dri/nv10: Use fast Z clears.

This commit is contained in:
Francisco Jerez 2010-09-25 15:29:02 +02:00
parent bdd19da218
commit 065163bcd2
6 changed files with 88 additions and 11 deletions

View file

@ -3191,6 +3191,12 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
#define NV17TCL_DMA_IN_MEMORY4 0x000001ac
#define NV17TCL_DMA_IN_MEMORY5 0x000001b0
#define NV17TCL_COLOR_MASK_ENABLE 0x000002bc
#define NV17TCL_ZCLEAR_ENABLE 0x000003f8
#define NV17TCL_ZCLEAR_VALUE 0x000003fc
#define NV17TCL_ZCLEAR_VALUE_DEPTH_SHIFT 8
#define NV17TCL_ZCLEAR_VALUE_DEPTH_MASK 0xffffff00
#define NV17TCL_ZCLEAR_VALUE_SEQUENCE_SHIFT 0
#define NV17TCL_ZCLEAR_VALUE_SEQUENCE_MASK 0x000000ff
#define NV17TCL_LMA_DEPTH_BUFFER_PITCH 0x00000d5c
#define NV17TCL_LMA_DEPTH_BUFFER_OFFSET 0x00000d60
#define NV17TCL_LMA_DEPTH_FILL_VALUE 0x00000d68

View file

@ -67,6 +67,11 @@ struct nouveau_context {
struct nouveau_hw_state hw;
struct nouveau_bo_state bo;
struct nouveau_render_state render;
struct {
GLboolean clear_blocked;
int clear_seq;
} hierz;
};
#define to_nouveau_context(ctx) ((struct nouveau_context *)(ctx))

View file

@ -29,8 +29,12 @@
struct nouveau_framebuffer {
struct gl_framebuffer base;
struct nouveau_bo *lma_bo;
GLboolean need_front;
struct {
struct nouveau_bo *bo;
uint32_t clear_value;
} hierz;
};
#define to_nouveau_framebuffer(x) ((struct nouveau_framebuffer *)(x))

View file

@ -40,9 +40,31 @@ static const struct dri_extension nv10_extensions[] = {
{ NULL, NULL }
};
static GLboolean
use_fast_zclear(GLcontext *ctx, GLbitfield buffers)
{
struct nouveau_context *nctx = to_nouveau_context(ctx);
struct gl_framebuffer *fb = ctx->DrawBuffer;
if (buffers & BUFFER_BIT_STENCIL) {
/*
* The stencil test is bypassed when fast Z clears are
* enabled.
*/
nctx->hierz.clear_blocked = GL_TRUE;
context_dirty(ctx, ZCLEAR);
return GL_FALSE;
}
return !nctx->hierz.clear_blocked &&
fb->_Xmax == fb->Width && fb->_Xmin == 0 &&
fb->_Ymax == fb->Height && fb->_Ymin == 0;
}
static void
nv10_clear(GLcontext *ctx, GLbitfield buffers)
{
struct nouveau_context *nctx = to_nouveau_context(ctx);
struct nouveau_channel *chan = context_chan(ctx);
struct nouveau_grobj *celsius = context_eng3d(ctx);
struct nouveau_framebuffer *nfb = to_nouveau_framebuffer(
@ -50,16 +72,28 @@ nv10_clear(GLcontext *ctx, GLbitfield buffers)
nouveau_validate_framebuffer(ctx);
/* Clear the LMA depth buffer, if present. */
if ((buffers & BUFFER_BIT_DEPTH) && ctx->Depth.Mask &&
nfb->lma_bo) {
if ((buffers & BUFFER_BIT_DEPTH) &&
ctx->Depth.Mask && nfb->hierz.bo) {
struct nouveau_surface *s = &to_nouveau_renderbuffer(
nfb->base._DepthBuffer->Wrapped)->surface;
/* Clear the hierarchical depth buffer */
BEGIN_RING(chan, celsius, NV17TCL_LMA_DEPTH_FILL_VALUE, 1);
OUT_RING(chan, pack_zs_f(s->format, ctx->Depth.Clear, 0));
BEGIN_RING(chan, celsius, NV17TCL_LMA_DEPTH_BUFFER_CLEAR, 1);
OUT_RING(chan, 1);
/* Mark the depth buffer as cleared */
if (use_fast_zclear(ctx, buffers)) {
if (nctx->hierz.clear_seq)
buffers &= ~BUFFER_BIT_DEPTH;
nfb->hierz.clear_value =
pack_zs_f(s->format, ctx->Depth.Clear, 0);
nctx->hierz.clear_seq++;
context_dirty(ctx, ZCLEAR);
}
}
nouveau_clear(ctx, buffers);
@ -423,7 +457,8 @@ const struct nouveau_driver nv10_driver = {
nv10_emit_tex_obj,
nouveau_emit_nothing,
nouveau_emit_nothing,
nv10_emit_viewport
nv10_emit_viewport,
nv10_emit_zclear
},
.num_emit = NUM_NOUVEAU_STATE,
.num_emit = NUM_NV10_STATE,
};

