ac/gpu_info: replace ib_alignment with per-IP IB base and size alignments

Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25043>

diff --git a/src/amd/common/ac_gpu_info.c b/src/amd/common/ac_gpu_info.c
index 4706edf..f3d7ae0 100644 (file)
--- a/src/amd/common/ac_gpu_info.c
+++ b/src/amd/common/ac_gpu_info.c
@@ -677,8 +677,8 @@ bool ac_query_gpu_info(int fd, void *dev_p, struct radeon_info *info,
             info->ip[AMD_IP_GFX].ver_minor = info->ip[AMD_IP_COMPUTE].ver_minor = 3;
       }
       info->ip[ip_type].num_queues = util_bitcount(ip_info.available_rings);
-      info->ib_alignment = MAX3(info->ib_alignment, ip_info.ib_start_alignment,
-                                ip_info.ib_size_alignment);
+      info->ip[ip_type].ib_base_alignment = ip_info.ib_start_alignment;
+      info->ip[ip_type].ib_size_alignment = ip_info.ib_size_alignment;
    }
 
    /* Only require gfx or compute. */
@@ -690,12 +690,6 @@ bool ac_query_gpu_info(int fd, void *dev_p, struct radeon_info *info,
    assert(util_is_power_of_two_or_zero(info->ip[AMD_IP_COMPUTE].num_queues));
    assert(util_is_power_of_two_or_zero(info->ip[AMD_IP_SDMA].num_queues));
 
-   /* The kernel pads gfx and compute IBs to 256 dwords since:
-    *   66f3b2d527154bd258a57c8815004b5964aa1cf5
-    * Do the same.
-    */
-   info->ib_alignment = MAX2(info->ib_alignment, 1024);
-
    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_ME, 0, 0, &info->me_fw_version,
                                      &info->me_fw_feature);
    if (r) {
@@ -1681,8 +1675,11 @@ void ac_print_gpu_info(const struct radeon_info *info, FILE *f)
 
    for (unsigned i = 0; i < AMD_NUM_IP_TYPES; i++) {
       if (info->ip[i].num_queues) {
-         fprintf(f, "    IP %-7s %2u.%u \tqueues:%u\n", ip_string[i],
-                 info->ip[i].ver_major, info->ip[i].ver_minor, info->ip[i].num_queues);
+         fprintf(f, "    IP %-7s %2u.%u   queues:%u   "
+                    "align(base:%u, size:%u)\n",
+                 ip_string[i], info->ip[i].ver_major, info->ip[i].ver_minor,
+                 info->ip[i].num_queues, info->ip[i].ib_base_alignment,
+                 info->ip[i].ib_size_alignment);
       }
    }
 
@@ -1756,7 +1753,6 @@ void ac_print_gpu_info(const struct radeon_info *info, FILE *f)
 
    fprintf(f, "CP info:\n");
    fprintf(f, "    gfx_ib_pad_with_type2 = %i\n", info->gfx_ib_pad_with_type2);
-   fprintf(f, "    ib_alignment = %u\n", info->ib_alignment);
    fprintf(f, "    me_fw_version = %i\n", info->me_fw_version);
    fprintf(f, "    me_fw_feature = %i\n", info->me_fw_feature);
    fprintf(f, "    mec_fw_version = %i\n", info->mec_fw_version);

diff --git a/src/amd/common/ac_gpu_info.h b/src/amd/common/ac_gpu_info.h
index 1e9e871..30f4bdd 100644 (file)
--- a/src/amd/common/ac_gpu_info.h
+++ b/src/amd/common/ac_gpu_info.h
@@ -26,6 +26,8 @@ struct amd_ip_info {
    uint8_t ver_minor;
    uint8_t ver_rev;
    uint8_t num_queues;
+   uint16_t ib_base_alignment;
+   uint16_t ib_size_alignment;
 };
 
 struct radeon_info {
@@ -160,7 +162,6 @@ struct radeon_info {
 
    /* CP info. */
    bool gfx_ib_pad_with_type2;
-   unsigned ib_alignment; /* both start and size alignment */
    uint32_t me_fw_version;
    uint32_t me_fw_feature;
    uint32_t mec_fw_version;

diff --git a/src/amd/vulkan/radv_device_generated_commands.c b/src/amd/vulkan/radv_device_generated_commands.c
index 18c0c6a..e085a63 100644 (file)
--- a/src/amd/vulkan/radv_device_generated_commands.c
+++ b/src/amd/vulkan/radv_device_generated_commands.c
@@ -1371,11 +1371,14 @@ radv_GetGeneratedCommandsMemoryRequirementsNV(VkDevice _device,
    VkDeviceSize cmd_buf_size =
       radv_align_cmdbuf_size(device, cmd_stride * pInfo->maxSequencesCount) + radv_dgc_preamble_cmdbuf_size(device);
    VkDeviceSize upload_buf_size = upload_stride * pInfo->maxSequencesCount;
+   unsigned ib_base_alignment = MAX2(device->physical_device->rad_info.ip[AMD_IP_GFX].ib_base_alignment,
+                                     device->physical_device->rad_info.ip[AMD_IP_COMPUTE].ib_base_alignment);
+   unsigned ib_size_alignment = MAX2(device->physical_device->rad_info.ip[AMD_IP_GFX].ib_size_alignment,
+                                     device->physical_device->rad_info.ip[AMD_IP_COMPUTE].ib_size_alignment);
 
    pMemoryRequirements->memoryRequirements.memoryTypeBits = device->physical_device->memory_types_32bit;
-   pMemoryRequirements->memoryRequirements.alignment = device->physical_device->rad_info.ib_alignment;
-   pMemoryRequirements->memoryRequirements.size =
-      align(cmd_buf_size + upload_buf_size, pMemoryRequirements->memoryRequirements.alignment);
+   pMemoryRequirements->memoryRequirements.alignment = ib_base_alignment;
+   pMemoryRequirements->memoryRequirements.size = align(cmd_buf_size + upload_buf_size, ib_size_alignment);
 }
 
