anv/pipeline: Handle output lowering in anv_pipeline instead of spirv_to_nir
[mesa.git] / src / vulkan / gen8_pipeline.c
index 16143300cf1fccd35a00d83f7199424a753aa5fb..dee3c4049c2bebc76e0fd76f6057155db095ebe4 100644 (file)
 #include "gen8_pack.h"
 #include "gen9_pack.h"
 
+#include "genX_pipeline_util.h"
+
 static void
 emit_vertex_input(struct anv_pipeline *pipeline,
-                  const VkPipelineVertexInputStateCreateInfo *info)
+                  const VkPipelineVertexInputStateCreateInfo *info,
+                  const struct anv_graphics_pipeline_create_info *extra)
 {
-   const uint32_t num_dwords = 1 + info->attributeCount * 2;
-   uint32_t *p;
-
    static_assert(ANV_GEN >= 8, "should be compiling this for gen < 8");
 
-   if (info->attributeCount > 0) {
+   uint32_t elements;
+   if (extra && extra->disable_vs) {
+      /* If the VS is disabled, just assume the user knows what they're
+       * doing and apply the layout blindly.  This can only come from
+       * meta, so this *should* be safe.
+       */
+      elements = 0;
+      for (uint32_t i = 0; i < info->vertexAttributeDescriptionCount; i++)
+         elements |= (1 << info->pVertexAttributeDescriptions[i].location);
+   } else {
+      /* Pull inputs_read out of the VS prog data */
+      uint64_t inputs_read = pipeline->vs_prog_data.inputs_read;
+      assert((inputs_read & ((1 << VERT_ATTRIB_GENERIC0) - 1)) == 0);
+      elements = inputs_read >> VERT_ATTRIB_GENERIC0;
+   }
+
+   const uint32_t num_dwords = 1 + __builtin_popcount(elements) * 2;
+
+   uint32_t *p;
+   if (elements != 0) {
       p = anv_batch_emitn(&pipeline->batch, num_dwords,
                           GENX(3DSTATE_VERTEX_ELEMENTS));
+      memset(p + 1, 0, (num_dwords - 1) * 4);
    }
 
-   for (uint32_t i = 0; i < info->attributeCount; i++) {
+   for (uint32_t i = 0; i < info->vertexAttributeDescriptionCount; i++) {
       const VkVertexInputAttributeDescription *desc =
          &info->pVertexAttributeDescriptions[i];
-      const struct anv_format *format = anv_format_for_vk_format(desc->format);
+      enum isl_format format = anv_get_isl_format(desc->format,
+                                                  VK_IMAGE_ASPECT_COLOR_BIT,
+                                                  VK_IMAGE_TILING_LINEAR);
+
+      assert(desc->binding < 32);
+
+      if ((elements & (1 << desc->location)) == 0)
+         continue; /* Binding unused */
+
+      uint32_t slot = __builtin_popcount(elements & ((1 << desc->location) - 1));
 
       struct GENX(VERTEX_ELEMENT_STATE) element = {
          .VertexBufferIndex = desc->binding,
          .Valid = true,
-         .SourceElementFormat = format->surface_format,
+         .SourceElementFormat = format,
          .EdgeFlagEnable = false,
-         .SourceElementOffset = desc->offsetInBytes,
-         .Component0Control = VFCOMP_STORE_SRC,
-         .Component1Control = format->num_channels >= 2 ? VFCOMP_STORE_SRC : VFCOMP_STORE_0,
-         .Component2Control = format->num_channels >= 3 ? VFCOMP_STORE_SRC : VFCOMP_STORE_0,
-         .Component3Control = format->num_channels >= 4 ? VFCOMP_STORE_SRC : VFCOMP_STORE_1_FP
+         .SourceElementOffset = desc->offset,
+         .Component0Control = vertex_element_comp_control(format, 0),
+         .Component1Control = vertex_element_comp_control(format, 1),
+         .Component2Control = vertex_element_comp_control(format, 2),
+         .Component3Control = vertex_element_comp_control(format, 3),
       };
-      GENX(VERTEX_ELEMENT_STATE_pack)(NULL, &p[1 + i * 2], &element);
+      GENX(VERTEX_ELEMENT_STATE_pack)(NULL, &p[1 + slot * 2], &element);
 
