r300g: implement hyper-z support. (v4)

This implements fast Z clear, Z compression, and HiZ support for r300->r500
GPUs.

It also allows cbzb clears when fast Z clears are being used for the ZB.

It requires a kernel with hyper-z support.

Thanks to Marek Olšák <maraeo@gmail.com>, who started this off, and Alex Deucher at AMD for providing lots of hints.

v2:
squashed zmask ram size fix]
squashed r300g/blitter: fix Z readback when compressed]

v3:
rebase around texture changes in master - .1 fix more bits

v4:
migrated to using u_mm in r300_texture to manage hiz/zmask rams consistently
disabled HiZ when using OQ
flush z-cache before turning hyper-z off
update hyper-z state on dsa state change
store depthclearvalue across cbzb clears and replace it afterwards.

Signed-off-by: Dave Airlie <airlied@redhat.com>

1 files changed, 235 insertions, 2 deletions

diff --git a/src/gallium/drivers/r300/r300_hyperz.c b/src/gallium/drivers/r300/r300_hyperz.c
index e952895601..e719342a46 100644
--- a/src/gallium/drivers/r300/r300_hyperz.c
+++ b/src/gallium/drivers/r300/r300_hyperz.c
@@ -21,25 +21,158 @@
  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
+#include "util/u_format.h"
+#include "util/u_mm.h"
 #include "r300_context.h"
 #include "r300_hyperz.h"
 #include "r300_reg.h"
 #include "r300_fs.h"
+#include "r300_emit.h"
+#include "r300_texture.h"
 
+/*
+  HiZ rules - taken from various docs 
+   1. HiZ only works on depth values
+   2. Cannot HiZ if stencil fail or zfail is !KEEP
+   3. on R300/400, HiZ is disabled if depth test is EQUAL
+   4. comparison changes without clears usually mean disabling HiZ
+*/
 /*****************************************************************************/
 /* The HyperZ setup                                                          */
 /*****************************************************************************/
 
+static bool r300_get_sc_hz_max(struct r300_context *r300)
+{
+    struct r300_dsa_state *dsa_state = r300->dsa_state.state;
+    int func = dsa_state->z_stencil_control & 0x7;
+    int ret = R300_SC_HYPERZ_MIN;
+
+    if (func >= 4 && func <= 7)
+       ret = R300_SC_HYPERZ_MAX;
+    return ret;
+}
+
+static bool r300_zfunc_same_direction(int func1, int func2)
+{
+    /* func1 is less/lessthan */
+    if (func1 == 1 || func1 == 2)
+        if (func2 == 3 || func2 == 4 || func2 == 5)
+            return FALSE;
+
+    if (func2 == 1 || func2 == 2)
+        if (func1 == 4 || func1 == 5)
+            return FALSE;
+    return TRUE;
+}
+    
+static int r300_get_hiz_min(struct r300_context *r300)
+{
+    struct r300_dsa_state *dsa_state = r300->dsa_state.state;
+    int func = dsa_state->z_stencil_control & 0x7;
+    int ret = R300_HIZ_MIN;
+
+    if (func == 1 || func == 2)
+       ret = R300_HIZ_MAX;
+    return ret;
+}
+
+static boolean r300_dsa_stencil_op_not_keep(struct pipe_stencil_state *s)
+{
+    if (s->enabled && (s->fail_op != PIPE_STENCIL_OP_KEEP ||
+                       s->zfail_op != PIPE_STENCIL_OP_KEEP))
+        return TRUE;
+    return FALSE;
+}
+
+static boolean r300_can_hiz(struct r300_context *r300)
+{
+    struct r300_dsa_state *dsa_state = r300->dsa_state.state;
+    struct pipe_depth_stencil_alpha_state *dsa = &dsa_state->dsa;
+    struct r300_screen* r300screen = r300->screen;
+    struct r300_hyperz_state *z = r300->hyperz_state.state;
+
+    /* shader writes depth - no HiZ */
+    if (r300_fragment_shader_writes_depth(r300_fs(r300))) /* (5) */
+        return FALSE;
+
+    if (r300->query_current)
+        return FALSE;
+    /* if stencil fail/zfail op is not KEEP */
+    if (r300_dsa_stencil_op_not_keep(&dsa->stencil[0]) ||
+        r300_dsa_stencil_op_not_keep(&dsa->stencil[1]))
+        return FALSE;
+
+    if (dsa->depth.enabled) {
+        /* if depth func is EQUAL pre-r500 */
+        if (dsa->depth.func == PIPE_FUNC_EQUAL && !r300screen->caps.is_r500)
+            return FALSE;
+        /* if depth func is NOTEQUAL */
+        if (dsa->depth.func == PIPE_FUNC_NOTEQUAL)
+            return FALSE;
+    }
+    /* depth comparison function - if just cleared save and return okay */
+    if (z->current_func == -1) {
+        int func = dsa_state->z_stencil_control & 0x7;
+        if (func != 0 && func != 7)
+            z->current_func = dsa_state->z_stencil_control & 0x7;
+    } else {
+        /* simple don't change */
+        if (!r300_zfunc_same_direction(z->current_func, (dsa_state->z_stencil_control & 0x7))) {
+            DBG(r300, DBG_HYPERZ, "z func changed direction - disabling hyper-z %d -> %d\n", z->current_func, dsa_state->z_stencil_control);
+            return FALSE;
+        }
+    }    
+    return TRUE;
+}
+
 static void r300_update_hyperz(struct r300_context* r300)
 {
     struct r300_hyperz_state *z =
         (struct r300_hyperz_state*)r300->hyperz_state.state;
 