 VKAPI_ATTR void VKAPI_CALL

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index a98b538..0420c85 100644 (file)
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -248,8 +248,9 @@ radv_amdgpu_cs_bo_create(struct radv_amdgpu_cs *cs, uint32_t ib_size)
    const enum radeon_bo_flag flags =
       RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_READ_ONLY | gtt_wc_flag;
 
-   return ws->buffer_create(ws, ib_size, cs->ws->info.ib_alignment, domain, flags, RADV_BO_PRIORITY_CS, 0,
-                            &cs->ib_buffer);
+   ib_size = align(ib_size, cs->ws->info.ip[cs->ib.ip_type].ib_size_alignment);
+   return ws->buffer_create(ws, ib_size, cs->ws->info.ip[cs->ib.ip_type].ib_base_alignment, domain, flags,
+                            RADV_BO_PRIORITY_CS, 0, &cs->ib_buffer);
 }
 
 static VkResult
@@ -1670,8 +1671,8 @@ radv_amdgpu_cs_submit(struct radv_amdgpu_ctx *ctx, struct radv_amdgpu_cs_request
       chunks[i].chunk_data = (uint64_t)(uintptr_t)&chunk_data[i];
 
       ib = &request->ibs[i];
-      assert(ib->ib_mc_address && ib->ib_mc_address % ctx->ws->info.ib_alignment == 0);
-      assert(ib->size);
+      assert(ib->ib_mc_address && ib->ib_mc_address % ctx->ws->info.ip[ib->ip_type].ib_base_alignment == 0);
+      assert(ib->size && (ib->size * 4) % ctx->ws->info.ip[ib->ip_type].ib_size_alignment == 0);
 
       chunk_data[i].ib_data._pad = 0;
       chunk_data[i].ib_data.va_start = ib->ib_mc_address;

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index c70c691..fd48e6c 100644 (file)
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -953,9 +953,11 @@ static void amdgpu_set_ib_size(struct radeon_cmdbuf *rcs, struct amdgpu_ib *ib)
 static void amdgpu_ib_finalize(struct amdgpu_winsys *ws, struct radeon_cmdbuf *rcs,
                                struct amdgpu_ib *ib)
 {
+   struct amdgpu_cs *cs = (struct amdgpu_cs*)ib;
+
    amdgpu_set_ib_size(rcs, ib);
    ib->used_ib_space += rcs->current.cdw * 4;
-   ib->used_ib_space = align(ib->used_ib_space, ws->info.ib_alignment);
+   ib->used_ib_space = align(ib->used_ib_space, ws->info.ip[cs->ip_type].ib_base_alignment);
    ib->max_ib_size = MAX2(ib->max_ib_size, rcs->prev_dw + rcs->current.cdw);
 }
 
@@ -1145,12 +1147,12 @@ amdgpu_cs_setup_preemption(struct radeon_cmdbuf *rcs, const uint32_t *preamble_i
    struct amdgpu_cs *cs = amdgpu_cs(rcs);
    struct amdgpu_winsys *ws = cs->ws;
    struct amdgpu_cs_context *csc[2] = {&cs->csc1, &cs->csc2};
-   unsigned size = align(preamble_num_dw * 4, ws->info.ib_alignment);
+   unsigned size = align(preamble_num_dw * 4, ws->info.ip[cs->ip_type].ib_size_alignment);
    struct pb_buffer *preamble_bo;
    uint32_t *map;
 
    /* Create the preamble IB buffer. */
-   preamble_bo = amdgpu_bo_create(ws, size, ws->info.ib_alignment,
+   preamble_bo = amdgpu_bo_create(ws, size, ws->info.ip[cs->ip_type].ib_base_alignment,
                                   RADEON_DOMAIN_VRAM,
                                   RADEON_FLAG_NO_INTERPROCESS_SHARING |
                                   RADEON_FLAG_GTT_WC |
@@ -1708,7 +1710,7 @@ static void amdgpu_cs_submit_ib(void *job, void *gdata, int thread_index)
 
    if (noop && acs->ip_type == AMD_IP_GFX) {
       /* Reduce the IB size and fill it with NOP to make it like an empty IB. */
-      unsigned noop_size = MIN2(cs->ib[IB_MAIN].ib_bytes, ws->info.ib_alignment);
+      unsigned noop_size = MIN2(cs->ib[IB_MAIN].ib_bytes, ws->info.ip[AMD_IP_GFX].ib_size_alignment);
 
       cs->ib_main_addr[0] = PKT3(PKT3_NOP, noop_size / 4 - 2, 0);
       cs->ib[IB_MAIN].ib_bytes = noop_size;

diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
index 18eb60b..990b8e9 100644 (file)
--- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
@@ -551,7 +551,11 @@ static bool do_winsys_init(struct radeon_drm_winsys *ws)
                                     (ws->info.family == CHIP_HAWAII &&
                                      ws->accel_working2 < 3);
    ws->info.tcc_cache_line_size = 64; /* TC L2 line size on GCN */
-   ws->info.ib_alignment = 4096;
+   for (unsigned i = 0; i < AMD_NUM_IP_TYPES; i++) {
+      /* This is probably too large. */
+      ws->info.ip[i].ib_base_alignment = 4096;
+      ws->info.ip[i].ib_size_alignment = 4096;
+   }
    ws->info.has_bo_metadata = false;
    ws->info.has_eqaa_surface_allocator = false;
    ws->info.has_sparse_vm_mappings = false;