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_INSTANCING),
                      .InstancingEnable = pipeline->instancing_enable[desc->binding],
-                     .VertexElementIndex = i,
+                     .VertexElementIndex = slot,
                      /* Vulkan so far doesn't have an instance divisor, so
                       * this is always 1 (ignored if not instancing). */
                      .InstanceDataStepRate = 1);
    }
 
+   const uint32_t id_slot = __builtin_popcount(elements);
    anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_SGVS),
                   .VertexIDEnable = pipeline->vs_prog_data.uses_vertexid,
                   .VertexIDComponentNumber = 2,
-                  .VertexIDElementOffset = info->bindingCount,
+                  .VertexIDElementOffset = id_slot,
                   .InstanceIDEnable = pipeline->vs_prog_data.uses_instanceid,
                   .InstanceIDComponentNumber = 3,
-                  .InstanceIDElementOffset = info->bindingCount);
+                  .InstanceIDElementOffset = id_slot);
 }
 
 static void
@@ -92,27 +122,9 @@ emit_ia_state(struct anv_pipeline *pipeline,
 
 static void
 emit_rs_state(struct anv_pipeline *pipeline,
-              const VkPipelineRasterStateCreateInfo *info,
+              const VkPipelineRasterizationStateCreateInfo *info,
               const struct anv_graphics_pipeline_create_info *extra)
 {
-   static const uint32_t vk_to_gen_cullmode[] = {
-      [VK_CULL_MODE_NONE]                       = CULLMODE_NONE,
-      [VK_CULL_MODE_FRONT_BIT]                  = CULLMODE_FRONT,
-      [VK_CULL_MODE_BACK_BIT]                   = CULLMODE_BACK,
-      [VK_CULL_MODE_FRONT_AND_BACK]             = CULLMODE_BOTH
-   };
-
-   static const uint32_t vk_to_gen_fillmode[] = {
-      [VK_POLYGON_MODE_FILL]                    = RASTER_SOLID,
-      [VK_POLYGON_MODE_LINE]                    = RASTER_WIREFRAME,
-      [VK_POLYGON_MODE_POINT]                   = RASTER_POINT,
-   };
-
-   static const uint32_t vk_to_gen_front_face[] = {
-      [VK_FRONT_FACE_COUNTER_CLOCKWISE]         = 1,
-      [VK_FRONT_FACE_CLOCKWISE]                 = 0
-   };
-
    struct GENX(3DSTATE_SF) sf = {
       GENX(3DSTATE_SF_header),
       .ViewportTransformEnable = !(extra && extra->disable_viewport),
@@ -135,11 +147,11 @@ emit_rs_state(struct anv_pipeline *pipeline,
       .BackFaceFillMode = vk_to_gen_fillmode[info->polygonMode],
       .ScissorRectangleEnable = !(extra && extra->disable_scissor),
 #if ANV_GEN == 8
-      .ViewportZClipTestEnable = info->depthClipEnable
+      .ViewportZClipTestEnable = true,
 #else
       /* GEN9+ splits ViewportZClipTestEnable into near and far enable bits */
-      .ViewportZFarClipTestEnable = info->depthClipEnable,
-      .ViewportZNearClipTestEnable = info->depthClipEnable,
+      .ViewportZFarClipTestEnable = true,
+      .ViewportZNearClipTestEnable = true,
 #endif
    };
 
@@ -148,66 +160,18 @@ emit_rs_state(struct anv_pipeline *pipeline,
 
 static void
 emit_cb_state(struct anv_pipeline *pipeline,
-              const VkPipelineColorBlendStateCreateInfo *info)
+              const VkPipelineColorBlendStateCreateInfo *info,
+              const VkPipelineMultisampleStateCreateInfo *ms_info)
 {
    struct anv_device *device = pipeline->device;
 