View file

@ -27,6 +27,11 @@
#ifndef __NV10_DRIVER_H__
#define __NV10_DRIVER_H__
enum {
NOUVEAU_STATE_ZCLEAR = NUM_NOUVEAU_STATE,
NUM_NV10_STATE
};
#define NV10_TEXTURE_UNITS 2
/* nv10_context.c */
@ -52,6 +57,9 @@ nv10_emit_scissor(GLcontext *ctx, int emit);
void
nv10_emit_viewport(GLcontext *ctx, int emit);
void
nv10_emit_zclear(GLcontext *ctx, int emit);
/* nv10_state_polygon.c */
void
nv10_emit_cull_face(GLcontext *ctx, int emit);

View file

@ -62,14 +62,14 @@ setup_lma_buffer(GLcontext *ctx)
height = align(fb->Height, 2),
size = pitch * height;
if (!nfb->lma_bo || nfb->lma_bo->size != size) {
nouveau_bo_ref(NULL, &nfb->lma_bo);
if (!nfb->hierz.bo || nfb->hierz.bo->size != size) {
nouveau_bo_ref(NULL, &nfb->hierz.bo);
nouveau_bo_new(context_dev(ctx), NOUVEAU_BO_VRAM, 0, size,
&nfb->lma_bo);
&nfb->hierz.bo);
}
nouveau_bo_markl(bctx, celsius, NV17TCL_LMA_DEPTH_BUFFER_OFFSET,
nfb->lma_bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
nfb->hierz.bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
WAIT_RING(chan, 9);
BEGIN_RING(chan, celsius, NV17TCL_LMA_DEPTH_WINDOW_X, 4);
@ -134,8 +134,10 @@ nv10_emit_framebuffer(GLcontext *ctx, int emit)
nouveau_bo_markl(bctx, celsius, NV10TCL_ZETA_OFFSET,
s->bo, 0, bo_flags);
if (context_chipset(ctx) >= 0x17)
if (context_chipset(ctx) >= 0x17) {
setup_lma_buffer(ctx);
context_dirty(ctx, ZCLEAR);
}
}
BEGIN_RING(chan, celsius, NV10TCL_RT_FORMAT, 2);
@ -187,3 +189,20 @@ nv10_emit_viewport(GLcontext *ctx, int emit)
context_dirty(ctx, PROJECTION);
}
void
nv10_emit_zclear(GLcontext *ctx, int emit)
{
struct nouveau_context *nctx = to_nouveau_context(ctx);
struct nouveau_channel *chan = context_chan(ctx);
struct nouveau_grobj *celsius = context_eng3d(ctx);
struct nouveau_framebuffer *nfb =
to_nouveau_framebuffer(ctx->DrawBuffer);
if (nfb->hierz.bo) {
BEGIN_RING(chan, celsius, NV17TCL_ZCLEAR_ENABLE, 2);
OUT_RING(chan, nctx->hierz.clear_blocked ? 0 : 1);
OUT_RING(chan, nfb->hierz.clear_value |
(nctx->hierz.clear_seq & 0xff));
}
}