llvmpipe: Generate the fragment pipeline into a single function.

Still hackish. Will document and optimize later.

1 files changed, 284 insertions, 110 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 521700acd5..9b0e7cdd37 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -38,9 +38,11 @@
 #include "tgsi/tgsi_parse.h"
 #include "lp_bld_type.h"
 #include "lp_bld_conv.h"
+#include "lp_bld_logic.h"
 #include "lp_bld_depth.h"
 #include "lp_bld_tgsi.h"
 #include "lp_bld_alpha.h"
+#include "lp_bld_blend.h"
 #include "lp_bld_swizzle.h"
 #include "lp_bld_flow.h"
 #include "lp_bld_debug.h"
@@ -55,13 +57,13 @@ static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
 
 
 static void
-setup_pos_vector(LLVMBuilderRef builder,
-                 LLVMValueRef x,
-                 LLVMValueRef y,
-                 LLVMValueRef a0_ptr,
-                 LLVMValueRef dadx_ptr,
-                 LLVMValueRef dady_ptr,
-                 LLVMValueRef *pos)
+generate_pos(LLVMBuilderRef builder,
+             LLVMValueRef x,
+             LLVMValueRef y,
+             LLVMValueRef a0_ptr,
+             LLVMValueRef dadx_ptr,
+             LLVMValueRef dady_ptr,
+             LLVMValueRef *pos)
 {
    LLVMTypeRef int_elem_type = LLVMInt32Type();
    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
@@ -110,13 +112,13 @@ setup_pos_vector(LLVMBuilderRef builder,
 
 
 static void
-depth_test_generate(struct llvmpipe_context *lp,
-                    LLVMBuilderRef builder,
-                    const struct pipe_depth_state *state,
-                    union lp_type src_type,
-                    struct lp_build_mask_context *mask,
-                    LLVMValueRef src,
-                    LLVMValueRef dst_ptr)
+generate_depth(struct llvmpipe_context *lp,
+               LLVMBuilderRef builder,
+               const struct pipe_depth_state *state,
+               union lp_type src_type,
+               struct lp_build_mask_context *mask,
+               LLVMValueRef src,
+               LLVMValueRef dst_ptr)
 {
    const struct util_format_description *format_desc;
    union lp_type dst_type;
@@ -151,18 +153,177 @@ depth_test_generate(struct llvmpipe_context *lp,
 }
 