-   static const uint32_t vk_to_gen_logic_op[] = {
-      [VK_LOGIC_OP_COPY]                        = LOGICOP_COPY,
-      [VK_LOGIC_OP_CLEAR]                       = LOGICOP_CLEAR,
-      [VK_LOGIC_OP_AND]                         = LOGICOP_AND,
-      [VK_LOGIC_OP_AND_REVERSE]                 = LOGICOP_AND_REVERSE,
-      [VK_LOGIC_OP_AND_INVERTED]                = LOGICOP_AND_INVERTED,
-      [VK_LOGIC_OP_NO_OP]                       = LOGICOP_NOOP,
-      [VK_LOGIC_OP_XOR]                         = LOGICOP_XOR,
-      [VK_LOGIC_OP_OR]                          = LOGICOP_OR,
-      [VK_LOGIC_OP_NOR]                         = LOGICOP_NOR,
-      [VK_LOGIC_OP_EQUIVALENT]                  = LOGICOP_EQUIV,
-      [VK_LOGIC_OP_INVERT]                      = LOGICOP_INVERT,
-      [VK_LOGIC_OP_OR_REVERSE]                  = LOGICOP_OR_REVERSE,
-      [VK_LOGIC_OP_COPY_INVERTED]               = LOGICOP_COPY_INVERTED,
-      [VK_LOGIC_OP_OR_INVERTED]                 = LOGICOP_OR_INVERTED,
-      [VK_LOGIC_OP_NAND]                        = LOGICOP_NAND,
-      [VK_LOGIC_OP_SET]                         = LOGICOP_SET,
-   };
-
-   static const uint32_t vk_to_gen_blend[] = {
-      [VK_BLEND_FACTOR_ZERO]                    = BLENDFACTOR_ZERO,
-      [VK_BLEND_FACTOR_ONE]                     = BLENDFACTOR_ONE,
-      [VK_BLEND_FACTOR_SRC_COLOR]               = BLENDFACTOR_SRC_COLOR,
-      [VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR]     = BLENDFACTOR_INV_SRC_COLOR,
-      [VK_BLEND_FACTOR_DST_COLOR]               = BLENDFACTOR_DST_COLOR,
-      [VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR]     = BLENDFACTOR_INV_DST_COLOR,
-      [VK_BLEND_FACTOR_SRC_ALPHA]               = BLENDFACTOR_SRC_ALPHA,
-      [VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA]     = BLENDFACTOR_INV_SRC_ALPHA,
-      [VK_BLEND_FACTOR_DST_ALPHA]               = BLENDFACTOR_DST_ALPHA,
-      [VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA]     = BLENDFACTOR_INV_DST_ALPHA,
-      [VK_BLEND_FACTOR_CONSTANT_COLOR]          = BLENDFACTOR_CONST_COLOR,
-      [VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR]= BLENDFACTOR_INV_CONST_COLOR,
-      [VK_BLEND_FACTOR_CONSTANT_ALPHA]          = BLENDFACTOR_CONST_ALPHA,
-      [VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA]= BLENDFACTOR_INV_CONST_ALPHA,
-      [VK_BLEND_FACTOR_SRC_ALPHA_SATURATE]      = BLENDFACTOR_SRC_ALPHA_SATURATE,
-      [VK_BLEND_FACTOR_SRC1_COLOR]              = BLENDFACTOR_SRC1_COLOR,
-      [VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR]    = BLENDFACTOR_INV_SRC1_COLOR,
-      [VK_BLEND_FACTOR_SRC1_ALPHA]              = BLENDFACTOR_SRC1_ALPHA,
-      [VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA]    = BLENDFACTOR_INV_SRC1_ALPHA,
-   };
-
-   static const uint32_t vk_to_gen_blend_op[] = {
-      [VK_BLEND_OP_ADD]                         = BLENDFUNCTION_ADD,
-      [VK_BLEND_OP_SUBTRACT]                    = BLENDFUNCTION_SUBTRACT,
-      [VK_BLEND_OP_REVERSE_SUBTRACT]            = BLENDFUNCTION_REVERSE_SUBTRACT,
-      [VK_BLEND_OP_MIN]                         = BLENDFUNCTION_MIN,
-      [VK_BLEND_OP_MAX]                         = BLENDFUNCTION_MAX,
-   };
-
    uint32_t num_dwords = GENX(BLEND_STATE_length);
    pipeline->blend_state =
       anv_state_pool_alloc(&device->dynamic_state_pool, num_dwords * 4, 64);
 
