#include "main/imports.h"
#include "main/mtypes.h"
+#include "main/macros.h"
#include "main/bufferobj.h"
#include "intel_context.h"
+#include "intel_blit.h"
#include "intel_buffer_objects.h"
#include "intel_batchbuffer.h"
#include "intel_regions.h"
+static GLboolean
+intel_bufferobj_unmap(GLcontext * ctx,
+ GLenum target, struct gl_buffer_object *obj);
/** Allocates a new dri_bo to store the data for the buffer object. */
static void
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
assert(intel_obj);
- assert(!obj->Pointer); /* Mesa should have unmapped it */
+
+ /* Buffer objects are automatically unmapped when deleting according
+ * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
+ * (though it does if you call glDeleteBuffers)
+ */
+ if (obj->Pointer)
+ intel_bufferobj_unmap(ctx, 0, obj);
_mesa_free(intel_obj->sys_buffer);
if (intel_obj->region) {
* Allocate space for and store data in a buffer object. Any data that was
* previously stored in the buffer object is lost. If data is NULL,
* memory will be allocated, but no copy will occur.
- * Called via glBufferDataARB().
+ * Called via ctx->Driver.BufferData().
+ * \return GL_TRUE for success, GL_FALSE if out of memory
*/
-static void
+static GLboolean
intel_bufferobj_data(GLcontext * ctx,
GLenum target,
GLsizeiptrARB size,
if (intel_obj->sys_buffer != NULL) {
if (data != NULL)
memcpy(intel_obj->sys_buffer, data, size);
- return;
+ return GL_TRUE;
}
}
#endif
intel_bufferobj_alloc_buffer(intel, intel_obj);
+ if (!intel_obj->buffer)
+ return GL_FALSE;
if (data != NULL)
dri_bo_subdata(intel_obj->buffer, 0, size, data);
}
+
+ return GL_TRUE;
}
if (intel_obj->sys_buffer)
memcpy((char *)intel_obj->sys_buffer + offset, data, size);
- else
+ else {
+ /* Flush any existing batchbuffer that might reference this data. */
+ intelFlush(ctx);
+
dri_bo_subdata(intel_obj->buffer, offset, size, data);
+ }
}
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
assert(intel_obj);
- dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
+ if (intel_obj->sys_buffer)
+ memcpy(data, (char *)intel_obj->sys_buffer + offset, size);
+ else
+ dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
}
return obj->Pointer;
}
+ /* Flush any existing batchbuffer that might have written to this
+ * buffer.
+ */
+ intelFlush(ctx);
+
if (intel_obj->region)
intel_bufferobj_cow(intel, intel_obj);
}
obj->Pointer = intel_obj->buffer->virtual;
+ obj->Length = obj->Size;
+ obj->Offset = 0;
+
return obj->Pointer;
}
+/**
+ * Called via glMapBufferRange().
+ *
+ * The goal of this extension is to allow apps to accumulate their rendering
+ * at the same time as they accumulate their buffer object. Without it,
+ * you'd end up blocking on execution of rendering every time you mapped
+ * the buffer to put new data in.
+ *
+ * We support it in 3 ways: If unsynchronized, then don't bother
+ * flushing the batchbuffer before mapping the buffer, which can save blocking
+ * in many cases. If we would still block, and they allow the whole buffer
+ * to be invalidated, then just allocate a new buffer to replace the old one.
+ * If not, and we'd block, and they allow the subrange of the buffer to be
+ * invalidated, then we can make a new little BO, let them write into that,
+ * and blit it into the real BO at unmap time.
+ */
+static void *
+intel_bufferobj_map_range(GLcontext * ctx,
+ GLenum target, GLintptr offset, GLsizeiptr length,
+ GLbitfield access, struct gl_buffer_object *obj)
+{
+ struct intel_context *intel = intel_context(ctx);
+ struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
+
+ assert(intel_obj);
+
+ /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
+ * internally uses our functions directly.
+ */
+ obj->Offset = offset;
+ obj->Length = length;
+ obj->AccessFlags = access;
+
+ if (intel_obj->sys_buffer) {
+ obj->Pointer = intel_obj->sys_buffer + offset;
+ return obj->Pointer;
+ }
+
+ if (intel_obj->region)
+ intel_bufferobj_cow(intel, intel_obj);
+
+ /* If the mapping is synchronized with other GL operations, flush
+ * the batchbuffer so that GEM knows about the buffer access for later
+ * syncing.
