X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fcell%2Fppu%2Fcell_state_emit.c;h=0a0af81f53fc478ad87191a773dde2c82722f97b;hb=7ba68693087618ababfe2c6f260330ec657c5534;hp=2da3097983c113c7a732f3e36cb006b40b5202bd;hpb=176c454765b88c71d8b1ef474bc0fd53cb253a08;p=mesa.git

diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c
index 2da3097983c..0a0af81f53f 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_emit.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c
@@ -25,18 +25,155 @@
  * 
  **************************************************************************/
 
+#include "pipe/p_inlines.h"
 #include "util/u_memory.h"
 #include "cell_context.h"
 #include "cell_gen_fragment.h"
 #include "cell_state.h"
 #include "cell_state_emit.h"
-#include "cell_state_per_fragment.h"
 #include "cell_batch.h"
 #include "cell_texture.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 
 
+/**
+ * Find/create a cell_command_fragment_ops object corresponding to the
+ * current blend/stencil/z/colormask/etc. state.
+ */
+static struct cell_command_fragment_ops *
+lookup_fragment_ops(struct cell_context *cell)
+{
+   struct cell_fragment_ops_key key;
+   struct cell_command_fragment_ops *ops;
+
+   /*
+    * Build key
+    */
+   memset(&key, 0, sizeof(key));
+   key.blend = *cell->blend;
+   key.blend_color = cell->blend_color;
+   key.dsa = *cell->depth_stencil;
+
+   if (cell->framebuffer.cbufs[0])
+      key.color_format = cell->framebuffer.cbufs[0]->format;
+   else
+      key.color_format = PIPE_FORMAT_NONE;
+
+   if (cell->framebuffer.zsbuf)
+      key.zs_format = cell->framebuffer.zsbuf->format;
+   else
+      key.zs_format = PIPE_FORMAT_NONE;
+
+   /*
+    * Look up key in cache.
+    */
+   ops = (struct cell_command_fragment_ops *)
+      util_keymap_lookup(cell->fragment_ops_cache, &key);
+
+   /*
+    * If not found, create/save new fragment ops command.
+    */
+   if (!ops) {
+      struct spe_function spe_code_front, spe_code_back;
+      unsigned int facing_dependent, total_code_size;
+
+      if (0)
+         debug_printf("**** Create New Fragment Ops\n");
+
+      /* Prepare the buffer that will hold the generated code.  The
+       * "0" passed in for the size means that the SPE code will
+       * use a default size.
+       */
+      spe_init_func(&spe_code_front, 0);
+      spe_init_func(&spe_code_back, 0);
+
+      /* Generate new code.  Always generate new code for both front-facing
+       * and back-facing fragments, even if it's the same code in both
+       * cases.
+       */
+      cell_gen_fragment_function(cell, CELL_FACING_FRONT, &spe_code_front);
+      cell_gen_fragment_function(cell, CELL_FACING_BACK, &spe_code_back);
+
+      /* Make sure the code is a multiple of 8 bytes long; this is
+       * required to ensure that the dual pipe instruction alignment
+       * is correct.  It's also important for the SPU unpacking,
+       * which assumes 8-byte boundaries.
+       */
+      unsigned int front_code_size = spe_code_size(&spe_code_front);
