i965/blorp: Do gen6 stencil offsets up-front
[mesa.git] / src / mesa / drivers / dri / i965 / brw_blorp_clear.cpp
index cd4c59019066cc30bb7d0bfcb14434d845d8ec0e..625baceee1c9c84e8cde7c62e224439afab2400e 100644 (file)
  * IN THE SOFTWARE.
  */
 
-extern "C" {
 #include "main/teximage.h"
 #include "main/blend.h"
 #include "main/fbobject.h"
 #include "main/renderbuffer.h"
-}
+#include "main/glformats.h"
 
-#include "glsl/ralloc.h"
+#include "util/ralloc.h"
 
 #include "intel_fbo.h"
 
 #include "brw_blorp.h"
+#include "brw_meta_util.h"
 #include "brw_context.h"
 #include "brw_eu.h"
 #include "brw_state.h"
 
+#include "nir_builder.h"
+
 #define FILE_DEBUG_FLAG DEBUG_BLORP
 
 struct brw_blorp_const_color_prog_key
@@ -45,176 +47,107 @@ struct brw_blorp_const_color_prog_key
    bool pad[3];
 };
 
-/**
- * Parameters for a blorp operation where the fragment shader outputs a
- * constant color.  This is used for both fast color clears and color
- * resolves.
- */
-class brw_blorp_const_color_params : public brw_blorp_params
-{
-public:
-   virtual uint32_t get_wm_prog(struct brw_context *brw,
-                                brw_blorp_prog_data **prog_data) const;
-
-protected:
-   brw_blorp_const_color_prog_key wm_prog_key;
-};
-
-class brw_blorp_clear_params : public brw_blorp_const_color_params
-{
-public:
-   brw_blorp_clear_params(struct brw_context *brw,
-                          struct gl_framebuffer *fb,
-                          struct gl_renderbuffer *rb,
-                          GLubyte *color_mask,
-                          bool partial_clear);
-};
-
-
-/**
- * Parameters for a blorp operation that performs a "render target resolve".
- * This is used to resolve pending fast clear pixels before a color buffer is
- * used for texturing, ReadPixels, or scanout.
- */
-class brw_blorp_rt_resolve_params : public brw_blorp_const_color_params
-{
-public:
-   brw_blorp_rt_resolve_params(struct brw_context *brw,
-                               struct intel_mipmap_tree *mt);
-};
-
-
-class brw_blorp_const_color_program
+static void
+brw_blorp_params_get_clear_kernel(struct brw_context *brw,
+                                  struct brw_blorp_params *params,
+                                  bool use_replicated_data)
 {
-public:
-   brw_blorp_const_color_program(struct brw_context *brw,
-                                 const brw_blorp_const_color_prog_key *key);
-   ~brw_blorp_const_color_program();
+   struct brw_blorp_const_color_prog_key blorp_key;
+   memset(&blorp_key, 0, sizeof(blorp_key));
+   blorp_key.use_simd16_replicated_data = use_replicated_data;
 
-   const GLuint *compile(struct brw_context *brw, GLuint *program_size);
+   if (brw_search_cache(&brw->cache, BRW_CACHE_BLORP_PROG,
+                        &blorp_key, sizeof(blorp_key),
+                        &params->wm_prog_kernel, &params->wm_prog_data))
+      return;
 
-   brw_blorp_prog_data prog_data;
+   void *mem_ctx = ralloc_context(NULL);
 
-private:
-   void alloc_regs();
+   nir_builder b;
+   nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+   b.shader->info.name = ralloc_strdup(b.shader, "BLORP-clear");
 
-   void *mem_ctx;
-   struct brw_context *brw;
-   const brw_blorp_const_color_prog_key *key;
-   struct brw_compile func;
+   nir_variable *v_color = nir_variable_create(b.shader, nir_var_shader_in,
+                                               glsl_vec4_type(), "v_color");
+   v_color->data.location = VARYING_SLOT_VAR0;
+   v_color->data.interpolation = INTERP_MODE_FLAT;
 
