drivers/svga/svga_shader.c

/**********************************************************
 * Copyright 2008-2012 VMware, Inc.  All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person
 * obtaining a copy of this software and associated documentation
 * files (the "Software"), to deal in the Software without
 * restriction, including without limitation the rights to use, copy,
 * modify, merge, publish, distribute, sublicense, and/or sell copies
 * of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be
 * included in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 **********************************************************/

#include "util/u_bitmask.h"
#include "util/u_memory.h"
#include "util/u_format.h"
#include "svga_context.h"
#include "svga_cmd.h"
#include "svga_format.h"
#include "svga_shader.h"
#include "svga_resource_texture.h"


/**
 * This bit isn't really used anywhere.  It only serves to help
 * generate a unique "signature" for the vertex shader output bitmask.
 * Shader input/output signatures are used to resolve shader linking
 * issues.
 */
#define FOG_GENERIC_BIT (((uint64_t) 1) << 63)


/**
 * Use the shader info to generate a bitmask indicating which generic
 * inputs are used by the shader.  A set bit indicates that GENERIC[i]
 * is used.
 */
uint64_t
svga_get_generic_inputs_mask(const struct tgsi_shader_info *info)
{
   unsigned i;
   uint64_t mask = 0x0;

   for (i = 0; i < info->num_inputs; i++) {
      if (info->input_semantic_name[i] == TGSI_SEMANTIC_GENERIC) {
         unsigned j = info->input_semantic_index[i];
         assert(j < sizeof(mask) * 8);
         mask |= ((uint64_t) 1) << j;
      }
   }

   return mask;
}


/**
 * Scan shader info to return a bitmask of written outputs.
 */
uint64_t
svga_get_generic_outputs_mask(const struct tgsi_shader_info *info)
{
   unsigned i;
   uint64_t mask = 0x0;

   for (i = 0; i < info->num_outputs; i++) {
      switch (info->output_semantic_name[i]) {
      case TGSI_SEMANTIC_GENERIC:
         {
            unsigned j = info->output_semantic_index[i];
            assert(j < sizeof(mask) * 8);
            mask |= ((uint64_t) 1) << j;
         }
         break;
      case TGSI_SEMANTIC_FOG:
         mask |= FOG_GENERIC_BIT;
         break;
      }
   }

   return mask;
}


/**
 * Given a mask of used generic variables (as returned by the above functions)
 * fill in a table which maps those indexes to small integers.
 * This table is used by the remap_generic_index() function in
 * svga_tgsi_decl_sm30.c
 * Example: if generics_mask = binary(1010) it means that GENERIC[1] and
 * GENERIC[3] are used.  The remap_table will contain:
 *   table[1] = 0;
 *   table[3] = 1;
 * The remaining table entries will be filled in with the next unused
 * generic index (in this example, 2).
 */
void
svga_remap_generics(uint64_t generics_mask,
                    int8_t remap_table[MAX_GENERIC_VARYING])
{
   /* Note texcoord[0] is reserved so start at 1 */
   unsigned count = 1, i;

   for (i = 0; i < MAX_GENERIC_VARYING; i++) {
      remap_table[i] = -1;
   }

   /* for each bit set in generic_mask */
   while (generics_mask) {
      unsigned index = ffsll(generics_mask) - 1;
      remap_table[index] = count++;
      generics_mask &= ~((uint64_t) 1 << index);
   }
}


/**
 * Use the generic remap table to map a TGSI generic varying variable
 * index to a small integer.  If the remapping table doesn't have a
 * valid value for the given index (the table entry is -1) it means
 * the fragment shader doesn't use that VS output.  Just allocate
 * the next free value in that case.  Alternately, we could cull
 * VS instructions that write to register, or replace the register
 * with a dummy temp register.
 * XXX TODO: we should do one of the later as it would save precious
 * texcoord registers.
 */
int
svga_remap_generic_index(int8_t remap_table[MAX_GENERIC_VARYING],
                         int generic_index)
{
   assert(generic_index < MAX_GENERIC_VARYING);

   if (generic_index >= MAX_GENERIC_VARYING) {
      /* just don't return a random/garbage value */
      generic_index = MAX_GENERIC_VARYING - 1;
   }

   if (remap_table[generic_index] == -1) {
      /* This is a VS output that has no matching PS input.  Find a
       * free index.
       */
      int i, max = 0;
      for (i = 0; i < MAX_GENERIC_VARYING; i++) {
         max = MAX2(max, remap_table[i]);
      }
      remap_table[generic_index] = max + 1;
   }

   return remap_table[generic_index];
}

static const enum pipe_swizzle copy_alpha[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_Y,
   PIPE_SWIZZLE_Z,
   PIPE_SWIZZLE_W,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};

static const enum pipe_swizzle set_alpha[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_Y,
   PIPE_SWIZZLE_Z,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};

static const enum pipe_swizzle set_000X[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};

static const enum pipe_swizzle set_XXXX[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};

static const enum pipe_swizzle set_XXX1[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};

static const enum pipe_swizzle set_XXXY[PIPE_SWIZZLE_MAX] = {
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_X,
   PIPE_SWIZZLE_Y,
   PIPE_SWIZZLE_0,
   PIPE_SWIZZLE_1,
   PIPE_SWIZZLE_NONE
};