    struct GENX(BLEND_STATE) blend_state = {
-      .AlphaToCoverageEnable = info->alphaToCoverageEnable,
-      .AlphaToOneEnable = info->alphaToOneEnable,
+      .AlphaToCoverageEnable = ms_info && ms_info->alphaToCoverageEnable,
+      .AlphaToOneEnable = ms_info && ms_info->alphaToOneEnable,
    };
 
    for (uint32_t i = 0; i < info->attachmentCount; i++) {
@@ -233,10 +197,10 @@ emit_cb_state(struct anv_pipeline *pipeline,
          .SourceAlphaBlendFactor = vk_to_gen_blend[a->srcAlphaBlendFactor],
          .DestinationAlphaBlendFactor = vk_to_gen_blend[a->dstAlphaBlendFactor],
          .AlphaBlendFunction = vk_to_gen_blend_op[a->alphaBlendOp],
-         .WriteDisableAlpha = !(a->channelWriteMask & VK_CHANNEL_A_BIT),
-         .WriteDisableRed = !(a->channelWriteMask & VK_CHANNEL_R_BIT),
-         .WriteDisableGreen = !(a->channelWriteMask & VK_CHANNEL_G_BIT),
-         .WriteDisableBlue = !(a->channelWriteMask & VK_CHANNEL_B_BIT),
+         .WriteDisableAlpha = !(a->colorWriteMask & VK_COLOR_COMPONENT_A_BIT),
+         .WriteDisableRed = !(a->colorWriteMask & VK_COLOR_COMPONENT_R_BIT),
+         .WriteDisableGreen = !(a->colorWriteMask & VK_COLOR_COMPONENT_G_BIT),
+         .WriteDisableBlue = !(a->colorWriteMask & VK_COLOR_COMPONENT_B_BIT),
       };
 
       /* Our hardware applies the blend factor prior to the blend function
@@ -258,34 +222,14 @@ emit_cb_state(struct anv_pipeline *pipeline,
    }
 
    GENX(BLEND_STATE_pack)(NULL, pipeline->blend_state.map, &blend_state);
+   if (!device->info.has_llc)
+      anv_state_clflush(pipeline->blend_state);
 
    anv_batch_emit(&pipeline->batch, GENX(3DSTATE_BLEND_STATE_POINTERS),
                   .BlendStatePointer = pipeline->blend_state.offset,
                   .BlendStatePointerValid = true);
 }
 
-static const uint32_t vk_to_gen_compare_op[] = {
-   [VK_COMPARE_OP_NEVER]                        = PREFILTEROPNEVER,
-   [VK_COMPARE_OP_LESS]                         = PREFILTEROPLESS,
-   [VK_COMPARE_OP_EQUAL]                        = PREFILTEROPEQUAL,
-   [VK_COMPARE_OP_LESS_OR_EQUAL]                = PREFILTEROPLEQUAL,
-   [VK_COMPARE_OP_GREATER]                      = PREFILTEROPGREATER,
-   [VK_COMPARE_OP_NOT_EQUAL]                    = PREFILTEROPNOTEQUAL,
-   [VK_COMPARE_OP_GREATER_OR_EQUAL]             = PREFILTEROPGEQUAL,
-   [VK_COMPARE_OP_ALWAYS]                       = PREFILTEROPALWAYS,
-};
-
-static const uint32_t vk_to_gen_stencil_op[] = {
-   [VK_STENCIL_OP_KEEP]                         = STENCILOP_KEEP,
-   [VK_STENCIL_OP_ZERO]                         = STENCILOP_ZERO,
-   [VK_STENCIL_OP_REPLACE]                      = STENCILOP_REPLACE,
-   [VK_STENCIL_OP_INCREMENT_AND_CLAMP]          = STENCILOP_INCRSAT,
-   [VK_STENCIL_OP_DECREMENT_AND_CLAMP]          = STENCILOP_DECRSAT,
-   [VK_STENCIL_OP_INVERT]                       = STENCILOP_INVERT,
-   [VK_STENCIL_OP_INCREMENT_AND_WRAP]           = STENCILOP_INCR,
-   [VK_STENCIL_OP_DECREMENT_AND_WRAP]           = STENCILOP_DECR,
-};
-
 static void
 emit_ds_state(struct anv_pipeline *pipeline,
               const VkPipelineDepthStencilStateCreateInfo *info)
@@ -329,8 +273,10 @@ emit_ds_state(struct anv_pipeline *pipeline,
 VkResult
 genX(graphics_pipeline_create)(
     VkDevice                                    _device,
+    struct anv_pipeline_cache *                 cache,
     const VkGraphicsPipelineCreateInfo*         pCreateInfo,
     const struct anv_graphics_pipeline_create_info *extra,
+    const VkAllocationCallbacks*                pAllocator,
     VkPipeline*                                 pPipeline)
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
@@ -340,32 +286,27 @@ genX(graphics_pipeline_create)(
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO);
 