+      while (front_code_size % 8 != 0) {
+         spe_lnop(&spe_code_front);
+         front_code_size = spe_code_size(&spe_code_front);
+      }
+      unsigned int back_code_size = spe_code_size(&spe_code_back);
+      while (back_code_size % 8 != 0) {
+         spe_lnop(&spe_code_back);
+         back_code_size = spe_code_size(&spe_code_back);
+      }
+
+      /* Determine whether the code we generated is facing-dependent, by
+       * determining whether the generated code is different for the front-
+       * and back-facing fragments.
+       */
+      if (front_code_size == back_code_size && memcmp(spe_code_front.store, spe_code_back.store, front_code_size) == 0) {
+         /* Code is identical; only need one copy. */
+         facing_dependent = 0;
+         total_code_size = front_code_size;
+      }
+      else {
+         /* Code is different for front-facing and back-facing fragments.
+          * Need to send both copies.
+          */
+         facing_dependent = 1;
+         total_code_size = front_code_size + back_code_size;
+      }
+
+      /* alloc new fragment ops command.  Note that this structure
+       * has variant length based on the total code size required.
+       */
+      ops = CALLOC_VARIANT_LENGTH_STRUCT(cell_command_fragment_ops, total_code_size);
+      /* populate the new cell_command_fragment_ops object */
+      ops->opcode = CELL_CMD_STATE_FRAGMENT_OPS;
+      ops->total_code_size = total_code_size;
+      ops->front_code_index = 0;
+      memcpy(ops->code, spe_code_front.store, front_code_size);
+      if (facing_dependent) {
+        /* We have separate front- and back-facing code.  Append the
+         * back-facing code to the buffer.  Be careful because the code
+         * size is in bytes, but the buffer is of unsigned elements.
+         */
+        ops->back_code_index = front_code_size / sizeof(spe_code_front.store[0]);
+        memcpy(ops->code + ops->back_code_index, spe_code_back.store, back_code_size);
+      }
+      else {
+        /* Use the same code for front- and back-facing fragments */
+        ops->back_code_index = ops->front_code_index;
+      }
+
+      /* Set the fields for the fallback case.  Note that these fields
+       * (and the whole fallback case) will eventually go away.
+       */
+      ops->dsa = *cell->depth_stencil;
+      ops->blend = *cell->blend;
+      ops->blend_color = cell->blend_color;
+
+      /* insert cell_command_fragment_ops object into keymap/cache */
+      util_keymap_insert(cell->fragment_ops_cache, &key, ops, NULL);
+
+      /* release rtasm buffer */
+      spe_release_func(&spe_code_front);
+      spe_release_func(&spe_code_back);
+   }
+   else {
+      if (0)
+         debug_printf("**** Re-use Fragment Ops\n");
+   }
+
+   return ops;
+}
+
+
+
 static void
 emit_state_cmd(struct cell_context *cell, uint cmd,
                const void *state, uint state_size)
@@ -73,6 +210,13 @@ cell_emit_state(struct cell_context *cell)
 #endif
    }
 