/**
 * Initialize the shader-neutral fields of svga_compile_key from context
 * state.  This is basically the texture-related state.
 */
void
svga_init_shader_key_common(const struct svga_context *svga,
                            enum pipe_shader_type shader,
                            struct svga_compile_key *key)
{
   unsigned i, idx = 0;

   assert(shader < ARRAY_SIZE(svga->curr.num_sampler_views));

   /* In case the number of samplers and sampler_views doesn't match,
    * loop over the lower of the two counts.
    */
   key->num_textures = MAX2(svga->curr.num_sampler_views[shader],
                            svga->curr.num_samplers[shader]);

   for (i = 0; i < key->num_textures; i++) {
      struct pipe_sampler_view *view = svga->curr.sampler_views[shader][i];
      const struct svga_sampler_state *sampler = svga->curr.sampler[shader][i];
      if (view) {
         assert(view->texture);
         assert(view->texture->target < (1 << 4)); /* texture_target:4 */

         /* 1D/2D array textures with one slice and cube map array textures
          * with one cube are treated as non-arrays by the SVGA3D device.
          * Set the is_array flag only if we know that we have more than 1
          * element.  This will be used to select shader instruction/resource
          * types during shader translation.
          */
         switch (view->texture->target) {
         case PIPE_TEXTURE_1D_ARRAY:
         case PIPE_TEXTURE_2D_ARRAY:
            key->tex[i].is_array = view->texture->array_size > 1;
            break;
         case PIPE_TEXTURE_CUBE_ARRAY:
            key->tex[i].is_array = view->texture->array_size > 6;
            break;
         default:
            ; /* nothing / silence compiler warning */
         }

         assert(view->texture->nr_samples < (1 << 5)); /* 5-bit field */
         key->tex[i].num_samples = view->texture->nr_samples;

         const enum pipe_swizzle *swizzle_tab;
         if (view->texture->target == PIPE_BUFFER) {
            SVGA3dSurfaceFormat svga_format;
            unsigned tf_flags;

            /* Apply any special swizzle mask for the view format if needed */

            svga_translate_texture_buffer_view_format(view->format,
                                                      &svga_format, &tf_flags);
            if (tf_flags & TF_000X)
               swizzle_tab = set_000X;
            else if (tf_flags & TF_XXXX)
               swizzle_tab = set_XXXX;
            else if (tf_flags & TF_XXX1)
               swizzle_tab = set_XXX1;
            else if (tf_flags & TF_XXXY)
               swizzle_tab = set_XXXY;
            else
               swizzle_tab = copy_alpha;
         }
         else {
            /* If we have a non-alpha view into an svga3d surface with an
             * alpha channel, then explicitly set the alpha channel to 1
             * when sampling. Note that we need to check the
             * actual device format to cover also imported surface cases.
             */
            swizzle_tab =
               (!util_format_has_alpha(view->format) &&
                svga_texture_device_format_has_alpha(view->texture)) ?
                set_alpha : copy_alpha;

            if (view->texture->format == PIPE_FORMAT_DXT1_RGB ||
                view->texture->format == PIPE_FORMAT_DXT1_SRGB)
               swizzle_tab = set_alpha;
         }

         key->tex[i].swizzle_r = swizzle_tab[view->swizzle_r];
         key->tex[i].swizzle_g = swizzle_tab[view->swizzle_g];
         key->tex[i].swizzle_b = swizzle_tab[view->swizzle_b];
         key->tex[i].swizzle_a = swizzle_tab[view->swizzle_a];
      }

      if (sampler) {
         if (!sampler->normalized_coords) {
            assert(idx < (1 << 5));  /* width_height_idx:5 bitfield */
            key->tex[i].width_height_idx = idx++;
            key->tex[i].unnormalized = TRUE;
            ++key->num_unnormalized_coords;

            if (sampler->magfilter == SVGA3D_TEX_FILTER_NEAREST ||
                sampler->minfilter == SVGA3D_TEX_FILTER_NEAREST) {
                key->tex[i].texel_bias = TRUE;
            }
         }
      }
   }
}