 
-static struct lp_fragment_shader_variant *
-shader_generate(struct llvmpipe_context *lp,
-                struct lp_fragment_shader *shader,
-                const struct lp_fragment_shader_variant_key *key)
+/**
+ * Generate the fragment shader, depth/stencil and alpha tests.
+ */
+static void
+generate_fs(struct llvmpipe_context *lp,
+            struct lp_fragment_shader *shader,
+            const struct lp_fragment_shader_variant_key *key,
+            LLVMBuilderRef builder,
+            union lp_type type,
+            unsigned i,
+            LLVMValueRef x,
+            LLVMValueRef y,
+            LLVMValueRef a0_ptr,
+            LLVMValueRef dadx_ptr,
+            LLVMValueRef dady_ptr,
+            LLVMValueRef consts_ptr,
+            LLVMValueRef *pmask,
+            LLVMValueRef *color,
+            LLVMValueRef depth_ptr,
+            LLVMValueRef samplers_ptr)
 {
-   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
-   struct lp_fragment_shader_variant *variant;
    const struct tgsi_token *tokens = shader->base.tokens;
-   union lp_type type;
    LLVMTypeRef elem_type;
    LLVMTypeRef vec_type;
    LLVMTypeRef int_vec_type;
+   LLVMValueRef pos[NUM_CHANNELS];
+   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
+   struct lp_build_mask_context mask;
+   boolean early_depth_test;
+   unsigned attrib;
+   unsigned chan;
+
+   elem_type = lp_build_elem_type(type);
+   vec_type = lp_build_vec_type(type);
+   int_vec_type = lp_build_int_vec_type(type);
+
+   generate_pos(builder, x, y, a0_ptr, dadx_ptr, dady_ptr, pos);
+
+   lp_build_mask_begin(&mask, builder, type, *pmask);
+
+   early_depth_test =
+      lp->depth_stencil->depth.enabled &&
+      lp->framebuffer.zsbuf &&
+      !lp->depth_stencil->alpha.enabled &&
+      !lp->fs->info.uses_kill &&
+      !lp->fs->info.writes_z;
+
+   if(early_depth_test)
+      generate_depth(lp, builder, &key->depth,
+                          type, &mask,
+                          pos[2], depth_ptr);
+
+   memset(outputs, 0, sizeof outputs);
+
+   lp_build_tgsi_soa(builder, tokens, type, &mask,
+                     pos, a0_ptr, dadx_ptr, dady_ptr,
+                     consts_ptr, outputs, samplers_ptr);
+
+   for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
+      for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+         if(outputs[attrib][chan]) {
+            lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
+
+            switch (shader->info.output_semantic_name[attrib]) {
+            case TGSI_SEMANTIC_COLOR:
+               {
+                  unsigned cbuf = shader->info.output_semantic_index[attrib];
+
+                  lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
+
+                  /* Alpha test */
+                  /* XXX: should the alpha reference value be passed separately? */
+                  if(cbuf == 0 && chan == 3)
+                     lp_build_alpha_test(builder, &key->alpha, type,
+                                         &mask,
+                                         outputs[attrib][chan]);
+
+                  if(cbuf == 0)
+                     color[chan] = outputs[attrib][chan];
+
+                  break;
+               }
+
+            case TGSI_SEMANTIC_POSITION:
+               if(chan == 2)
+                  pos[2] = outputs[attrib][chan];
+               break;
+            }
+         }
+      }
+   }
+
+   if(!early_depth_test)
+      generate_depth(lp, builder, &key->depth,
+                          type, &mask,
+                          pos[2], depth_ptr);
+
+   lp_build_mask_end(&mask);
+
+   *pmask = mask.value;
+
+}
+
+
+/**
+ * Generate blending code according to blend->base state.
+ * The blend function will look like:
+ *    blend(mask, src_color, constant color, dst_color)
+ * dst_color will be modified and contain the result of the blend func.
+ */
+static void
+generate_blend(const struct pipe_blend_state *blend,
+               LLVMBuilderRef builder,
+               union lp_type type,
+               LLVMValueRef mask,
+               LLVMValueRef *src,
+               LLVMValueRef const_ptr,
+               LLVMValueRef dst_ptr)
+{
+   struct lp_build_context bld;
+   LLVMTypeRef vec_type;
+   LLVMTypeRef int_vec_type;
+   LLVMValueRef con[4];
+   LLVMValueRef dst[4];
+   LLVMValueRef res[4];
+   unsigned chan;
+
+   vec_type = lp_build_vec_type(type);
+   int_vec_type = lp_build_int_vec_type(type);
+
+   lp_build_context_init(&bld, builder, type);
+
+   for(chan = 0; chan < 4; ++chan) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
+
+      if(const_ptr)
+         con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
+      else
+         con[chan] = LLVMGetUndef(vec_type); /* FIXME */
+
+      dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
+
+      lp_build_name(con[chan], "con.%c", "rgba"[chan]);
+      lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
+   }
+
+   lp_build_blend_soa(builder, blend, type, src, dst, con, res);
+
+   for(chan = 0; chan < 4; ++chan) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
+      lp_build_name(res[chan], "res.%c", "rgba"[chan]);
+      res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
+      LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
+   }
+}
+
+
+static struct lp_fragment_shader_variant *
+generate_fragment(struct llvmpipe_context *lp,
+                  struct lp_fragment_shader *shader,
+                  const struct lp_fragment_shader_variant_key *key)
+{
+   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
+   struct lp_fragment_shader_variant *variant;
+   union lp_type fs_type;
+   union lp_type blend_type;
+   LLVMTypeRef fs_elem_type;
+   LLVMTypeRef fs_vec_type;
+   LLVMTypeRef fs_int_vec_type;
+   LLVMTypeRef blend_vec_type;
+   LLVMTypeRef blend_int_vec_type;
    LLVMTypeRef arg_types[10];
    LLVMTypeRef func_type;
    LLVMValueRef x;
@@ -177,24 +338,38 @@ shader_generate(struct llvmpipe_context *lp,
    LLVMValueRef samplers_ptr;
    LLVMBasicBlockRef block;
    LLVMBuilderRef builder;
-   LLVMValueRef pos[NUM_CHANNELS];
-   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
-   struct lp_build_mask_context mask;
-   boolean early_depth_test;
+   LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
+   LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
+   LLVMValueRef blend_mask;
+   LLVMValueRef blend_in_color[NUM_CHANNELS];
    LLVMValueRef fetch_texel;
+   unsigned num_fs;
    unsigned i;
-   unsigned attrib;
    unsigned chan;
 
