anv: device: calculate compute thread numbers using subslices numbers

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

6 files changed, 74 insertions, 18 deletions

diff --git a/src/intel/vulkan/anv_allocator.c b/src/intel/vulkan/anv_allocator.c
index c1687b9897..f694aee92d 100644
--- a/src/intel/vulkan/anv_allocator.c
+++ b/src/intel/vulkan/anv_allocator.c
@@ -924,14 +924,15 @@ anv_scratch_pool_alloc(struct anv_device *device, struct anv_scratch_pool *pool,
    if (size == 0) {
       /* We own the lock.  Allocate a buffer */
 
-      struct gen_device_info *devinfo = &device->info;
+      struct anv_physical_device *physical_device =
+         &device->instance->physicalDevice;
       uint32_t max_threads[] = {
-         [MESA_SHADER_VERTEX]                  = devinfo->max_vs_threads,
-         [MESA_SHADER_TESS_CTRL]               = devinfo->max_hs_threads,
-         [MESA_SHADER_TESS_EVAL]               = devinfo->max_ds_threads,
-         [MESA_SHADER_GEOMETRY]                = devinfo->max_gs_threads,
-         [MESA_SHADER_FRAGMENT]                = devinfo->max_wm_threads,
-         [MESA_SHADER_COMPUTE]                 = devinfo->max_cs_threads,
+         [MESA_SHADER_VERTEX]           = physical_device->max_vs_threads,
+         [MESA_SHADER_TESS_CTRL]        = physical_device->max_hs_threads,
+         [MESA_SHADER_TESS_EVAL]        = physical_device->max_ds_threads,
+         [MESA_SHADER_GEOMETRY]         = physical_device->max_gs_threads,
+         [MESA_SHADER_FRAGMENT]         = physical_device->max_wm_threads,
+         [MESA_SHADER_COMPUTE]          = physical_device->max_cs_threads,
       };
 
       size = per_thread_scratch * max_threads[stage];
diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index e66f81252d..fecb8505b8 100644
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -136,6 +136,41 @@ anv_physical_device_init(struct anv_physical_device *device,
 
    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 
+   device->max_vs_threads = device->info->max_vs_threads;
+   device->max_hs_threads = device->info->max_hs_threads;
+   device->max_ds_threads = device->info->max_ds_threads;
+   device->max_gs_threads = device->info->max_gs_threads;
+   device->max_wm_threads = device->info->max_wm_threads;
+
+   /* GENs prior to 8 do not support EU/Subslice info */
+   if (device->info->gen >= 8) {
+      device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
+      device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
+
+      /* Without this information, we cannot get the right Braswell
+       * brandstrings, and we have to use conservative numbers for GPGPU on
+       * many platforms, but otherwise, things will just work.
+       */
+      if (device->subslice_total < 1 || device->eu_total < 1) {
+         fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
+                         " query GPU properties.\n");
+      }
+   } else if (device->info->gen == 7) {
+      device->subslice_total = 1 << (device->info->gt - 1);
+   }
+
+   if (device->info->is_cherryview &&
+       device->subslice_total > 0 && device->eu_total > 0) {
+      /* Logical CS threads = EUs per subslice * 7 threads per EU */
+      device->max_cs_threads = device->eu_total / device->subslice_total * 7;
+
+      /* Fuse configurations may give more threads than expected, never less. */
+      if (device->max_cs_threads < device->info->max_cs_threads)
+         device->max_cs_threads = device->info->max_cs_threads;
+   } else {
+      device->max_cs_threads = device->info->max_cs_threads;
+   }
+
    close(fd);
 
    brw_process_intel_debug_variable();
@@ -503,11 +538,11 @@ void anv_GetPhysicalDeviceProperties(
       .maxFragmentCombinedOutputResources       = 8,
       .maxComputeSharedMemorySize               = 32768,
       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
-      .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
+      .maxComputeWorkGroupInvocations           = 16 * pdevice->max_cs_threads,
       .maxComputeWorkGroupSize = {
-         16 * devinfo->max_cs_threads,
-         16 * devinfo->max_cs_threads,
-         16 * devinfo->max_cs_threads,
+         16 * pdevice->max_cs_threads,
+         16 * pdevice->max_cs_threads,
+         16 * pdevice->max_cs_threads,
       },
       .subPixelPrecisionBits                    = 4 /* FIXME */,
       .subTexelPrecisionBits                    = 4 /* FIXME */,
diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index a083f93fc6..f578a9d9a8 100644
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -570,6 +570,20 @@ struct anv_physical_device {
     struct isl_device                           isl_dev;
     int                                         cmd_parser_version;
 
+    uint32_t                                    eu_total;
+    uint32_t                                    subslice_total;
+
+    /**
+     * Platform specific constants containing the maximum number of threads
+     * for each pipeline stage.
+     */
+    uint32_t                                    max_vs_threads;
+    uint32_t                                    max_hs_threads;
+    uint32_t                                    max_ds_threads;
+    uint32_t                                    max_gs_threads;
+    uint32_t                                    max_wm_threads;
+    uint32_t                                    max_cs_threads;
+
     struct anv_wsi_interface *                  wsi[VK_ICD_WSI_PLATFORM_MAX];
 };
 