-   pipeline = anv_device_alloc(device, sizeof(*pipeline), 8,
-                               VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
+   pipeline = anv_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
+                         VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (pipeline == NULL)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   result = anv_pipeline_init(pipeline, device, pCreateInfo, extra);
-   if (result != VK_SUCCESS)
+   result = anv_pipeline_init(pipeline, device, cache,
+                              pCreateInfo, extra, pAllocator);
+   if (result != VK_SUCCESS) {
+      anv_free2(&device->alloc, pAllocator, pipeline);
       return result;
-
-   /* FIXME: The compiler dead-codes FS inputs when we don't have a VS, so we
-    * hard code this to num_attributes - 2. This is because the attributes
-    * include VUE header and position, which aren't counted as varying
-    * inputs. */
-   if (pipeline->vs_simd8 == NO_KERNEL) {
-      pipeline->wm_prog_data.num_varying_inputs =
-         pCreateInfo->pVertexInputState->attributeCount - 2;
    }
 
    assert(pCreateInfo->pVertexInputState);
-   emit_vertex_input(pipeline, pCreateInfo->pVertexInputState);
+   emit_vertex_input(pipeline, pCreateInfo->pVertexInputState, extra);
    assert(pCreateInfo->pInputAssemblyState);
    emit_ia_state(pipeline, pCreateInfo->pInputAssemblyState, extra);
-   assert(pCreateInfo->pRasterState);
-   emit_rs_state(pipeline, pCreateInfo->pRasterState, extra);
+   assert(pCreateInfo->pRasterizationState);
+   emit_rs_state(pipeline, pCreateInfo->pRasterizationState, extra);
    emit_ds_state(pipeline, pCreateInfo->pDepthStencilState);
-   emit_cb_state(pipeline, pCreateInfo->pColorBlendState);
+   emit_cb_state(pipeline, pCreateInfo->pColorBlendState,
+                           pCreateInfo->pMultisampleState);
 
    anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_STATISTICS),
                    .StatisticsEnable = true);
@@ -440,23 +381,24 @@ genX(graphics_pipeline_create)(
    offset = 1;
    length = (gs_prog_data->base.vue_map.num_slots + 1) / 2 - offset;
 
-   if (pipeline->gs_vec4 == NO_KERNEL)
+   if (pipeline->gs_kernel == NO_KERNEL)
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_GS), .Enable = false);
    else
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_GS),
                      .SingleProgramFlow = false,
-                     .KernelStartPointer = pipeline->gs_vec4,
+                     .KernelStartPointer = pipeline->gs_kernel,
                      .VectorMaskEnable = Dmask,
                      .SamplerCount = 0,
                      .BindingTableEntryCount = 0,
                      .ExpectedVertexCount = pipeline->gs_vertex_count,
 
-                     .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_GEOMETRY],
+                     .ScratchSpaceBasePointer = pipeline->scratch_start[MESA_SHADER_GEOMETRY],
                      .PerThreadScratchSpace = ffs(gs_prog_data->base.base.total_scratch / 2048),
 