 #ifdef DEBUG
    tgsi_dump(shader->base.tokens, 0);
-   debug_printf("depth.enabled = %u\n", key->depth.enabled);
-   debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
-   debug_printf("depth.writemask = %u\n", key->depth.writemask);
-   debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count);
-   debug_printf("alpha.enabled = %u\n", key->alpha.enabled);
-   debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
-   debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
+   if(key->depth.enabled) {
+      debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
+      debug_printf("depth.writemask = %u\n", key->depth.writemask);
+      debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count);
+   }
+   if(key->alpha.enabled) {
+      debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
+      debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
+   }
+   if(key->blend.logicop_enable) {
+      debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
+   }
+   else if(key->blend.blend_enable) {
+      debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
+      debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
+      debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
+      debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
+      debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
+      debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
+   }
+   debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 #endif
 
    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
@@ -204,26 +379,37 @@ shader_generate(struct llvmpipe_context *lp,
    variant->shader = shader;
    memcpy(&variant->key, key, sizeof *key);
 
-   type.value = 0;
-   type.floating = TRUE; /* floating point values */
-   type.sign = TRUE;     /* values are signed */
-   type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
-   type.width = 32;      /* 32-bit float */
-   type.length = 4;      /* 4 element per vector */
+   fs_type.value = 0;
+   fs_type.floating = TRUE; /* floating point values */
+   fs_type.sign = TRUE;     /* values are signed */
+   fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
+   fs_type.width = 32;      /* 32-bit float */
+   fs_type.length = 4;      /* 4 element per vector */
+   num_fs = 4;
 
-   elem_type = lp_build_elem_type(type);
-   vec_type = lp_build_vec_type(type);
-   int_vec_type = lp_build_int_vec_type(type);
+   blend_type.value = 0;
+   blend_type.floating = FALSE; /* values are integers */
+   blend_type.sign = FALSE;     /* values are unsigned */
+   blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
+   blend_type.width = 8;        /* 8-bit ubyte values */
+   blend_type.length = 16;      /* 16 elements per vector */
+
+   fs_elem_type = lp_build_elem_type(fs_type);
+   fs_vec_type = lp_build_vec_type(fs_type);
+   fs_int_vec_type = lp_build_int_vec_type(fs_type);
+
+   blend_vec_type = lp_build_vec_type(blend_type);
+   blend_int_vec_type = lp_build_int_vec_type(blend_type);
 
    arg_types[0] = LLVMInt32Type();                     /* x */
    arg_types[1] = LLVMInt32Type();                     /* y */
-   arg_types[2] = LLVMPointerType(elem_type, 0);       /* a0 */
-   arg_types[3] = LLVMPointerType(elem_type, 0);       /* dadx */
-   arg_types[4] = LLVMPointerType(elem_type, 0);       /* dady */
-   arg_types[5] = LLVMPointerType(elem_type, 0);       /* consts */
-   arg_types[6] = LLVMPointerType(int_vec_type, 0);    /* mask */
-   arg_types[7] = LLVMPointerType(vec_type, 0);        /* color */
-   arg_types[8] = LLVMPointerType(int_vec_type, 0);    /* depth */
+   arg_types[2] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
+   arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
+   arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dady */
+   arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* consts */
+   arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
+   arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
+   arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
    arg_types[9] = LLVMPointerType(LLVMInt8Type(), 0);  /* samplers */
 
    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
@@ -260,70 +446,57 @@ shader_generate(struct llvmpipe_context *lp,
    builder = LLVMCreateBuilder();
    LLVMPositionBuilderAtEnd(builder, block);
 