/** Search for a compiled shader variant with the same compile key */
struct svga_shader_variant *
svga_search_shader_key(const struct svga_shader *shader,
                       const struct svga_compile_key *key)
{
   struct svga_shader_variant *variant = shader->variants;

   assert(key);

   for ( ; variant; variant = variant->next) {
      if (svga_compile_keys_equal(key, &variant->key))
         return variant;
   }
   return NULL;
}

/** Search for a shader with the same token key */
struct svga_shader *
svga_search_shader_token_key(struct svga_shader *pshader,
                             const struct svga_token_key *key)
{
   struct svga_shader *shader = pshader;

   assert(key);

   for ( ; shader; shader = shader->next) {
      if (memcmp(key, &shader->token_key, sizeof(struct svga_token_key)) == 0)
         return shader;
   }
   return NULL;
}

/**
 * Helper function to define a gb shader for non-vgpu10 device
 */
static enum pipe_error
define_gb_shader_vgpu9(struct svga_context *svga,
                       SVGA3dShaderType type,
                       struct svga_shader_variant *variant,
                       unsigned codeLen)
{
   struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
   enum pipe_error ret;

   /**
    * Create gb memory for the shader and upload the shader code.
    * Kernel module will allocate an id for the shader and issue
    * the DefineGBShader command.
    */
   variant->gb_shader = sws->shader_create(sws, type,
                                           variant->tokens, codeLen);

   if (!variant->gb_shader)
      return PIPE_ERROR_OUT_OF_MEMORY;

   ret = SVGA3D_BindGBShader(svga->swc, variant->gb_shader);

   return ret;
}

/**
 * Helper function to define a gb shader for vgpu10 device
 */
static enum pipe_error
define_gb_shader_vgpu10(struct svga_context *svga,
                        SVGA3dShaderType type,
                        struct svga_shader_variant *variant,
                        unsigned codeLen)
{
   struct svga_winsys_context *swc = svga->swc;
   enum pipe_error ret;

   /**
    * Shaders in VGPU10 enabled device reside in the device COTable.
    * SVGA driver will allocate an integer ID for the shader and
    * issue DXDefineShader and DXBindShader commands.
    */
   variant->id = util_bitmask_add(svga->shader_id_bm);
   if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
      return PIPE_ERROR_OUT_OF_MEMORY;
   }

   /* Create gb memory for the shader and upload the shader code */
   variant->gb_shader = swc->shader_create(swc,
                                           variant->id, type,
                                           variant->tokens, codeLen);

   if (!variant->gb_shader) {
      /* Free the shader ID */
      assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
      goto fail_no_allocation;
   }

   /**
    * Since we don't want to do any flush within state emission to avoid
    * partial state in a command buffer, it's important to make sure that
    * there is enough room to send both the DXDefineShader & DXBindShader
    * commands in the same command buffer. So let's send both
    * commands in one command reservation. If it fails, we'll undo
    * the shader creation and return an error.
    */
   ret = SVGA3D_vgpu10_DefineAndBindShader(swc, variant->gb_shader,
                                           variant->id, type, codeLen);

   if (ret != PIPE_OK)
      goto fail;

   return PIPE_OK;

fail:
   swc->shader_destroy(swc, variant->gb_shader);
   variant->gb_shader = NULL;

fail_no_allocation:
   util_bitmask_clear(svga->shader_id_bm, variant->id);
   variant->id = UTIL_BITMASK_INVALID_INDEX;

   return PIPE_ERROR_OUT_OF_MEMORY;
}

/**
 * Issue the SVGA3D commands to define a new shader.
 * \param variant  contains the shader tokens, etc.  The result->id field will
 *                 be set here.
 */
enum pipe_error
svga_define_shader(struct svga_context *svga,
                   SVGA3dShaderType type,
                   struct svga_shader_variant *variant)
{
   unsigned codeLen = variant->nr_tokens * sizeof(variant->tokens[0]);
   enum pipe_error ret;

   SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_DEFINESHADER);

   variant->id = UTIL_BITMASK_INVALID_INDEX;

   if (svga_have_gb_objects(svga)) {
      if (svga_have_vgpu10(svga))
         ret = define_gb_shader_vgpu10(svga, type, variant, codeLen);
      else
         ret = define_gb_shader_vgpu9(svga, type, variant, codeLen);
   }
   else {
      /* Allocate an integer ID for the shader */
      variant->id = util_bitmask_add(svga->shader_id_bm);
      if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
         ret = PIPE_ERROR_OUT_OF_MEMORY;
         goto done;
      }