+ */
+ if (!(access & GL_MAP_UNSYNCHRONIZED_BIT))
+ intelFlush(ctx);
+
+ if (intel_obj->buffer == NULL) {
+ obj->Pointer = NULL;
+ return NULL;
+ }
+
+ /* If the user doesn't care about existing buffer contents and mapping
+ * would cause us to block, then throw out the old buffer.
+ */
+ if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
+ (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
+ drm_intel_bo_busy(intel_obj->buffer)) {
+ drm_intel_bo_unreference(intel_obj->buffer);
+ intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
+ intel_obj->Base.Size, 64);
+ }
+
+ /* If the user is mapping a range of an active buffer object but
+ * doesn't require the current contents of that range, make a new
+ * BO, and we'll copy what they put in there out at unmap or
+ * FlushRange time.
+ */
+ if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
+ drm_intel_bo_busy(intel_obj->buffer)) {
+ if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
+ intel_obj->range_map_buffer = _mesa_malloc(length);
+ obj->Pointer = intel_obj->range_map_buffer;
+ } else {
+ intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
+ "range map",
+ length, 64);
+ if (!(access & GL_MAP_READ_BIT) &&
+ intel->intelScreen->kernel_exec_fencing) {
+ drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
+ intel_obj->mapped_gtt = GL_TRUE;
+ } else {
+ drm_intel_bo_map(intel_obj->range_map_bo,
+ (access & GL_MAP_WRITE_BIT) != 0);
+ intel_obj->mapped_gtt = GL_FALSE;
+ }
+ obj->Pointer = intel_obj->range_map_bo->virtual;
+ }
+ return obj->Pointer;
+ }
+
+ if (!(access & GL_MAP_READ_BIT) &&
+ intel->intelScreen->kernel_exec_fencing) {
+ drm_intel_gem_bo_map_gtt(intel_obj->buffer);
+ intel_obj->mapped_gtt = GL_TRUE;
+ } else {
+ drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
+ intel_obj->mapped_gtt = GL_FALSE;
+ }
+
+ obj->Pointer = intel_obj->buffer->virtual + offset;
+ return obj->Pointer;
+}
+
+/* Ideally we'd use a BO to avoid taking up cache space for the temporary
+ * data, but FlushMappedBufferRange may be followed by further writes to
+ * the pointer, so we would have to re-map after emitting our blit, which
+ * would defeat the point.
+ */
+static void
+intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
+ GLintptr offset, GLsizeiptr length,
+ struct gl_buffer_object *obj)
+{
+ struct intel_context *intel = intel_context(ctx);
+ struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
+ drm_intel_bo *temp_bo;
+
+ /* Unless we're in the range map using a temporary system buffer,
+ * there's no work to do.
+ */
+ if (intel_obj->range_map_buffer == NULL)
+ return;
+
+ temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
+
+ drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
+
+ intel_emit_linear_blit(intel,
+ intel_obj->buffer, obj->Offset + offset,
+ temp_bo, 0,
+ length);
+
+ drm_intel_bo_unreference(temp_bo);
+}
+
/**
- * Called via glMapBufferARB().
+ * Called via glUnmapBuffer().
*/
static GLboolean
intel_bufferobj_unmap(GLcontext * ctx,
GLenum target, struct gl_buffer_object *obj)
{
+ struct intel_context *intel = intel_context(ctx);
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
assert(intel_obj);
+ assert(obj->Pointer);
if (intel_obj->sys_buffer != NULL) {
- assert(obj->Pointer);
- obj->Pointer = NULL;
+ /* always keep the mapping around. */
+ } else if (intel_obj->range_map_buffer != NULL) {
+ /* Since we've emitted some blits to buffers that will (likely) be used
+ * in rendering operations in other cache domains in this batch, emit a
+ * flush. Once again, we wish for a domain tracker in libdrm to cover
+ * usage inside of a batchbuffer.
+ */
+ intel_batchbuffer_emit_mi_flush(intel->batch);
+ free(intel_obj->range_map_buffer);
+ intel_obj->range_map_buffer = NULL;
+ } else if (intel_obj->range_map_bo != NULL) {
+ if (intel_obj->mapped_gtt) {
+ drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
+ } else {
+ drm_intel_bo_unmap(intel_obj->range_map_bo);
+ }
+
+ intel_emit_linear_blit(intel,
+ intel_obj->buffer, obj->Offset,
+ intel_obj->range_map_bo, 0,
+ obj->Length);
+
+ /* Since we've emitted some blits to buffers that will (likely) be used
+ * in rendering operations in other cache domains in this batch, emit a
+ * flush. Once again, we wish for a domain tracker in libdrm to cover
+ * usage inside of a batchbuffer.
