u_upload_mgr: pass alignment to u_upload_alloc manually

The fixed alignment of u_upload_mgr will go away.
This is the first step.

The motivation is that one u_upload_mgr can have multiple users,
each allocating from the same buffer, but requiring a different alignment.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

src/gallium/auxiliary/hud/hud_context.c

@@ -431,7 +431,7 @@ hud_alloc_vertices(struct hud_context *hud, struct vertex_queue *v,
    v->max_num_vertices = num_vertices;
    v->vbuf.stride = stride;
    u_upload_alloc(hud->uploader, 0, v->vbuf.stride * v->max_num_vertices,
-                  &v->vbuf.buffer_offset, &v->vbuf.buffer,
+                  16, &v->vbuf.buffer_offset, &v->vbuf.buffer,
                   (void**)&v->vertices);
 }
 

src/gallium/auxiliary/indices/u_primconvert.c

@@ -156,7 +156,7 @@ util_primconvert_draw_vbo(struct primconvert_context *pc,
       pc->upload = u_upload_create(pc->pipe, 4096, 4, PIPE_BIND_INDEX_BUFFER);
    }
 
-   u_upload_alloc(pc->upload, 0, new_ib.index_size * new_info.count,
+   u_upload_alloc(pc->upload, 0, new_ib.index_size * new_info.count, 4,
                   &new_ib.offset, &new_ib.buffer, &dst);
 
    if (info->indexed) {

src/gallium/auxiliary/util/u_upload_mgr.c

@@ -181,11 +181,11 @@ void
 u_upload_alloc(struct u_upload_mgr *upload,
                unsigned min_out_offset,
                unsigned size,
+               unsigned alignment,
                unsigned *out_offset,
                struct pipe_resource **outbuf,
                void **ptr)
 {
-   unsigned alignment = upload->alignment;
    unsigned buffer_size = upload->buffer ? upload->buffer->width0 : 0;
    unsigned offset;
 
@@ -249,7 +249,7 @@ void u_upload_data(struct u_upload_mgr *upload,
 {
    uint8_t *ptr;
 
-   u_upload_alloc(upload, min_out_offset, size,
+   u_upload_alloc(upload, min_out_offset, size, upload->alignment,
                   out_offset, outbuf,
                   (void**)&ptr);
    if (ptr)

src/gallium/auxiliary/util/u_upload_mgr.h

@@ -74,6 +74,7 @@ void u_upload_unmap( struct u_upload_mgr *upload );
  * \param upload           Upload manager
  * \param min_out_offset   Minimum offset that should be returned in out_offset.
  * \param size             Size of the allocation.
+ * \param alignment        Alignment of the suballocation within the buffer
  * \param out_offset       Pointer to where the new buffer offset will be returned.
  * \param outbuf           Pointer to where the upload buffer will be returned.
  * \param ptr              Pointer to the allocated memory that is returned.
@@ -81,6 +82,7 @@ void u_upload_unmap( struct u_upload_mgr *upload );
 void u_upload_alloc(struct u_upload_mgr *upload,
                     unsigned min_out_offset,
                     unsigned size,
+                    unsigned alignment,
                     unsigned *out_offset,
                     struct pipe_resource **outbuf,
                     void **ptr);

src/gallium/auxiliary/util/u_vbuf.c

@@ -454,7 +454,7 @@ u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key,
 
       /* Create and map the output buffer. */
       u_upload_alloc(mgr->uploader, 0,
-                     key->output_stride * num_indices,
+                     key->output_stride * num_indices, 4,
                      &out_offset, &out_buffer,
                      (void**)&out_map);
       if (!out_buffer)
@@ -487,7 +487,7 @@ u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key,
       /* Create and map the output buffer. */
       u_upload_alloc(mgr->uploader,
                      key->output_stride * start_vertex,
-                     key->output_stride * num_vertices,
+                     key->output_stride * num_vertices, 4,
                      &out_offset, &out_buffer,
                      (void**)&out_map);
       if (!out_buffer)