      /* Issue SVGA3D device command to define the shader */
      ret = SVGA3D_DefineShader(svga->swc,
                                variant->id,
                                type,
                                variant->tokens,
                                codeLen);
      if (ret != PIPE_OK) {
         /* free the ID */
         assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
         util_bitmask_clear(svga->shader_id_bm, variant->id);
         variant->id = UTIL_BITMASK_INVALID_INDEX;
      }
   }

done:
   SVGA_STATS_TIME_POP(svga_sws(svga));
   return ret;
}


/**
 * Issue the SVGA3D commands to set/bind a shader.
 * \param result  the shader to bind.
 */
enum pipe_error
svga_set_shader(struct svga_context *svga,
                SVGA3dShaderType type,
                struct svga_shader_variant *variant)
{
   enum pipe_error ret;
   unsigned id = variant ? variant->id : SVGA3D_INVALID_ID;

   assert(type == SVGA3D_SHADERTYPE_VS ||
          type == SVGA3D_SHADERTYPE_GS ||
          type == SVGA3D_SHADERTYPE_PS);

   if (svga_have_gb_objects(svga)) {
      struct svga_winsys_gb_shader *gbshader =
         variant ? variant->gb_shader : NULL;

      if (svga_have_vgpu10(svga))
         ret = SVGA3D_vgpu10_SetShader(svga->swc, type, gbshader, id);
      else
         ret = SVGA3D_SetGBShader(svga->swc, type, gbshader);
   }
   else {
      ret = SVGA3D_SetShader(svga->swc, type, id);
   }

   return ret;
}


struct svga_shader_variant *
svga_new_shader_variant(struct svga_context *svga)
{
   svga->hud.num_shaders++;
   return CALLOC_STRUCT(svga_shader_variant);
}


void
svga_destroy_shader_variant(struct svga_context *svga,
                            SVGA3dShaderType type,
                            struct svga_shader_variant *variant)
{
   enum pipe_error ret = PIPE_OK;

   if (svga_have_gb_objects(svga) && variant->gb_shader) {
      if (svga_have_vgpu10(svga)) {
         struct svga_winsys_context *swc = svga->swc;
         swc->shader_destroy(swc, variant->gb_shader);
         ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
         if (ret != PIPE_OK) {
            /* flush and try again */
            svga_context_flush(svga, NULL);
            ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
            assert(ret == PIPE_OK);
         }
         util_bitmask_clear(svga->shader_id_bm, variant->id);
      }
      else {
         struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
         sws->shader_destroy(sws, variant->gb_shader);
      }
      variant->gb_shader = NULL;
   }
   else {
      if (variant->id != UTIL_BITMASK_INVALID_INDEX) {
         ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
         if (ret != PIPE_OK) {
            /* flush and try again */
            svga_context_flush(svga, NULL);
            ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
            assert(ret == PIPE_OK);
         }
         util_bitmask_clear(svga->shader_id_bm, variant->id);
      }
   }

   FREE((unsigned *)variant->tokens);
   FREE(variant);

   svga->hud.num_shaders--;
}

/*
 * Rebind shaders.
 * Called at the beginning of every new command buffer to ensure that
 * shaders are properly paged-in. Instead of sending the SetShader
 * command, this function sends a private allocation command to
 * page in a shader. This avoids emitting redundant state to the device
 * just to page in a resource.
 */
enum pipe_error
svga_rebind_shaders(struct svga_context *svga)
{
   struct svga_winsys_context *swc = svga->swc;
   struct svga_hw_draw_state *hw = &svga->state.hw_draw;
   enum pipe_error ret;

   assert(svga_have_vgpu10(svga));

   /**
    * If the underlying winsys layer does not need resource rebinding,
    * just clear the rebind flags and return.
    */
   if (swc->resource_rebind == NULL) {
      svga->rebind.flags.vs = 0;
      svga->rebind.flags.gs = 0;
      svga->rebind.flags.fs = 0;

      return PIPE_OK;
   }

   if (svga->rebind.flags.vs && hw->vs && hw->vs->gb_shader) {
      ret = swc->resource_rebind(swc, NULL, hw->vs->gb_shader, SVGA_RELOC_READ);
      if (ret != PIPE_OK)
         return ret;
   }
   svga->rebind.flags.vs = 0;

   if (svga->rebind.flags.gs && hw->gs && hw->gs->gb_shader) {
      ret = swc->resource_rebind(swc, NULL, hw->gs->gb_shader, SVGA_RELOC_READ);
      if (ret != PIPE_OK)
         return ret;
   }
   svga->rebind.flags.gs = 0;

   if (svga->rebind.flags.fs && hw->fs && hw->fs->gb_shader) {
      ret = swc->resource_rebind(swc, NULL, hw->fs->gb_shader, SVGA_RELOC_READ);
      if (ret != PIPE_OK)
         return ret;
   }
   svga->rebind.flags.fs = 0;

   return PIPE_OK;
}