-   setup_pos_vector(builder, x, y, a0_ptr, dadx_ptr, dady_ptr, pos);
-
-   lp_build_mask_begin(&mask, builder, type,
-                       LLVMBuildLoad(builder, mask_ptr, ""));
-
-   early_depth_test =
-      lp->depth_stencil->depth.enabled &&
-      lp->framebuffer.zsbuf &&
-      !lp->depth_stencil->alpha.enabled &&
-      !lp->fs->info.uses_kill &&
-      !lp->fs->info.writes_z;
-
-   if(early_depth_test)
-      depth_test_generate(lp, builder, &key->depth,
-                          type, &mask,
-                          pos[2], depth_ptr);
-
-   memset(outputs, 0, sizeof outputs);
-
-   lp_build_tgsi_soa(builder, tokens, type, &mask,
-                     pos, a0_ptr, dadx_ptr, dady_ptr,
-                     consts_ptr, outputs, samplers_ptr);
-
-   for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
-      for(chan = 0; chan < NUM_CHANNELS; ++chan) {
-         if(outputs[attrib][chan]) {
-            lp_build_name(outputs[attrib][chan], "output%u.%c", attrib, "xyzw"[chan]);
-
-            switch (shader->info.output_semantic_name[attrib]) {
-            case TGSI_SEMANTIC_COLOR:
-               {
-                  unsigned cbuf = shader->info.output_semantic_index[attrib];
-                  LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf*NUM_CHANNELS + chan, 0);
-                  LLVMValueRef output_ptr = LLVMBuildGEP(builder, color_ptr, &index, 1, "");
-                  lp_build_name(outputs[attrib][chan], "color%u.%c", attrib, "rgba"[chan]);
-                  LLVMBuildStore(builder, outputs[attrib][chan], output_ptr);
-
-                  /* Alpha test */
-                  /* XXX: should the alpha reference value be passed separately? */
-                  if(cbuf == 0 && chan == 3)
-                     lp_build_alpha_test(builder, &key->alpha, type,
-                                         &mask,
-                                         outputs[attrib][chan]);
-
-                  break;
-               }
+   for(i = 0; i < num_fs; ++i) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef out_color[NUM_CHANNELS];
+      LLVMValueRef x_i;
+      LLVMValueRef depth_ptr_i;
+
+      /* TODO: Reuse position interpolation */
+      x_i = LLVMBuildAdd(builder, x, LLVMConstInt(LLVMInt32Type(), 2*i, 0), "");
+
+      fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
+      depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
+
+      generate_fs(lp,
+                  shader,
+                  key,
+                  builder,
+                  fs_type,
+                  i,
+                  x_i,
+                  y,
+                  a0_ptr,
+                  dadx_ptr,
+                  dady_ptr,
+                  consts_ptr,
+                  &fs_mask[i],
+                  out_color,
+                  depth_ptr_i,
+                  samplers_ptr);
+
+      for(chan = 0; chan < NUM_CHANNELS; ++chan)
+         fs_out_color[chan][i] = out_color[chan];
+   }
 
-            case TGSI_SEMANTIC_POSITION:
-               if(chan == 2)
-                  pos[2] = outputs[attrib][chan];
-               break;
-            }
-         }
-      }
+   for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+      lp_build_conv(builder, fs_type, blend_type,
+                    fs_out_color[chan], num_fs,
+                    &blend_in_color[chan], 1);
+      lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
    }
 
-   if(!early_depth_test)
-      depth_test_generate(lp, builder, &key->depth,
-                          type, &mask,
-                          pos[2], depth_ptr);
+   lp_build_conv_mask(builder, fs_type, blend_type,
+                               fs_mask, num_fs,
+                               &blend_mask, 1);
 
-   lp_build_mask_end(&mask);
-   if(mask.value)
-      LLVMBuildStore(builder, mask.value, mask_ptr);
+   generate_blend(&key->blend,
+                  builder,
+                  blend_type,
+                  blend_mask,
+                  blend_in_color,
+                  NULL /* FIXME: blend_const_color */,
+                  color_ptr);
 
    LLVMBuildRetVoid(builder);;
 
@@ -512,6 +685,7 @@ void llvmpipe_update_fs(struct llvmpipe_context *lp)
    memset(&key, 0, sizeof key);
    memcpy(&key.depth, &lp->depth_stencil->depth, sizeof &key.depth);
    memcpy(&key.alpha, &lp->depth_stencil->alpha, sizeof &key.alpha);
+   memcpy(&key.blend, &lp->blend->base, sizeof &key.blend);
 
    variant = shader->variants;
    while(variant) {
@@ -522,7 +696,7 @@ void llvmpipe_update_fs(struct llvmpipe_context *lp)
    }
 
    if(!variant)
-      variant = shader_generate(lp, shader, &key);
+      variant = generate_fragment(lp, shader, &key);
 
    shader->current = variant;
 }