src/gallium/auxiliary/vl/vl_compositor.c

@@ -716,6 +716,7 @@ gen_vertex_data(struct vl_compositor *c, struct vl_compositor_state *s, struct u
    /* Allocate new memory for vertices. */
    u_upload_alloc(c->upload, 0,
                   c->vertex_buf.stride * VL_COMPOSITOR_MAX_LAYERS * 4, /* size */
+                  4, /* alignment */
                   &c->vertex_buf.buffer_offset, &c->vertex_buf.buffer,
                   (void**)&vb);
 

src/gallium/drivers/freedreno/a3xx/fd3_context.c

@@ -172,7 +172,7 @@ fd3_context_create(struct pipe_screen *pscreen, void *priv, unsigned flags)
 	fd3_query_context_init(pctx);
 
 	fd3_ctx->border_color_uploader = u_upload_create(pctx, 4096,
-			2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE, 0);
+			BORDER_COLOR_UPLOAD_SIZE, 0);
 
 	return pctx;
 }

src/gallium/drivers/freedreno/a3xx/fd3_emit.c

@@ -145,7 +145,8 @@ emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
 	void *ptr;
 
 	u_upload_alloc(fd3_ctx->border_color_uploader,
-			0, 2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE, &off,
+			0, BORDER_COLOR_UPLOAD_SIZE,
+		       BORDER_COLOR_UPLOAD_SIZE, &off,
 			&fd3_ctx->border_color_buf,
 			&ptr);
 

src/gallium/drivers/freedreno/a4xx/fd4_context.c

@@ -172,7 +172,7 @@ fd4_context_create(struct pipe_screen *pscreen, void *priv, unsigned flags)
 	fd4_query_context_init(pctx);
 
 	fd4_ctx->border_color_uploader = u_upload_create(pctx, 4096,
-			2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE, 0);
+			BORDER_COLOR_UPLOAD_SIZE, 0);
 
 	return pctx;
 }

src/gallium/drivers/freedreno/a4xx/fd4_emit.c

@@ -133,7 +133,8 @@ emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
 	void *ptr;
 
 	u_upload_alloc(fd4_ctx->border_color_uploader,
-			0, 2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE, &off,
+			0, BORDER_COLOR_UPLOAD_SIZE,
+		       BORDER_COLOR_UPLOAD_SIZE, &off,
 			&fd4_ctx->border_color_buf,
 			&ptr);
 

src/gallium/drivers/freedreno/freedreno_context.h

@@ -40,6 +40,8 @@
 #include "freedreno_gmem.h"
 #include "freedreno_util.h"
 
+#define BORDER_COLOR_UPLOAD_SIZE (2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE)
+
 struct fd_vertex_stateobj;
 
 struct fd_texture_stateobj {

src/gallium/drivers/r300/r300_render_translate.c

@@ -37,7 +37,7 @@ void r300_translate_index_buffer(struct r300_context *r300,
     switch (*index_size) {
     case 1:
         *out_buffer = NULL;
-        u_upload_alloc(r300->uploader, 0, count * 2,
+        u_upload_alloc(r300->uploader, 0, count * 2, 4,
                        &out_offset, out_buffer, &ptr);
 
         util_shorten_ubyte_elts_to_userptr(
@@ -51,7 +51,7 @@ void r300_translate_index_buffer(struct r300_context *r300,
     case 2:
         if (index_offset) {
             *out_buffer = NULL;
-            u_upload_alloc(r300->uploader, 0, count * 2,
+            u_upload_alloc(r300->uploader, 0, count * 2, 4,
                            &out_offset, out_buffer, &ptr);
 
             util_rebuild_ushort_elts_to_userptr(&r300->context, ib,
@@ -65,7 +65,7 @@ void r300_translate_index_buffer(struct r300_context *r300,
     case 4:
         if (index_offset) {
             *out_buffer = NULL;
-            u_upload_alloc(r300->uploader, 0, count * 4,
+            u_upload_alloc(r300->uploader, 0, count * 4, 4,
                            &out_offset, out_buffer, &ptr);
 
             util_rebuild_uint_elts_to_userptr(&r300->context, ib,

src/gallium/drivers/r600/r600_state_common.c

@@ -1732,7 +1732,7 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 				}
 			}
 