+    z->gb_z_peq_config = 0;
     z->zb_bw_cntl = 0;
     z->sc_hyperz = R300_SC_HYPERZ_ADJ_2;
+    z->flush = 0;
 
-    if (r300->cbzb_clear)
+    if (r300->cbzb_clear) {
         z->zb_bw_cntl |= R300_ZB_CB_CLEAR_CACHE_LINE_WRITE_ONLY;
+        return;
+    }
+
+    /* Zbuffer compression. */
+    if (r300->z_compression) {
+        z->zb_bw_cntl |= R300_RD_COMP_ENABLE;
+        if (r300->z_decomp_rd == false)
+            z->zb_bw_cntl |= R300_WR_COMP_ENABLE;
+       /* RV350 and up optimizations. */
+       if (r300->z_compression == RV350_Z_COMPRESS_88)
+           z->gb_z_peq_config |= R300_GB_Z_PEQ_CONFIG_Z_PEQ_SIZE_8_8;
+    }
+
+    /* Z fastfill. */
+    if (r300->z_fastfill) {
+        z->zb_bw_cntl |= R300_FAST_FILL_ENABLE; /*  | R300_FORCE_COMPRESSED_STENCIL_VALUE_ENABLE;*/
+    }
+
+    if (r300->hiz_enable) {
+        bool can_hiz = r300_can_hiz(r300);
+        if (can_hiz) {
+            z->zb_bw_cntl |= R300_HIZ_ENABLE;
+            z->sc_hyperz |= R300_SC_HYPERZ_ENABLE;
+            z->sc_hyperz |= r300_get_sc_hz_max(r300);
+            z->zb_bw_cntl |= r300_get_hiz_min(r300);
+        }
+    }
+
+    if (r300->screen->caps.is_r500) {
+        /* XXX Are these bits really available on RV350? */
+        z->zb_bw_cntl |= R500_HIZ_FP_EXP_BITS_3;
+        z->zb_bw_cntl |=
+                R500_HIZ_EQUAL_REJECT_ENABLE |
+                R500_PEQ_PACKING_ENABLE |
+                R500_COVERED_PTR_MASKING_ENABLE;
+    }
 }
 
 /*****************************************************************************/
@@ -126,15 +259,115 @@ static void r300_update_ztop(struct r300_context* r300)
     } else {
         ztop_state->z_buffer_top = R300_ZTOP_ENABLE;
     }
-
     if (ztop_state->z_buffer_top != old_ztop)
         r300->ztop_state.dirty = TRUE;
 }
 