-   /* Thread dispatch header */
-   struct brw_reg R0;
+   nir_variable *frag_color = nir_variable_create(b.shader, nir_var_shader_out,
+                                                  glsl_vec4_type(),
+                                                  "gl_FragColor");
+   frag_color->data.location = FRAG_RESULT_COLOR;
 
-   /* Pixel X/Y coordinates (always in R1). */
-   struct brw_reg R1;
+   nir_copy_var(&b, frag_color, v_color);
 
-   /* Register with push constants (a single vec4) */
-   struct brw_reg clear_rgba;
+   struct brw_wm_prog_key wm_key;
+   brw_blorp_init_wm_prog_key(&wm_key);
 
-   /* MRF used for render target writes */
-   GLuint base_mrf;
-};
+   struct brw_blorp_prog_data prog_data;
+   unsigned program_size;
+   const unsigned *program =
+      brw_blorp_compile_nir_shader(brw, b.shader, &wm_key, use_replicated_data,
+                                   &prog_data, &program_size);
 
-brw_blorp_const_color_program::brw_blorp_const_color_program(
-      struct brw_context *brw,
-      const brw_blorp_const_color_prog_key *key)
-   : mem_ctx(ralloc_context(NULL)),
-     brw(brw),
-     key(key),
-     R0(),
-     R1(),
-     clear_rgba(),
-     base_mrf(0)
-{
-   brw_init_compile(brw, &func, mem_ctx);
-}
+   brw_upload_cache(&brw->cache, BRW_CACHE_BLORP_PROG,
+                    &blorp_key, sizeof(blorp_key),
+                    program, program_size,
+                    &prog_data, sizeof(prog_data),
+                    &params->wm_prog_kernel, &params->wm_prog_data);
 
-brw_blorp_const_color_program::~brw_blorp_const_color_program()
-{
    ralloc_free(mem_ctx);
 }
 
-
-/**
- * Determine if fast color clear supports the given clear color.
- *
- * Fast color clear can only clear to color values of 1.0 or 0.0.  At the
- * moment we only support floating point, unorm, and snorm buffers.
- */
 static bool
-is_color_fast_clear_compatible(struct brw_context *brw,
-                               gl_format format,
-                               const union gl_color_union *color)
+set_write_disables(const struct intel_renderbuffer *irb, 
+                   const GLubyte *color_mask, bool *color_write_disable)
 {
-   if (_mesa_is_format_integer_color(format))
-      return false;
-
-   for (int i = 0; i < 4; i++) {
-      if (color->f[i] != 0.0 && color->f[i] != 1.0) {
-         perf_debug("Clear color unsupported by fast color clear.  "
-                    "Falling back to slow clear.\n");
-         return false;
-      }
-   }
-   return true;
-}
+   /* Format information in the renderbuffer represents the requirements
+    * given by the client. There are cases where the backing miptree uses,
+    * for example, RGBA to represent RGBX. Since the client is only expecting
+    * RGB we can treat alpha as not used and write whatever we like into it.
+    */
+   const GLenum base_format = irb->Base.Base._BaseFormat;
+   const int components = _mesa_base_format_component_count(base_format);
+   bool disables = false;
 
+   assert(components > 0);
 
-/**
- * Convert the given color to a bitfield suitable for ORing into DWORD 7 of
- * SURFACE_STATE.
- */
-static uint32_t
-compute_fast_clear_color_bits(const union gl_color_union *color)
-{
-   uint32_t bits = 0;
-   for (int i = 0; i < 4; i++) {
-      if (color->f[i] != 0.0)
-         bits |= 1 << (GEN7_SURFACE_CLEAR_COLOR_SHIFT + (3 - i));
+   for (int i = 0; i < components; i++) {
+      color_write_disable[i] = !color_mask[i];
+      disables = disables || !color_mask[i];
    }
-   return bits;
-}
 
+   return disables;
+}
 
-brw_blorp_clear_params::brw_blorp_clear_params(struct brw_context *brw,
-                                               struct gl_framebuffer *fb,
-                                               struct gl_renderbuffer *rb,
-                                               GLubyte *color_mask,
-                                               bool partial_clear)
+static bool
+do_single_blorp_clear(struct brw_context *brw, struct gl_framebuffer *fb,
+                      struct gl_renderbuffer *rb, unsigned buf,
+                      bool partial_clear, bool encode_srgb, unsigned layer)
 {
-   struct intel_context *intel = &brw->intel;
-   struct gl_context *ctx = &intel->ctx;
+   struct gl_context *ctx = &brw->ctx;
    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
+   mesa_format format = irb->mt->format;
 