-			u_upload_alloc(rctx->b.uploader, start, count * 2,
+			u_upload_alloc(rctx->b.uploader, start, count * 2, 256,
 				       &out_offset, &out_buffer, &ptr);
 
 			util_shorten_ubyte_elts_to_userptr(

src/gallium/drivers/radeon/r600_buffer_common.c

@@ -298,7 +298,7 @@ static void *r600_buffer_transfer_map(struct pipe_context *ctx,
 			struct r600_resource *staging = NULL;
 
 			u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
-				       &offset, (struct pipe_resource**)&staging, (void**)&data);
+				       256, &offset, (struct pipe_resource**)&staging, (void**)&data);
 
 			if (staging) {
 				data += box->x % R600_MAP_BUFFER_ALIGNMENT;

src/gallium/drivers/radeon/r600_pipe_common.c

@@ -85,7 +85,7 @@ void r600_draw_rectangle(struct blitter_context *blitter,
 	/* Upload vertices. The hw rectangle has only 3 vertices,
 	 * I guess the 4th one is derived from the first 3.
 	 * The vertex specification should match u_blitter's vertex element state. */
-	u_upload_alloc(rctx->uploader, 0, sizeof(float) * 24, &offset, &buf, (void**)&vb);
+	u_upload_alloc(rctx->uploader, 0, sizeof(float) * 24, 256, &offset, &buf, (void**)&vb);
 	if (!buf)
 		return;
 

src/gallium/drivers/radeonsi/si_descriptors.c

@@ -109,7 +109,7 @@ static bool si_upload_descriptors(struct si_context *sctx,
 	if (!desc->list_dirty)
 		return true;
 
-	u_upload_alloc(sctx->b.uploader, 0, list_size,
+	u_upload_alloc(sctx->b.uploader, 0, list_size, 256,
 		       &desc->buffer_offset,
 		       (struct pipe_resource**)&desc->buffer, &ptr);
 	if (!desc->buffer)
@@ -391,7 +391,7 @@ static bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 	 * directly through a staging buffer and don't go through
 	 * the fine-grained upload path.
 	 */
-	u_upload_alloc(sctx->b.uploader, 0, count * 16, &desc->buffer_offset,
+	u_upload_alloc(sctx->b.uploader, 0, count * 16, 256, &desc->buffer_offset,
 		       (struct pipe_resource**)&desc->buffer, (void**)&ptr);
 	if (!desc->buffer)
 		return false;
@@ -465,7 +465,7 @@ void si_upload_const_buffer(struct si_context *sctx, struct r600_resource **rbuf
 {
 	void *tmp;
 
-	u_upload_alloc(sctx->b.uploader, 0, size, const_offset,
+	u_upload_alloc(sctx->b.uploader, 0, size, 256, const_offset,
 		       (struct pipe_resource**)rbuffer, &tmp);
 	if (rbuffer)
 		util_memcpy_cpu_to_le32(tmp, ptr, size);

src/gallium/drivers/radeonsi/si_state_draw.c

@@ -818,7 +818,7 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 			si_get_draw_start_count(sctx, info, &start, &count);
 			start_offset = start * ib.index_size;
 
-			u_upload_alloc(sctx->b.uploader, start_offset, count * 2,
+			u_upload_alloc(sctx->b.uploader, start_offset, count * 2, 256,
 				       &out_offset, &out_buffer, &ptr);
 			if (!out_buffer) {
 				pipe_resource_reference(&ib.buffer, NULL);

src/gallium/drivers/svga/svga_context.c

@@ -46,7 +46,6 @@
 #include "svga_winsys.h"
 
