pvr: Add routine for filling out usc_mrt_setup from dynamic rendering state

Signed-off-by: Ella Stanforth <ella@igalia.com>
Co-authored-by: Luigi Santivetti <luigi.santivetti@imgtec.com>
Acked-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38744>

src/imagination/vulkan/meson.build

@@ -43,6 +43,7 @@ pvr_files = files(
   'pvr_job_context.c',
   'pvr_job_render.c',
   'pvr_job_transfer.c',
+  'pvr_mrt.c',
   'pvr_pass.c',
   'pvr_physical_device.c',
   'pvr_pipeline.c',

src/imagination/vulkan/pvr_hw_pass.h

@@ -31,12 +31,7 @@
 struct pvr_device;
 struct pvr_render_pass;
 
-/* Specifies the location of render target writes. */
-enum usc_mrt_resource_type {
-   USC_MRT_RESOURCE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid. */
-   USC_MRT_RESOURCE_TYPE_OUTPUT_REG,
-   USC_MRT_RESOURCE_TYPE_MEMORY,
-};
+#include "pvr_mrt.h"
 
 enum pvr_resolve_type {
    PVR_RESOLVE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid. */
@@ -55,79 +50,6 @@ enum pvr_renderpass_hwsetup_input_access {
    PVR_RENDERPASS_HWSETUP_INPUT_ACCESS_ONCHIP_ZREPLICATE,
 };
 
-#define PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS (4)
-
-struct usc_mrt_desc {
-   /* Size (in bytes) of the intermediate storage required for each pixel in the
-    * render target.
-    */
-   uint32_t intermediate_size;
-
-   /* Mask of the bits from each dword which are read by the PBE. */
-   uint32_t valid_mask[PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS];
-
-   /* Higher number = higher priority. Used to decide which render targets get
-    * allocated dedicated output registers.
-    */
-   uint32_t priority;
-};
-
-struct usc_mrt_resource {
-   /* Input description of render target. */
-   struct usc_mrt_desc mrt_desc;
-
-   /* Resource type allocated for render target. */
-   enum usc_mrt_resource_type type;
-
-   /* Intermediate pixel size (in bytes). */
-   uint32_t intermediate_size;
-
-   union {
-      /* If type == USC_MRT_RESOURCE_TYPE_OUTPUT_REG. */
-      struct {
-         /* The output register to use. */
-         uint32_t output_reg;
-
-         /* The offset in bytes into the output register. */
-         uint32_t offset;
-      } reg;
-
-      /* If type == USC_MRT_RESOURCE_TYPE_MEMORY. */
-      struct {
-         /* The index of the tile buffer to use. */
-         uint32_t tile_buffer;
-
-         /* The offset in dwords within the tile buffer. */
-         uint32_t offset_dw;
-      } mem;
-   };
-};
-
-struct usc_mrt_setup {
-   /* Number of render targets present. */
-   uint32_t num_render_targets;
-
-   /* Number of output registers used per-pixel (1, 2 or 4). */
-   uint32_t num_output_regs;
-
-   /* Number of tile buffers used. */
-   uint32_t num_tile_buffers;
-
-   /* Size of a tile buffer in bytes. */
-   uint32_t tile_buffer_size;
-
-   /* Array of MRT resources allocated for each render target. The number of
-    * elements is determined by usc_mrt_setup::num_render_targets.
-    */
-   struct usc_mrt_resource *mrt_resources;
-
-   /* Don't set up source pos in emit. */
-   bool disable_source_pos_override;
-
-   /* Hash unique to this particular setup. */
-   uint32_t hash;
-};
-
 struct pvr_renderpass_hwsetup_eot_surface {
    /* MRT index to store from. Also used to index into
     * usc_mrt_setup::mrt_resources.