                      .OutputVertexSize = gs_prog_data->output_vertex_size_hwords * 2 - 1,
                      .OutputTopology = gs_prog_data->output_topology,
                      .VertexURBEntryReadLength = gs_prog_data->base.urb_read_length,
+                     .IncludeVertexHandles = gs_prog_data->base.include_vue_handles,
                      .DispatchGRFStartRegisterForURBData =
                         gs_prog_data->base.base.dispatch_grf_start_reg,
 
@@ -488,7 +430,10 @@ genX(graphics_pipeline_create)(
    offset = 1;
    length = (vue_prog_data->vue_map.num_slots + 1) / 2 - offset;
 
-   if (pipeline->vs_simd8 == NO_KERNEL || (extra && extra->disable_vs))
+   uint32_t vs_start = pipeline->vs_simd8 != NO_KERNEL ? pipeline->vs_simd8 :
+                                                         pipeline->vs_vec4;
+
+   if (vs_start == NO_KERNEL || (extra && extra->disable_vs))
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VS),
                      .FunctionEnable = false,
                      /* Even if VS is disabled, SBE still gets the amount of
@@ -497,7 +442,7 @@ genX(graphics_pipeline_create)(
                      .VertexURBEntryOutputLength = length);
    else
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VS),
-                     .KernelStartPointer = pipeline->vs_simd8,
+                     .KernelStartPointer = vs_start,
                      .SingleVertexDispatch = Multiple,
                      .VectorMaskEnable = Dmask,
                      .SamplerCount = 0,
@@ -509,7 +454,7 @@ genX(graphics_pipeline_create)(
                      .AccessesUAV = false,
                      .SoftwareExceptionEnable = false,
 
-                     .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_VERTEX],
+                     .ScratchSpaceBasePointer = pipeline->scratch_start[MESA_SHADER_VERTEX],
                      .PerThreadScratchSpace = ffs(vue_prog_data->base.total_scratch / 2048),
 
                      .DispatchGRFStartRegisterForURBData =
@@ -519,7 +464,7 @@ genX(graphics_pipeline_create)(
 
                      .MaximumNumberofThreads = device->info.max_vs_threads - 1,
                      .StatisticsEnable = false,
-                     .SIMD8DispatchEnable = true,
+                     .SIMD8DispatchEnable = pipeline->vs_simd8 != NO_KERNEL,
                      .VertexCacheDisable = false,
                      .FunctionEnable = true,
 
@@ -534,7 +479,7 @@ genX(graphics_pipeline_create)(
     * shared with other gens.
     */
    const struct brw_vue_map *fs_input_map;
-   if (pipeline->gs_vec4 == NO_KERNEL)
+   if (pipeline->gs_kernel == NO_KERNEL)
       fs_input_map = &vue_prog_data->vue_map;
    else
       fs_input_map = &gs_prog_data->base.vue_map;
@@ -550,16 +495,30 @@ genX(graphics_pipeline_create)(
       if (input_index < 0)
         continue;
 
-      /* We have to subtract two slots to accout for the URB entry output
-       * read offset in the VS and GS stages.
-       */
-      int source_attr = fs_input_map->varying_to_slot[attr] - 2;
+      int source_attr = fs_input_map->varying_to_slot[attr];
       max_source_attr = MAX2(max_source_attr, source_attr);
 
       if (input_index >= 16)
         continue;
 
-      swiz.Attribute[input_index].SourceAttribute = source_attr;
+      if (source_attr == -1) {
+         /* This attribute does not exist in the VUE--that means that the
+          * vertex shader did not write to it.  It could be that it's a
+          * regular varying read by the fragment shader but not written by the
+          * vertex shader or it's gl_PrimitiveID. In the first case the value
+          * is undefined, in the second it needs to be gl_PrimitiveID.
+          */
+         swiz.Attribute[input_index].ConstantSource = PRIM_ID;
+         swiz.Attribute[input_index].ComponentOverrideX = true;
+         swiz.Attribute[input_index].ComponentOverrideY = true;
+         swiz.Attribute[input_index].ComponentOverrideZ = true;
+         swiz.Attribute[input_index].ComponentOverrideW = true;
+      } else {
+         /* We have to subtract two slots to accout for the URB entry output
+          * read offset in the VS and GS stages.
+          */
+         swiz.Attribute[input_index].SourceAttribute = source_attr - 2;
+      }
    }
 
    anv_batch_emit(&pipeline->batch, GENX(3DSTATE_SBE),
@@ -621,7 +580,7 @@ genX(graphics_pipeline_create)(
                   .VectorMaskEnable = true,
                   .SamplerCount = 1,
 