diff --git a/src/intel/vulkan/gen7_pipeline.c b/src/intel/vulkan/gen7_pipeline.c
index 48457aa9a3..878308b7fb 100644
--- a/src/intel/vulkan/gen7_pipeline.c
+++ b/src/intel/vulkan/gen7_pipeline.c
@@ -45,6 +45,8 @@ genX(graphics_pipeline_create)(
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
    ANV_FROM_HANDLE(anv_render_pass, pass, pCreateInfo->renderPass);
+   struct anv_physical_device *physical_device =
+      &device->instance->physicalDevice;
    struct anv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
    struct anv_pipeline *pipeline;
    VkResult result;
@@ -123,7 +125,7 @@ genX(graphics_pipeline_create)(
 
          vs.VertexURBEntryReadLength   = vs_prog_data->base.urb_read_length;
          vs.VertexURBEntryReadOffset   = 0;
-         vs.MaximumNumberofThreads     = device->info.max_vs_threads - 1;
+         vs.MaximumNumberofThreads     = physical_device->max_vs_threads - 1;
          vs.StatisticsEnable           = true;
          vs.VSFunctionEnable           = true;
       }
@@ -152,7 +154,7 @@ genX(graphics_pipeline_create)(
          gs.DispatchGRFStartRegisterforURBData =
             gs_prog_data->base.base.dispatch_grf_start_reg;
 
-         gs.MaximumNumberofThreads     = device->info.max_gs_threads - 1;
+         gs.MaximumNumberofThreads     = physical_device->max_gs_threads - 1;
          /* This in the next dword on HSW. */
          gs.ControlDataFormat          = gs_prog_data->control_data_format;
          gs.ControlDataHeaderSize      = gs_prog_data->control_data_header_size_hwords;
@@ -185,7 +187,7 @@ genX(graphics_pipeline_create)(
        * don't at least set the maximum number of threads.
        */
       anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS), ps) {
-         ps.MaximumNumberofThreads = device->info.max_wm_threads - 1;
+         ps.MaximumNumberofThreads = physical_device->max_wm_threads - 1;
       }
    } else {
       const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
@@ -207,7 +209,7 @@ genX(graphics_pipeline_create)(
             .offset = 0,
          };
          ps.PerThreadScratchSpace         = scratch_space(&wm_prog_data->base);
-         ps.MaximumNumberofThreads        = device->info.max_wm_threads - 1;
+         ps.MaximumNumberofThreads        = physical_device->max_wm_threads - 1;
          ps.PushConstantEnable            = wm_prog_data->base.nr_params > 0;
          ps.AttributeEnable               = wm_prog_data->num_varying_inputs > 0;
          ps.oMaskPresenttoRenderTarget    = wm_prog_data->uses_omask;
diff --git a/src/intel/vulkan/gen8_pipeline.c b/src/intel/vulkan/gen8_pipeline.c
index 445177a84f..9f16960ccc 100644
--- a/src/intel/vulkan/gen8_pipeline.c
+++ b/src/intel/vulkan/gen8_pipeline.c
@@ -55,6 +55,8 @@ genX(graphics_pipeline_create)(
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
    ANV_FROM_HANDLE(anv_render_pass, pass, pCreateInfo->renderPass);
+   struct anv_physical_device *physical_device =
+      &device->instance->physicalDevice;
    struct anv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
    struct anv_pipeline *pipeline;
    VkResult result;
@@ -142,7 +144,7 @@ genX(graphics_pipeline_create)(
          gs.DispatchGRFStartRegisterForURBData =
             gs_prog_data->base.base.dispatch_grf_start_reg;
 
-         gs.MaximumNumberofThreads  = device->info.max_gs_threads / 2 - 1;
+         gs.MaximumNumberofThreads  = physical_device->max_gs_threads / 2 - 1;
          gs.ControlDataHeaderSize   = gs_prog_data->control_data_header_size_hwords;
          gs.DispatchMode            = gs_prog_data->base.dispatch_mode;
          gs.StatisticsEnable        = true;
@@ -213,7 +215,7 @@ genX(graphics_pipeline_create)(
          vs.VertexURBEntryReadLength      = vs_prog_data->base.urb_read_length;
          vs.VertexURBEntryReadOffset      = 0;
 
-         vs.MaximumNumberofThreads        = device->info.max_vs_threads - 1;
+         vs.MaximumNumberofThreads        = physical_device->max_vs_threads - 1;
          vs.StatisticsEnable              = false;
          vs.SIMD8DispatchEnable           = pipeline->vs_simd8 != NO_KERNEL;
          vs.VertexCacheDisable            = false;
diff --git a/src/intel/vulkan/genX_pipeline.c b/src/intel/vulkan/genX_pipeline.c
index 5a3e1ab252..3ecc29e953 100644
--- a/src/intel/vulkan/genX_pipeline.c
+++ b/src/intel/vulkan/genX_pipeline.c
@@ -35,6 +35,8 @@ genX(compute_pipeline_create)(
     VkPipeline*                                 pPipeline)
 {
    ANV_FROM_HANDLE(anv_device, device, _device);
+   struct anv_physical_device *physical_device =
+      &device->instance->physicalDevice;
    struct anv_pipeline *pipeline;
    VkResult result;
 
@@ -115,7 +117,7 @@ genX(compute_pipeline_create)(
 #else
       vfe.GPGPUMode              = true;
 #endif
-      vfe.MaximumNumberofThreads = device->info.max_cs_threads - 1;
+      vfe.MaximumNumberofThreads = physical_device->max_cs_threads - 1;
       vfe.NumberofURBEntries     = GEN_GEN <= 7 ? 0 : 2;
       vfe.ResetGatewayTimer      = true;
 #if GEN_GEN <= 8