src/imagination/vulkan/pvr_mrt.c

@@ -0,0 +1,237 @@
+/*
+ * Copyright © 2025 Imagination Technologies Ltd.
+ *
+ * SPDX-License-Identifier: MIT
+ */
+
+#include "pvr_mrt.h"
+
+#include "vk_log.h"
+
+#include "pvr_device.h"
+#include "pvr_formats.h"
+#include "pvr_physical_device.h"
+
+#include "hwdef/rogue_hw_utils.h"
+#include "util/macros.h"
+
+#include "pvr_mrt.h"
+
+/* Which parts of the output registers/a tile buffer are currently allocated. */
+struct pvr_mrt_alloc_mask {
+   /* Bit array. A bit is set if the corresponding dword is allocated. */
+   BITSET_DECLARE(allocs, 8U);
+};
+
+struct pvr_mrt_alloc_ctx {
+   /* Which pixel output registers are allocated. */
+   struct pvr_mrt_alloc_mask output_reg;
+
+   /* Range of allocated output registers. */
+   uint32_t output_regs_count;
+
+   /* Number of tile buffers allocated. */
+   uint32_t tile_buffers_count;
+
+   /* Which parts of each tile buffer are allocated. Length is
+    * tile_buffers_count.
+    */
+   struct pvr_mrt_alloc_mask tile_buffers[PVR_MAX_TILE_BUFFER_COUNT];
+};
+
+static uint32_t pvr_get_accum_format_bitsize(VkFormat vk_format)
+{
+   return pvr_get_pbe_accum_format_size_in_bytes(vk_format) * 8;
+}
+
+/**
+ * Check if there is space in a buffer for storing a render target of a
+ * specified size.
+ */
+static int32_t
+pvr_mrt_alloc_from_buffer(const struct pvr_device_info *dev_info,
+                          struct pvr_mrt_alloc_mask *buffer,
+                          uint32_t pixel_size)
+{
+   const uint32_t max_out_regs = rogue_get_max_output_regs_per_pixel(dev_info);
+   uint32_t alignment = 1U;
+
+   if (PVR_HAS_FEATURE(dev_info, pbe2_in_xe)) {
+      /* For a 64-bit/128-bit source format: the start offset must be even. */
+      if (pixel_size == 2U || pixel_size == 4U)
+         alignment = 2U;
+   }
+
+   assert(pixel_size <= max_out_regs);
+
+   for (uint32_t i = 0U; i <= (max_out_regs - pixel_size); i += alignment) {
+      if (!BITSET_TEST_RANGE(buffer->allocs, i, i + pixel_size - 1U)) {
+         BITSET_SET_RANGE(buffer->allocs, i, i + pixel_size - 1U);
+         return i;
+      }
+   }
+
+   return -1;
+}
+
+void
+pvr_init_mrt_desc(VkFormat format, struct usc_mrt_desc *desc)
+{
+   uint32_t pixel_size_in_chunks;
+   uint32_t pixel_size_in_bits;
+
+   /* TODO: Add support for packing multiple attachments into the same
+    * register
+    */
+   const uint32_t part_bits = 0;
+   if (vk_format_is_color(format) &&
+       pvr_get_pbe_accum_format(format) == PVR_PBE_ACCUM_FORMAT_INVALID) {
+      /* The VkFormat is not supported as a color attachment so `0`.
+       * vulkan doesn't seem to restrict vkCreateRenderPass() to supported
+       * formats only.
+       */
+      pixel_size_in_bits = 0;
+   } else {
+      /* TODO: handle IMG_PIXFMT_A8_UNORM
+       *  For alpha only formats alpha is still placed in channel 3, so channels
+       *  0-2 need to be allocated but are left unused
+       */
+      pixel_size_in_bits = pvr_get_accum_format_bitsize(format);
+   }
+
+   desc->intermediate_size = DIV_ROUND_UP(pixel_size_in_bits, CHAR_BIT);
+
+   pixel_size_in_chunks = DIV_ROUND_UP(pixel_size_in_bits, 32U);
+   for (uint32_t j = 0U; j < pixel_size_in_chunks; j++)
+      desc->valid_mask[j] = ~0;
+
+   if (part_bits > 0U)
+      desc->valid_mask[pixel_size_in_chunks] = BITFIELD_MASK(part_bits);
+}
+
+static VkResult pvr_alloc_mrt(const struct pvr_device_info *dev_info,
+                              struct pvr_mrt_alloc_ctx *alloc,
+                              struct usc_mrt_setup *setup,
+                              unsigned rt,
+                              VkFormat format)
+{
+   struct usc_mrt_resource *resource = &setup->mrt_resources[rt];
+
+   const uint32_t pixel_size =
+      DIV_ROUND_UP(pvr_get_accum_format_bitsize(format), 32U);
+
+   const int32_t output_reg =