-   dst.set(brw, irb->mt, irb->mt_level, irb->mt_layer);
+   struct brw_blorp_params params;
+   brw_blorp_params_init(&params);
 
-   /* Override the surface format according to the context's sRGB rules. */
-   gl_format format = _mesa_get_render_format(ctx, irb->mt->format);
-   dst.brw_surfaceformat = brw->render_target_format[format];
+   if (!encode_srgb && _mesa_get_format_color_encoding(format) == GL_SRGB)
+      format = _mesa_get_srgb_format_linear(format);
 
-   x0 = fb->_Xmin;
-   x1 = fb->_Xmax;
+   params.x0 = fb->_Xmin;
+   params.x1 = fb->_Xmax;
    if (rb->Name != 0) {
-      y0 = fb->_Ymin;
-      y1 = fb->_Ymax;
+      params.y0 = fb->_Ymin;
+      params.y1 = fb->_Ymax;
    } else {
-      y0 = rb->Height - fb->_Ymax;
-      y1 = rb->Height - fb->_Ymin;
+      params.y0 = rb->Height - fb->_Ymax;
+      params.y1 = rb->Height - fb->_Ymin;
    }
 
-   float *push_consts = (float *)&wm_push_consts;
-
-   push_consts[0] = ctx->Color.ClearColor.f[0];
-   push_consts[1] = ctx->Color.ClearColor.f[1];
-   push_consts[2] = ctx->Color.ClearColor.f[2];
-   push_consts[3] = ctx->Color.ClearColor.f[3];
+   memcpy(&params.wm_inputs, ctx->Color.ClearColor.f, sizeof(float) * 4);
 
-   use_wm_prog = true;
-
-   memset(&wm_prog_key, 0, sizeof(wm_prog_key));
-
-   wm_prog_key.use_simd16_replicated_data = true;
+   bool use_simd16_replicated_data = true;
 
    /* From the SNB PRM (Vol4_Part1):
     *
@@ -222,232 +155,113 @@ brw_blorp_clear_params::brw_blorp_clear_params(struct brw_context *brw,
     *      accessing tiled memory.  Using this Message Type to access linear
     *      (untiled) memory is UNDEFINED."
     */
-   if (irb->mt->region->tiling == I915_TILING_NONE)
-      wm_prog_key.use_simd16_replicated_data = false;
+   if (irb->mt->tiling == I915_TILING_NONE)
+      use_simd16_replicated_data = false;
 