-                  .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_FRAGMENT],
+                  .ScratchSpaceBasePointer = pipeline->scratch_start[MESA_SHADER_FRAGMENT],
                   .PerThreadScratchSpace = ffs(wm_prog_data->base.total_scratch / 2048),
 
                   .MaximumNumberofThreadsPerPSD = 64 - num_thread_bias,
@@ -660,7 +619,9 @@ genX(graphics_pipeline_create)(
 
 VkResult genX(compute_pipeline_create)(
     VkDevice                                    _device,
+    struct anv_pipeline_cache *                 cache,
     const VkComputePipelineCreateInfo*          pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
     VkPipeline*                                 pPipeline)
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
@@ -669,8 +630,8 @@ VkResult genX(compute_pipeline_create)(
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO);
 
-   pipeline = anv_device_alloc(device, sizeof(*pipeline), 8,
-                               VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
+   pipeline = anv_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
+                         VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (pipeline == NULL)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
@@ -679,18 +640,16 @@ VkResult genX(compute_pipeline_create)(
 
    pipeline->blend_state.map = NULL;
 
-   result = anv_reloc_list_init(&pipeline->batch_relocs, device);
+   result = anv_reloc_list_init(&pipeline->batch_relocs,
+                                pAllocator ? pAllocator : &device->alloc);
    if (result != VK_SUCCESS) {
-      anv_device_free(device, pipeline);
+      anv_free2(&device->alloc, pAllocator, pipeline);
       return result;
    }
    pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
    pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
    pipeline->batch.relocs = &pipeline->batch_relocs;
 
-   anv_state_stream_init(&pipeline->program_stream,
-                         &device->instruction_block_pool);
-
    /* When we free the pipeline, we detect stages based on the NULL status
     * of various prog_data pointers.  Make them NULL by default.
     */
@@ -699,21 +658,22 @@ VkResult genX(compute_pipeline_create)(
 
    pipeline->vs_simd8 = NO_KERNEL;
    pipeline->vs_vec4 = NO_KERNEL;
-   pipeline->gs_vec4 = NO_KERNEL;
+   pipeline->gs_kernel = NO_KERNEL;
 
    pipeline->active_stages = 0;
    pipeline->total_scratch = 0;
 
-   assert(pCreateInfo->stage.stage == VK_SHADER_STAGE_COMPUTE);
-   ANV_FROM_HANDLE(anv_shader, shader,  pCreateInfo->stage.shader);
-   anv_pipeline_compile_cs(pipeline, pCreateInfo, shader);
+   assert(pCreateInfo->stage.stage == VK_SHADER_STAGE_COMPUTE_BIT);
+   ANV_FROM_HANDLE(anv_shader_module, module,  pCreateInfo->stage.module);
+   anv_pipeline_compile_cs(pipeline, cache, pCreateInfo, module,
+                           pCreateInfo->stage.pName);
 
    pipeline->use_repclear = false;
 
    const struct brw_cs_prog_data *cs_prog_data = &pipeline->cs_prog_data;
 
    anv_batch_emit(&pipeline->batch, GENX(MEDIA_VFE_STATE),
-                  .ScratchSpaceBasePointer = pipeline->scratch_start[VK_SHADER_STAGE_COMPUTE],
+                  .ScratchSpaceBasePointer = pipeline->scratch_start[MESA_SHADER_COMPUTE],
                   .PerThreadScratchSpace = ffs(cs_prog_data->base.total_scratch / 2048),
                   .ScratchSpaceBasePointerHigh = 0,
                   .StackSize = 0,