+      pvr_mrt_alloc_from_buffer(dev_info, &alloc->output_reg, pixel_size);
+
+   if (output_reg != -1) {
+      resource->type = USC_MRT_RESOURCE_TYPE_OUTPUT_REG;
+      resource->reg.output_reg = output_reg;
+      resource->reg.offset = 0;
+
+      alloc->output_regs_count =
+         MAX2(alloc->output_regs_count, resource->reg.output_reg + pixel_size);
+   } else {
+      resource->type = USC_MRT_RESOURCE_TYPE_MEMORY;
+
+      unsigned tib = 0;
+      for (; tib < alloc->tile_buffers_count; tib++) {
+         struct pvr_mrt_alloc_mask *tib_alloc =
+            &alloc->tile_buffers[tib];
+
+         const int32_t tile_buffer_offset =
+            pvr_mrt_alloc_from_buffer(dev_info, tib_alloc, pixel_size);
+
+         if (tile_buffer_offset != -1) {
+            resource->mem.tile_buffer = tib;
+            resource->mem.offset_dw = tile_buffer_offset;
+            break;
+         }
+      }
+
+      if (tib == alloc->tile_buffers_count) {
+         if (alloc->tile_buffers_count == PVR_MAX_TILE_BUFFER_COUNT)
+            return vk_error(NULL, VK_ERROR_TOO_MANY_OBJECTS);
+
+         resource->mem.tile_buffer = alloc->tile_buffers_count;
+         resource->mem.offset_dw = 0;
+      }
+
+      /* If needed a new tile buffer than those that were allocated, then wipe
+       * it bump the global count.
+       */
+      if (resource->mem.tile_buffer >= alloc->tile_buffers_count) {
+         memset(
+            &alloc->tile_buffers[alloc->tile_buffers_count],
+            0U,
+            sizeof(alloc->tile_buffers[0U]) *
+               (resource->mem.tile_buffer + 1U - alloc->tile_buffers_count));
+         alloc->tile_buffers_count = resource->mem.tile_buffer + 1U;
+      }
+
+      /* The hardware makes the bit depth of the on-chip storage and memory
+       * storage the same so make sure the memory storage is large enough to
+       * accommodate the largest render target.
+       */
+      alloc->output_regs_count =
+         MAX2(alloc->output_regs_count, resource->mem.offset_dw + pixel_size);
+   }
+
+   pvr_init_mrt_desc(format, &resource->mrt_desc);
+   resource->intermediate_size = resource->mrt_desc.intermediate_size;
+
+   setup->num_render_targets++;
+
+   return VK_SUCCESS;
+}
+
+VkResult
+pvr_init_usc_mrt_setup(struct pvr_device *device,
+                       uint32_t attachment_count,
+                       const VkFormat attachment_formats[attachment_count],
+                       struct usc_mrt_setup *setup)
+{
+   const struct pvr_device_info *dev_info = &device->pdevice->dev_info;
+   struct pvr_mrt_alloc_ctx alloc = { 0 };
+   VkResult result;
+
+   memset(setup, 0, sizeof(*setup));
+
+   if (!attachment_count)
+      goto early_exit;
+
+   setup->mrt_resources =
+      vk_alloc(&device->vk.alloc,
+               sizeof(*setup->mrt_resources) * attachment_count, 8U,
+               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (!setup->mrt_resources)
+      return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   for (unsigned i = 0; i < attachment_count; i++) {
+      VkFormat att_format = attachment_formats[i];
+      assert(att_format != VK_FORMAT_UNDEFINED);
+
+      result = pvr_alloc_mrt(dev_info, &alloc, setup, i, att_format);
+      if (result != VK_SUCCESS) {
+         result = vk_error(NULL, VK_ERROR_OUT_OF_DEVICE_MEMORY);
+         goto fail;
+      }
+   }
+
+early_exit:
+   setup->num_output_regs = util_next_power_of_two(alloc.output_regs_count);
+   setup->num_tile_buffers = alloc.tile_buffers_count;
+   return VK_SUCCESS;
+fail:
+   vk_free(&device->vk.alloc, setup->mrt_resources);
+   return result;
+}
+
+void
+pvr_destroy_mrt_setup(const struct pvr_device *device,
+                      struct usc_mrt_setup *setup)
+{
+   if (!setup)
+      return;
+
+   vk_free(&device->vk.alloc, setup->mrt_resources);
+}