    /* Constant color writes ignore everyting in blend and color calculator
     * state.  This is not documented.
     */
-   for (int i = 0; i < 4; i++) {
-      if (!color_mask[i]) {
-         color_write_disable[i] = true;
-         wm_prog_key.use_simd16_replicated_data = false;
-      }
-   }
-
-   /* If we can do this as a fast color clear, do so. */
-   if (irb->mt->mcs_state != INTEL_MCS_STATE_NONE && !partial_clear &&
-       wm_prog_key.use_simd16_replicated_data &&
-       is_color_fast_clear_compatible(brw, format, &ctx->Color.ClearColor)) {
-      memset(push_consts, 0xff, 4*sizeof(float));
-      fast_clear_op = GEN7_FAST_CLEAR_OP_FAST_CLEAR;
-
-      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
-       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
-       *
-       *     Clear pass must have a clear rectangle that must follow alignment
-       *     rules in terms of pixels and lines as shown in the table
-       *     below. Further, the clear-rectangle height and width must be
-       *     multiple of the following dimensions. If the height and width of
-       *     the render target being cleared do not meet these requirements,
-       *     an MCS buffer can be created such that it follows the requirement
-       *     and covers the RT.
-       *
-       * The alignment size in the table that follows is related to the
-       * alignment size returned by intel_get_non_msrt_mcs_alignment(), but
-       * with X alignment multiplied by 16 and Y alignment multiplied by 32.
-       */
-      unsigned x_align, y_align;
-      intel_get_non_msrt_mcs_alignment(brw, irb->mt, &x_align, &y_align);
-      x_align *= 16;
-      y_align *= 32;
-      x0 = ROUND_DOWN_TO(x0, x_align);
-      y0 = ROUND_DOWN_TO(y0, y_align);
-      x1 = ALIGN(x1, x_align);
-      y1 = ALIGN(y1, y_align);
-
-      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
-       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
-       *
-       *     In order to optimize the performance MCS buffer (when bound to 1X
-       *     RT) clear similarly to MCS buffer clear for MSRT case, clear rect
-       *     is required to be scaled by the following factors in the
-       *     horizontal and vertical directions:
-       *
-       * The X and Y scale down factors in the table that follows are each
-       * equal to half the alignment value computed above.
-       */
-      unsigned x_scaledown = x_align / 2;
-      unsigned y_scaledown = y_align / 2;
-      x0 /= x_scaledown;
-      y0 /= y_scaledown;
-      x1 /= x_scaledown;
-      y1 /= y_scaledown;
+   if (set_write_disables(irb, ctx->Color.ColorMask[buf],
+                          params.color_write_disable))
+      use_simd16_replicated_data = false;
+
+   if (irb->mt->fast_clear_state != INTEL_FAST_CLEAR_STATE_NO_MCS &&
+       !partial_clear && use_simd16_replicated_data &&
+       brw_is_color_fast_clear_compatible(brw, irb->mt,
+                                          &ctx->Color.ClearColor)) {
+      memset(&params.wm_inputs, 0xff, 4*sizeof(float));
+      params.fast_clear_op = GEN7_PS_RENDER_TARGET_FAST_CLEAR_ENABLE;
+
+      brw_get_fast_clear_rect(brw, fb, irb->mt, &params.x0, &params.y0,
+                              &params.x1, &params.y1);
+   } else {
+      brw_meta_get_buffer_rect(fb, &params.x0, &params.y0,
+                               &params.x1, &params.y1);
    }
-}
-
-
-brw_blorp_rt_resolve_params::brw_blorp_rt_resolve_params(
-      struct brw_context *brw,
-      struct intel_mipmap_tree *mt)
-{
-   dst.set(brw, mt, 0 /* level */, 0 /* layer */);
-
-   /* From the Ivy Bridge PRM, Vol2 Part1 11.9 "Render Target Resolve":
-    *
-    *     A rectangle primitive must be scaled down by the following factors
-    *     with respect to render target being resolved.
-    *
-    * The scaledown factors in the table that follows are related to the
-    * alignment size returned by intel_get_non_msrt_mcs_alignment(), but with
-    * X and Y alignment each divided by 2.
-    */
-   unsigned x_align, y_align;
-   intel_get_non_msrt_mcs_alignment(brw, mt, &x_align, &y_align);
-   unsigned x_scaledown = x_align / 2;
-   unsigned y_scaledown = y_align / 2;
-   x0 = y0 = 0;
-   x1 = ALIGN(mt->logical_width0, x_scaledown) / x_scaledown;
-   y1 = ALIGN(mt->logical_height0, y_scaledown) / y_scaledown;
 
-   fast_clear_op = GEN7_FAST_CLEAR_OP_RESOLVE;
+   brw_blorp_params_get_clear_kernel(brw, &params, use_simd16_replicated_data);
 
-   /* Note: there is no need to initialize push constants because it doesn't
-    * matter what data gets dispatched to the render target.  However, we must
-    * ensure that the fragment shader delivers the data using the "replicated
-    * color" message.
-    */
-   use_wm_prog = true;
-   memset(&wm_prog_key, 0, sizeof(wm_prog_key));
-   wm_prog_key.use_simd16_replicated_data = true;
-}
+   const bool is_fast_clear =
+      params.fast_clear_op == GEN7_PS_RENDER_TARGET_FAST_CLEAR_ENABLE;
+   if (is_fast_clear) {
+      /* Record the clear color in the miptree so that it will be
+       * programmed in SURFACE_STATE by later rendering and resolve
+       * operations.
+       */
+      const bool color_updated = brw_meta_set_fast_clear_color(
+                                    brw, irb->mt, &ctx->Color.ClearColor);
 