+#define ALIGN_DIVUP(x, y) (((x) + (y) - 1) / (y))
+
+static void r300_update_hiz_clear(struct r300_context *r300)
+{
+    struct pipe_framebuffer_state *fb =
+        (struct pipe_framebuffer_state*)r300->fb_state.state;
+    uint32_t height;
+
+    height = ALIGN_DIVUP(fb->zsbuf->height, 4);
+    r300->hiz_clear.size = height * 4;
+}
+
+static void r300_update_zmask_clear(struct r300_context *r300)
+{
+    struct pipe_framebuffer_state *fb =
+        (struct pipe_framebuffer_state*)r300->fb_state.state;
+    uint32_t height;
+    int mult;
+
+    if (r300->z_compression == RV350_Z_COMPRESS_88)
+        mult = 8;
+    else
+        mult = 4;
+
+    height = ALIGN_DIVUP(fb->zsbuf->height, mult);
+
+    r300->zmask_clear.size = height * 4;
+}
+
 void r300_update_hyperz_state(struct r300_context* r300)
 {
     r300_update_ztop(r300);
     if (r300->hyperz_state.dirty) {
         r300_update_hyperz(r300);
     }
+
+    if (r300->hiz_clear.dirty) {
+       r300_update_hiz_clear(r300);
+    }
+    if (r300->zmask_clear.dirty) {
+       r300_update_zmask_clear(r300);
+    }
+}
+
+void r300_hiz_alloc_block(struct r300_context *r300, struct r300_surface *surf)
+{
+    struct r300_texture *tex;
+    uint32_t zsize, ndw;
+    int level = surf->base.level;
+
+    tex = r300_texture(surf->base.texture);
+
+    if (tex->hiz_mem[level])
+        return;
+
+    zsize = tex->desc.layer_size_in_bytes[level];
+    zsize /= util_format_get_blocksize(tex->desc.b.b.format);
+    ndw = ALIGN_DIVUP(zsize, 64);
+
+    tex->hiz_mem[level] = u_mmAllocMem(r300->hiz_mm, ndw, 0, 0);
+    return;
+}
+
+void r300_zmask_alloc_block(struct r300_context *r300, struct r300_surface *surf, int compress)
+{
+    int bsize = 256;
+    uint32_t zsize, ndw;
+    int level = surf->base.level;
+    struct r300_texture *tex;
+
+    tex = r300_texture(surf->base.texture);
+
+    if (tex->zmask_mem[level])
+        return;
+
+    zsize = tex->desc.layer_size_in_bytes[level];
+    zsize /= util_format_get_blocksize(tex->desc.b.b.format);
+
+    /* each zmask dword represents 16 4x4 blocks - which is 256 pixels
+       or 16 8x8 depending on the gb peq flag = 1024 pixels */
+    if (compress == RV350_Z_COMPRESS_88)
+        bsize = 1024;
+
+    ndw = ALIGN_DIVUP(zsize, bsize);
+    tex->zmask_mem[level] = u_mmAllocMem(r300->zmask_mm, ndw, 0, 0);
+    return;
+}
+
+void r300_hyperz_init_mm(struct r300_context *r300)
+{
+    struct r300_screen* r300screen = r300->screen;
+    int frag_pipes = r300screen->caps.num_frag_pipes;
+
+    if (r300screen->caps.hiz_ram)
+      r300->hiz_mm = u_mmInit(0, r300screen->caps.hiz_ram * frag_pipes);
+
+    r300->zmask_mm = u_mmInit(0, r300screen->caps.zmask_ram * frag_pipes);
+}
+
+void r300_hyperz_destroy_mm(struct r300_context *r300)
+{
+    struct r300_screen* r300screen = r300->screen;
+
+    if (r300screen->caps.hiz_ram)
+      u_mmDestroy(r300->hiz_mm);
+
+    u_mmDestroy(r300->zmask_mm);
 }