src/imagination/vulkan/pvr_mrt.h

@@ -0,0 +1,108 @@
+/*
+ * Copyright © 2025 Imagination Technologies Ltd.
+ *
+ * SPDX-License-Identifier: MIT
+ */
+
+#ifndef PVR_MRT_H
+#define PVR_MRT_H
+
+#include <stdbool.h>
+#include <stdint.h>
+#include <vulkan/vulkan.h>
+
+struct pvr_device;
+
+/* Specifies the location of render target writes. */
+enum usc_mrt_resource_type {
+   USC_MRT_RESOURCE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid. */
+   USC_MRT_RESOURCE_TYPE_OUTPUT_REG,
+   USC_MRT_RESOURCE_TYPE_MEMORY,
+};
+
+#define PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS (4)
+
+struct usc_mrt_desc {
+   /* Size (in bytes) of the intermediate storage required for each pixel in the
+    * render target.
+    */
+   uint32_t intermediate_size;
+
+   /* Mask of the bits from each dword which are read by the PBE. */
+   uint32_t valid_mask[PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS];
+
+   /* Higher number = higher priority. Used to decide which render targets get
+    * allocated dedicated output registers.
+    */
+   uint32_t priority;
+};
+
+struct usc_mrt_resource {
+   /* Input description of render target. */
+   struct usc_mrt_desc mrt_desc;
+
+   /* Resource type allocated for render target. */
+   enum usc_mrt_resource_type type;
+
+   /* Intermediate pixel size (in bytes). */
+   uint32_t intermediate_size;
+
+   union {
+      /* If type == USC_MRT_RESOURCE_TYPE_OUTPUT_REG. */
+      struct {
+         /* The output register to use. */
+         uint32_t output_reg;
+
+         /* The offset in bytes into the output register. */
+         uint32_t offset;
+      } reg;
+
+      /* If type == USC_MRT_RESOURCE_TYPE_MEMORY. */
+      struct {
+         /* The index of the tile buffer to use. */
+         uint32_t tile_buffer;
+
+         /* The offset in dwords within the tile buffer. */
+         uint32_t offset_dw;
+      } mem;
+   };
+};
+
+struct usc_mrt_setup {
+   /* Number of render targets present. */
+   uint32_t num_render_targets;
+
+   /* Number of output registers used per-pixel (1, 2 or 4). */
+   uint32_t num_output_regs;
+
+   /* Number of tile buffers used. */
+   uint32_t num_tile_buffers;
+
+   /* Size of a tile buffer in bytes. */
+   uint32_t tile_buffer_size;
+
+   /* Array of MRT resources allocated for each render target. The number of
+    * elements is determined by usc_mrt_setup::num_render_targets.
+    */
+   struct usc_mrt_resource *mrt_resources;
+
+   /* Don't set up source pos in emit. */
+   bool disable_source_pos_override;
+
+   /* Hash unique to this particular setup. */
+   uint32_t hash;
+};
+
+VkResult
+pvr_init_usc_mrt_setup(struct pvr_device *device,
+                       uint32_t attachment_count,
+                       const VkFormat attachment_formats[attachment_count],
+                       struct usc_mrt_setup *setup);
+
+void
+pvr_destroy_mrt_setup(const struct pvr_device *device,
+                      struct usc_mrt_setup *setup);
+
+void pvr_init_mrt_desc(VkFormat format, struct usc_mrt_desc *desc);
+
+#endif