+      /* If the buffer is already in INTEL_FAST_CLEAR_STATE_CLEAR, the clear
+       * is redundant and can be skipped.
+       */
+      if (!color_updated &&
+          irb->mt->fast_clear_state == INTEL_FAST_CLEAR_STATE_CLEAR)
+         return true;
 
-uint32_t
-brw_blorp_const_color_params::get_wm_prog(struct brw_context *brw,
-                                          brw_blorp_prog_data **prog_data)
-   const
-{
-   uint32_t prog_offset;
-   if (!brw_search_cache(&brw->cache, BRW_BLORP_CONST_COLOR_PROG,
-                         &this->wm_prog_key, sizeof(this->wm_prog_key),
-                         &prog_offset, prog_data)) {
-      brw_blorp_const_color_program prog(brw, &this->wm_prog_key);
-      GLuint program_size;
-      const GLuint *program = prog.compile(brw, &program_size);
-      brw_upload_cache(&brw->cache, BRW_BLORP_CONST_COLOR_PROG,
-                       &this->wm_prog_key, sizeof(this->wm_prog_key),
-                       program, program_size,
-                       &prog.prog_data, sizeof(prog.prog_data),
-                       &prog_offset, prog_data);
+      /* If the MCS buffer hasn't been allocated yet, we need to allocate
+       * it now.
+       */
+      if (!irb->mt->mcs_mt) {
+         if (!intel_miptree_alloc_non_msrt_mcs(brw, irb->mt)) {
+            /* MCS allocation failed--probably this will only happen in
+             * out-of-memory conditions.  But in any case, try to recover
+             * by falling back to a non-blorp clear technique.
+             */
+            return false;
+         }
+      }
    }
-   return prog_offset;
-}
 
-void
-brw_blorp_const_color_program::alloc_regs()
-{
-   int reg = 0;
-   this->R0 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW);
-   this->R1 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW);
-
-   prog_data.first_curbe_grf = reg;
-   clear_rgba = retype(brw_vec4_grf(reg++, 0), BRW_REGISTER_TYPE_F);
-   reg += BRW_BLORP_NUM_PUSH_CONST_REGS;
-
-   /* Make sure we didn't run out of registers */
-   assert(reg <= GEN7_MRF_HACK_START);
+   brw_blorp_surface_info_init(brw, &params.dst, irb->mt, irb->mt_level,
+                               layer, format, true);
 
-   this->base_mrf = 2;
-}
-
-const GLuint *
-brw_blorp_const_color_program::compile(struct brw_context *brw,
-                                       GLuint *program_size)
-{
-   /* Set up prog_data */
-   memset(&prog_data, 0, sizeof(prog_data));
-   prog_data.persample_msaa_dispatch = false;
+   /* Override the surface format according to the context's sRGB rules. */
+   params.dst.view.format = (enum isl_format)brw->render_target_format[format];
 
-   alloc_regs();
+   const char *clear_type;
+   if (is_fast_clear)
+      clear_type = "fast";
+   else if (use_simd16_replicated_data)
+      clear_type = "replicated";
+   else
+      clear_type = "slow";
 
-   brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
+   DBG("%s (%s) to mt %p level %d layer %d\n", __FUNCTION__, clear_type,
+       irb->mt, irb->mt_level, irb->mt_layer);
 
-   struct brw_reg mrf_rt_write =
-      retype(vec16(brw_message_reg(base_mrf)), BRW_REGISTER_TYPE_F);
+   brw_blorp_exec(brw, &params);
 