+ */
+ intel_batchbuffer_emit_mi_flush(intel->batch);
+
+ drm_intel_bo_unreference(intel_obj->range_map_bo);
+ intel_obj->range_map_bo = NULL;
} else if (intel_obj->buffer != NULL) {
- assert(obj->Pointer);
if (intel_obj->mapped_gtt) {
drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
} else {
drm_intel_bo_unmap(intel_obj->buffer);
}
- obj->Pointer = NULL;
}
+ obj->Pointer = NULL;
+ obj->Offset = 0;
+ obj->Length = 0;
+
return GL_TRUE;
}
}
if (intel_obj->buffer == NULL) {
+ void *sys_buffer = intel_obj->sys_buffer;
+
+ /* only one of buffer and sys_buffer could be non-NULL */
intel_bufferobj_alloc_buffer(intel, intel_obj);
+ intel_obj->sys_buffer = NULL;
+
intel_bufferobj_subdata(&intel->ctx,
GL_ARRAY_BUFFER_ARB,
0,
intel_obj->Base.Size,
- intel_obj->sys_buffer,
+ sys_buffer,
&intel_obj->Base);
- _mesa_free(intel_obj->sys_buffer);
+ _mesa_free(sys_buffer);
intel_obj->sys_buffer = NULL;
}
return intel_obj->buffer;
}
+static void
+intel_bufferobj_copy_subdata(GLcontext *ctx,
+ struct gl_buffer_object *src,
+ struct gl_buffer_object *dst,
+ GLintptr read_offset, GLintptr write_offset,
+ GLsizeiptr size)
+{
+ struct intel_context *intel = intel_context(ctx);
+ struct intel_buffer_object *intel_src = intel_buffer_object(src);
+ struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
+ drm_intel_bo *src_bo, *dst_bo;
+
+ if (size == 0)
+ return;
+
+ /* If we're in system memory, just map and memcpy. */
+ if (intel_src->sys_buffer || intel_dst->sys_buffer) {
+ /* The same buffer may be used, but note that regions copied may
+ * not overlap.
+ */
+ if (src == dst) {
+ char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
+ GL_READ_WRITE, dst);
+ memcpy(ptr + write_offset, ptr + read_offset, size);
+ intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
+ } else {
+ const char *src_ptr;
+ char *dst_ptr;
+
+ src_ptr = intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
+ GL_READ_ONLY, src);
+ dst_ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
+ GL_WRITE_ONLY, dst);
+
+ memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
+
+ intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
+ intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
+ }
+ }
+
+ /* Otherwise, we have real BOs, so blit them. */
+
+ dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
+ src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
+
+ intel_emit_linear_blit(intel,
+ dst_bo, write_offset,
+ src_bo, read_offset, size);
+
+ /* Since we've emitted some blits to buffers that will (likely) be used
+ * in rendering operations in other cache domains in this batch, emit a
+ * flush. Once again, we wish for a domain tracker in libdrm to cover
+ * usage inside of a batchbuffer.
+ */
+ intel_batchbuffer_emit_mi_flush(intel->batch);
+}
+
void
-intel_bufferobj_init(struct intel_context *intel)
+intelInitBufferObjectFuncs(struct dd_function_table *functions)
{
- GLcontext *ctx = &intel->ctx;
-
- ctx->Driver.NewBufferObject = intel_bufferobj_alloc;
- ctx->Driver.DeleteBuffer = intel_bufferobj_free;
- ctx->Driver.BufferData = intel_bufferobj_data;
- ctx->Driver.BufferSubData = intel_bufferobj_subdata;
- ctx->Driver.GetBufferSubData = intel_bufferobj_get_subdata;
- ctx->Driver.MapBuffer = intel_bufferobj_map;
- ctx->Driver.UnmapBuffer = intel_bufferobj_unmap;
+ functions->NewBufferObject = intel_bufferobj_alloc;
+ functions->DeleteBuffer = intel_bufferobj_free;
+ functions->BufferData = intel_bufferobj_data;
+ functions->BufferSubData = intel_bufferobj_subdata;
+ functions->GetBufferSubData = intel_bufferobj_get_subdata;
+ functions->MapBuffer = intel_bufferobj_map;
+ functions->MapBufferRange = intel_bufferobj_map_range;
+ functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
+ functions->UnmapBuffer = intel_bufferobj_unmap;
+ functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
}
#include "util/u_tile.h"
#include "util/u_blit.h"
#include "util/u_surface.h"
+#include "util/u_math.h"
#define DBG if (0) printf
}
-static INLINE unsigned
-logbase2(unsigned n)
-{
- unsigned log2 = 0;
- while (n >>= 1)
- ++log2;
- return log2;
-}
-
-
/**
* Return default texture usage bitmask for the given texture format.