+   if (cell->dirty & (CELL_NEW_RASTERIZER)) {
+      struct cell_command_rasterizer *rast =
+         cell_batch_alloc(cell, sizeof(*rast));
+      rast->opcode = CELL_CMD_STATE_RASTERIZER;
+      rast->rasterizer = *cell->rasterizer;
+   }
+
    if (cell->dirty & (CELL_NEW_FS)) {
       /* Send new fragment program to SPUs */
       struct cell_command_fragment_program *fp
@@ -90,59 +234,81 @@ cell_emit_state(struct cell_context *cell)
       }
    }
 
+   if (cell->dirty & (CELL_NEW_FS_CONSTANTS)) {
+      const uint shader = PIPE_SHADER_FRAGMENT;
+      const uint num_const = cell->constants[shader].size / sizeof(float);
+      uint i, j;
+      float *buf = cell_batch_alloc(cell, 16 + num_const * sizeof(float));
+      uint64_t *ibuf = (uint64_t *) buf;
+      const float *constants = pipe_buffer_map(cell->pipe.screen,
+                                               cell->constants[shader].buffer,
+                                               PIPE_BUFFER_USAGE_CPU_READ);
+      ibuf[0] = CELL_CMD_STATE_FS_CONSTANTS;
+      ibuf[1] = num_const;
+      j = 4;
+      for (i = 0; i < num_const; i++) {
+         buf[j++] = constants[i];
+      }
+      pipe_buffer_unmap(cell->pipe.screen, cell->constants[shader].buffer);
+   }
+
    if (cell->dirty & (CELL_NEW_FRAMEBUFFER |
                       CELL_NEW_DEPTH_STENCIL |
                       CELL_NEW_BLEND)) {
-      /* XXX we don't want to always do codegen here.  We should have
-       * a hash/lookup table to cache previous results...
-       */
-      struct cell_command_fragment_ops *fops
-            = cell_batch_alloc(cell, sizeof(*fops));
-      struct spe_function spe_code;
-
-      /* generate new code */
-      cell_gen_fragment_function(cell, &spe_code);
-      /* put the new code into the batch buffer */
-      fops->opcode = CELL_CMD_STATE_FRAGMENT_OPS;
-      memcpy(&fops->code, spe_code.store,
-             SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE);
-      fops->dsa = cell->depth_stencil->base;
-      fops->blend = cell->blend->base;
-      /* free codegen buffer */
-      spe_release_func(&spe_code);
+      struct cell_command_fragment_ops *fops, *fops_cmd;
+      /* Note that cell_command_fragment_ops is a variant-sized record */
+      fops = lookup_fragment_ops(cell);
+      fops_cmd = cell_batch_alloc(cell, sizeof(*fops_cmd) + fops->total_code_size);
+      memcpy(fops_cmd, fops, sizeof(*fops) + fops->total_code_size);
    }
 
    if (cell->dirty & CELL_NEW_SAMPLER) {
       uint i;
       for (i = 0; i < CELL_MAX_SAMPLERS; i++) {
-         if (cell->sampler[i]) {
-            struct cell_command_sampler *sampler
-               = cell_batch_alloc(cell, sizeof(*sampler));
-            sampler->opcode = CELL_CMD_STATE_SAMPLER;
-            sampler->unit = i;
-            sampler->state = *cell->sampler[i];
+         if (cell->dirty_samplers & (1 << i)) {
+            if (cell->sampler[i]) {
+               struct cell_command_sampler *sampler
+                  = cell_batch_alloc(cell, sizeof(*sampler));
+               sampler->opcode = CELL_CMD_STATE_SAMPLER;
+               sampler->unit = i;
+               sampler->state = *cell->sampler[i];
+            }
          }
       }
+      cell->dirty_samplers = 0x0;
    }
 
    if (cell->dirty & CELL_NEW_TEXTURE) {
       uint i;
       for (i = 0;i < CELL_MAX_SAMPLERS; i++) {
-         struct cell_command_texture *texture
-            =  cell_batch_alloc(cell, sizeof(*texture));
-         texture->opcode = CELL_CMD_STATE_TEXTURE;
-         texture->unit = i;
-         if (cell->texture[i]) {
-            texture->start = cell->texture[i]->tiled_data;
-            texture->width = cell->texture[i]->base.width[0];
-            texture->height = cell->texture[i]->base.height[0];
-         }
-         else {
-            texture->start = NULL;
-            texture->width = 1;
-            texture->height = 1;
+         if (cell->dirty_textures & (1 << i)) {
+            struct cell_command_texture *texture
+               =  cell_batch_alloc(cell, sizeof(*texture));
+            texture->opcode = CELL_CMD_STATE_TEXTURE;
+            texture->unit = i;
+            if (cell->texture[i]) {
+               uint level;
+               for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
+                  texture->start[level] = cell->texture[i]->tiled_mapped[level];
+                  texture->width[level] = cell->texture[i]->base.width[level];
+                  texture->height[level] = cell->texture[i]->base.height[level];
+                  texture->depth[level] = cell->texture[i]->base.depth[level];
+               }
+               texture->target = cell->texture[i]->base.target;
+            }
+            else {
+               uint level;
+               for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
+                  texture->start[level] = NULL;
+                  texture->width[level] = 0;
+                  texture->height[level] = 0;
+                  texture->depth[level] = 0;
+               }
+               texture->target = 0;
+            }
          }
       }
+      cell->dirty_textures = 0x0;
    }
 
    if (cell->dirty & CELL_NEW_VERTEX_INFO) {