-   uint32_t mlen, msg_type;
-   if (key->use_simd16_replicated_data) {
-      /* The message payload is a single register with the low 4 floats/ints
-       * filled with the constant clear color.
+   if (is_fast_clear) {
+      /* Now that the fast clear has occurred, put the buffer in
+       * INTEL_FAST_CLEAR_STATE_CLEAR so that we won't waste time doing
+       * redundant clears.
        */
-      brw_set_mask_control(&func, BRW_MASK_DISABLE);
-      brw_MOV(&func, vec4(brw_message_reg(base_mrf)), clear_rgba);
-      brw_set_mask_control(&func, BRW_MASK_ENABLE);
-
-      msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED;
-      mlen = 1;
-   } else {
-      for (int i = 0; i < 4; i++) {
-         /* The message payload is pairs of registers for 16 pixels each of r,
-          * g, b, and a.
-          */
-         brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
-         brw_MOV(&func,
-                 brw_message_reg(base_mrf + i * 2),
-                 brw_vec1_grf(clear_rgba.nr, i));
-         brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
-      }
-
-      msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
-      mlen = 8;
+      irb->mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_CLEAR;
+   } else if (intel_miptree_is_lossless_compressed(brw, irb->mt)) {
+      /* Compressed buffers can be cleared also using normal rep-clear. In
+       * such case they bahave such as if they were drawn using normal 3D
+       * render pipeline, and we simply mark the mcs as dirty.
+       */
+      assert(partial_clear);
+      irb->mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_UNRESOLVED;
    }
 
-   /* Now write to the render target and terminate the thread */
-   brw_fb_WRITE(&func,
-                16 /* dispatch_width */,
-                base_mrf /* msg_reg_nr */,
-                mrf_rt_write /* src0 */,
-                msg_type,
-                BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX,
-                mlen,
-                0 /* response_length */,
-                true /* eot */,
-                false /* header present */);
-
-   if (unlikely(INTEL_DEBUG & DEBUG_BLORP)) {
-      printf("Native code for BLORP clear:\n");
-      brw_dump_compile(&func, stdout, 0, func.next_insn_offset);
-      printf("\n");
-   }
-   return brw_get_program(&func, program_size);
+   return true;
 }
 
+
 extern "C" {
 bool
 brw_blorp_clear_color(struct brw_context *brw, struct gl_framebuffer *fb,
-                      bool partial_clear)
+                      GLbitfield mask, bool partial_clear, bool encode_srgb)
 {
-   struct gl_context *ctx = &brw->intel.ctx;
-
-   /* The constant color clear code doesn't work for multisampled surfaces, so
-    * we need to support falling back to other clear mechanisms.
-    * Unfortunately, our clear code is based on a bitmask that doesn't
-    * distinguish individual color attachments, so we walk the attachments to
-    * see if any require fallback, and fall back for all if any of them need
-    * to.
-    */
-   for (unsigned buf = 0; buf < ctx->DrawBuffer->_NumColorDrawBuffers; buf++) {
-      struct gl_renderbuffer *rb = ctx->DrawBuffer->_ColorDrawBuffers[buf];
+   for (unsigned buf = 0; buf < fb->_NumColorDrawBuffers; buf++) {
+      struct gl_renderbuffer *rb = fb->_ColorDrawBuffers[buf];
       struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 
-      if (irb && irb->mt->msaa_layout != INTEL_MSAA_LAYOUT_NONE)
-         return false;
-   }
-
-   for (unsigned buf = 0; buf < ctx->DrawBuffer->_NumColorDrawBuffers; buf++) {
-      struct gl_renderbuffer *rb = ctx->DrawBuffer->_ColorDrawBuffers[buf];
-      struct intel_renderbuffer *irb = intel_renderbuffer(rb);
+      /* Only clear the buffers present in the provided mask */
+      if (((1 << fb->_ColorDrawBufferIndexes[buf]) & mask) == 0)
+         continue;
 
       /* If this is an ES2 context or GL_ARB_ES2_compatibility is supported,
        * the framebuffer can be complete with some attachments missing.  In
@@ -456,56 +270,27 @@ brw_blorp_clear_color(struct brw_context *brw, struct gl_framebuffer *fb,
       if (rb == NULL)
          continue;
 