*/
lastLevel = firstLevel;
}
else {
- GLuint l2width = logbase2(width);
- GLuint l2height = logbase2(height);
- GLuint l2depth = logbase2(depth);
+ GLuint l2width = util_logbase2(width);
+ GLuint l2height = util_logbase2(height);
+ GLuint l2depth = util_logbase2(depth);
lastLevel = firstLevel + MAX2(MAX2(l2width, l2height), l2depth);
}
DBG("%s target %s level %d %dx%dx%d border %d\n", __FUNCTION__,
_mesa_lookup_enum_by_nr(target), level, width, height, depth, border);
+ /* switch to "normal" */
+ if (stObj->surface_based) {
+ _mesa_clear_texture_object(ctx, texObj);
+ stObj->surface_based = GL_FALSE;
+ }
+
/* gallium does not support texture borders, strip it off */
if (border) {
strip_texture_border(border, &width, &height, &depth, unpack, &unpackNB);
PIPE_TRANSFER_READ,
0, 0, width, height);
- pixels = _mesa_map_readpix_pbo(ctx, &ctx->Pack, pixels);
+ pixels = _mesa_map_pbo_dest(ctx, &ctx->Pack, pixels);
/* copy/pack data into user buffer */
if (st_equal_formats(stImage->pt->format, format, type)) {
}
}
- _mesa_unmap_readpix_pbo(ctx, &ctx->Pack);
+ _mesa_unmap_pbo_dest(ctx, &ctx->Pack);
/* destroy the temp / dest surface */
util_destroy_rgba_surface(dst_texture, dst_surface);
}
- /**
- * Compute which mipmap levels that really need to be sent to the hardware.
- * This depends on the base image size, GL_TEXTURE_MIN_LOD,
- * GL_TEXTURE_MAX_LOD, GL_TEXTURE_BASE_LEVEL, and GL_TEXTURE_MAX_LEVEL.
- */
- static void
- calculate_first_last_level(struct st_texture_object *stObj)
- {
- struct gl_texture_object *tObj = &stObj->base;
-
- /* These must be signed values. MinLod and MaxLod can be negative numbers,
- * and having firstLevel and lastLevel as signed prevents the need for
- * extra sign checks.
- */
- GLint firstLevel;
- GLint lastLevel;
-
- /* Yes, this looks overly complicated, but it's all needed.
- */
- switch (tObj->Target) {
- case GL_TEXTURE_1D:
- case GL_TEXTURE_2D:
- case GL_TEXTURE_3D:
- case GL_TEXTURE_CUBE_MAP:
- if (tObj->MinFilter == GL_NEAREST || tObj->MinFilter == GL_LINEAR) {
- /* GL_NEAREST and GL_LINEAR only care about GL_TEXTURE_BASE_LEVEL.
- */
- firstLevel = lastLevel = tObj->BaseLevel;
- }
- else {
- firstLevel = 0;
- lastLevel = MIN2(tObj->MaxLevel,
- (int) tObj->Image[0][tObj->BaseLevel]->WidthLog2);
- }
- break;
- case GL_TEXTURE_RECTANGLE_NV:
- case GL_TEXTURE_4D_SGIS:
- firstLevel = lastLevel = 0;
- break;
- default:
- return;
- }
-
- stObj->lastLevel = lastLevel;
- }
-
-
static void
copy_image_data_to_texture(struct st_context *st,
struct st_texture_object *stObj,
*needFlush = GL_FALSE;
- /* We know/require this is true by now:
- */
- assert(stObj->base._Complete);
+ if (stObj->base._Complete) {
+ /* The texture is complete and we know exactly how many mipmap levels
+ * are present/needed. This is conditional because we may be called
+ * from the st_generate_mipmap() function when the texture object is
+ * incomplete. In that case, we'll have set stObj->lastLevel before
+ * we get here.
+ */
+ stObj->lastLevel = stObj->base._MaxLevel - stObj->base.BaseLevel;
+ }
- /* What levels must the texture include at a minimum?
- */
- calculate_first_last_level(stObj);
firstImage = st_texture_image(stObj->base.Image[0][stObj->base.BaseLevel]);
/* If both firstImage and stObj point to a texture which can contain