32_t append_bitmask,
+				 uint32_t flags,
                                  uint32_t num_targets,
                                  uint32_t *handles);