-      brw_blorp_clear_params params(brw, fb, rb, ctx->Color.ColorMask[buf],
-                                    partial_clear);
-
-      bool is_fast_clear =
-         (params.fast_clear_op == GEN7_FAST_CLEAR_OP_FAST_CLEAR);
-      if (is_fast_clear) {
-         /* Record the clear color in the miptree so that it will be
-          * programmed in SURFACE_STATE by later rendering and resolve
-          * operations.
-          */
-         uint32_t new_color_value =
-            compute_fast_clear_color_bits(&ctx->Color.ClearColor);
-         if (irb->mt->fast_clear_color_value != new_color_value) {
-            irb->mt->fast_clear_color_value = new_color_value;
-            brw->state.dirty.brw |= BRW_NEW_SURFACES;
-         }
-
-         /* If the buffer is already in INTEL_MCS_STATE_CLEAR, the clear is
-          * redundant and can be skipped.
-          */
-         if (irb->mt->mcs_state == INTEL_MCS_STATE_CLEAR)
-            continue;
-
-         /* If the MCS buffer hasn't been allocated yet, we need to allocate
-          * it now.
-          */
-         if (!irb->mt->mcs_mt) {
-            if (!intel_miptree_alloc_non_msrt_mcs(brw, irb->mt)) {
-               /* MCS allocation failed--probably this will only happen in
-                * out-of-memory conditions.  But in any case, try to recover
-                * by falling back to a non-blorp clear technique.
-                */
+      if (fb->MaxNumLayers > 0) {
+         unsigned layer_multiplier =
+            (irb->mt->msaa_layout == INTEL_MSAA_LAYOUT_UMS ||
+             irb->mt->msaa_layout == INTEL_MSAA_LAYOUT_CMS) ?
+            irb->mt->num_samples : 1;
+         unsigned num_layers = irb->layer_count;
+         for (unsigned layer = 0; layer < num_layers; layer++) {
+            if (!do_single_blorp_clear(
+                    brw, fb, rb, buf, partial_clear, encode_srgb,
+                    irb->mt_layer + layer * layer_multiplier)) {
                return false;
             }
-            brw->state.dirty.brw |= BRW_NEW_SURFACES;
          }
+      } else {
+         unsigned layer = irb->mt_layer;
+         if (!do_single_blorp_clear(brw, fb, rb, buf, partial_clear,
+                                    encode_srgb, layer))
+            return false;
       }
 
-      DBG("%s to mt %p level %d layer %d\n", __FUNCTION__,
-          irb->mt, irb->mt_level, irb->mt_layer);
-
-      brw_blorp_exec(brw, &params);
-
-      if (is_fast_clear) {
-         /* Now that the fast clear has occurred, put the buffer in
-          * INTEL_MCS_STATE_CLEAR so that we won't waste time doing redundant
-          * clears.
-          */
-         irb->mt->mcs_state = INTEL_MCS_STATE_CLEAR;
-      }
+      irb->need_downsample = true;
    }
 
    return true;
@@ -516,9 +301,32 @@ brw_blorp_resolve_color(struct brw_context *brw, struct intel_mipmap_tree *mt)
 {
    DBG("%s to mt %p\n", __FUNCTION__, mt);
 
-   brw_blorp_rt_resolve_params params(brw, mt);
+   const mesa_format format = _mesa_get_srgb_format_linear(mt->format);
+
+   struct brw_blorp_params params;
+   brw_blorp_params_init(&params);
+
+   brw_blorp_surface_info_init(brw, &params.dst, mt,
+                               0 /* level */, 0 /* layer */, format, true);
+
+   brw_get_resolve_rect(brw, mt, &params.x0, &params.y0,
+                        &params.x1, &params.y1);
+
+   if (intel_miptree_is_lossless_compressed(brw, mt))
+      params.resolve_type = GEN9_PS_RENDER_TARGET_RESOLVE_FULL;
+   else
+      params.resolve_type = GEN7_PS_RENDER_TARGET_RESOLVE_ENABLE;
+
+   /* Note: there is no need to initialize push constants because it doesn't
+    * matter what data gets dispatched to the render target.  However, we must
+    * ensure that the fragment shader delivers the data using the "replicated
+    * color" message.
+    */
+
+   brw_blorp_params_get_clear_kernel(brw, &params, true);
+
    brw_blorp_exec(brw, &params);
-   mt->mcs_state = INTEL_MCS_STATE_RESOLVED;
+   mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_RESOLVED;
 }
 
 } /* extern "C" */