 #define CONST0_UPLOAD_DEFAULT_SIZE 65536
-#define CONST0_UPLOAD_ALIGNMENT 256
 
 DEBUG_GET_ONCE_BOOL_OPTION(no_swtnl, "SVGA_NO_SWTNL", FALSE)
 DEBUG_GET_ONCE_BOOL_OPTION(force_swtnl, "SVGA_FORCE_SWTNL", FALSE);

src/gallium/drivers/svga/svga_context.h

@@ -74,6 +74,8 @@
  */
 #define SVGA_MAX_CONST_BUF_SIZE (4096 * 4 * sizeof(int))
 
+#define CONST0_UPLOAD_ALIGNMENT 256
+
 struct draw_vertex_shader;
 struct draw_fragment_shader;
 struct svga_shader_variant;

src/gallium/drivers/svga/svga_state_constants.c

@@ -613,7 +613,8 @@ emit_constbuf_vgpu10(struct svga_context *svga, unsigned shader)
     */
    new_buf_size = align(new_buf_size, 16);
 
-   u_upload_alloc(svga->const0_upload, 0, new_buf_size, &offset,
+   u_upload_alloc(svga->const0_upload, 0, new_buf_size,
+                  CONST0_UPLOAD_ALIGNMENT, &offset,
                   &dst_buffer, &dst_map);
    if (!dst_map) {
       if (src_map)

src/gallium/drivers/vc4/vc4_resource.c

@@ -921,7 +921,7 @@ vc4_get_shadow_index_buffer(struct pipe_context *pctx,
 
         void *data;
         struct pipe_resource *shadow_rsc = NULL;
-        u_upload_alloc(vc4->uploader, 0, count * 2,
+        u_upload_alloc(vc4->uploader, 0, count * 2, 4,;
                        shadow_offset, &shadow_rsc, &data);
         uint16_t *dst = data;
 

src/mesa/state_tracker/st_cb_bitmap.c

@@ -204,7 +204,7 @@ setup_bitmap_vertex_data(struct st_context *st, bool normalized,
       tBot = (GLfloat) height;
    }
 
-   u_upload_alloc(st->uploader, 0, 4 * sizeof(vertices[0]),
+   u_upload_alloc(st->uploader, 0, 4 * sizeof(vertices[0]), 4,
                   vbuf_offset, vbuf, (void **) &vertices);
    if (!*vbuf) {
       return;

src/mesa/state_tracker/st_cb_clear.c

@@ -184,7 +184,7 @@ draw_quad(struct st_context *st,
 
    vb.stride = 8 * sizeof(float);
 
-   u_upload_alloc(st->uploader, 0, 4 * sizeof(vertices[0]),
+   u_upload_alloc(st->uploader, 0, 4 * sizeof(vertices[0]), 4,
                   &vb.buffer_offset, &vb.buffer,
                   (void **) &vertices);
    if (!vb.buffer) {

src/mesa/state_tracker/st_cb_drawpixels.c

@@ -457,7 +457,7 @@ draw_quad(struct gl_context *ctx, GLfloat x0, GLfloat y0, GLfloat z,
    struct pipe_resource *buf = NULL;
    unsigned offset;
 
-   u_upload_alloc(st->uploader, 0, 4 * sizeof(verts[0]), &offset,
+   u_upload_alloc(st->uploader, 0, 4 * sizeof(verts[0]), 4, &offset,
                   &buf, (void **) &verts);
    if (!buf) {
       return;

src/mesa/state_tracker/st_cb_drawtex.c

@@ -150,7 +150,7 @@ st_DrawTex(struct gl_context *ctx, GLfloat x, GLfloat y, GLfloat z,
       GLuint attr;
 
       u_upload_alloc(st->uploader, 0,
-                     numAttribs * 4 * 4 * sizeof(GLfloat),
+                     numAttribs * 4 * 4 * sizeof(GLfloat), 4,
                      &offset, &vbuffer, (void **) &vbuf);
       if (!vbuffer) {
          return;