intel: Add INTEL_DEBUG=no-vrt

Add support for disabling the VRT (Variable Register Thread) feature.
The strategy here is to force the old BRW_MAX_GRF limit for the
register allocator (locks the upper limit) and make sure
ptl_register_blocks() always return that amount of blocks (locks
the lower limit).

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35781>

src/gallium/drivers/iris/iris_state.c

@@ -1413,7 +1413,7 @@ iris_init_render_context(struct iris_batch *batch)
 #if GFX_VER >= 30
    iris_emit_cmd(batch, GENX(STATE_COMPUTE_MODE), cm) {
       cm.EnableVariableRegisterSizeAllocationMask = 1;
-      cm.EnableVariableRegisterSizeAllocation = true;
+      cm.EnableVariableRegisterSizeAllocation = !INTEL_DEBUG(DEBUG_NO_VRT);
    }
 #endif
 
@@ -1549,7 +1549,7 @@ iris_init_compute_context(struct iris_batch *batch)
    iris_emit_cmd(batch, GENX(STATE_COMPUTE_MODE), cm) {
 #if GFX_VER >= 30
       cm.EnableVariableRegisterSizeAllocationMask = 1;
-      cm.EnableVariableRegisterSizeAllocation = true;
+      cm.EnableVariableRegisterSizeAllocation = !INTEL_DEBUG(DEBUG_NO_VRT);
 #endif
 #if GFX_VER >= 20
       cm.AsyncComputeThreadLimit = pixel_async_compute_thread_limit;

src/intel/common/intel_common.c

@@ -5,6 +5,8 @@
 
 #include <stdlib.h>
 
+#include "dev/intel_debug.h"
+
 #include "intel_common.h"
 
 #include "intel_engine.h"
@@ -47,7 +49,7 @@ intel_compute_engine_async_threads_limit(const struct intel_device_info *devinfo
    uint8_t pixel_async_compute_thread_limit = 2;
    uint8_t z_pass_async_compute_thread_limit = 0;
    uint8_t np_z_async_throttle_settings = 0;
-   bool has_vrt = devinfo->verx10 >= 300;
+   bool has_vrt = devinfo->verx10 >= 300 && !INTEL_DEBUG(DEBUG_NO_VRT);
 
    /* When VRT is enabled async threads limits don't have effect */
    if (!slm_or_barrier_enabled || has_vrt) {

src/intel/compiler/brw_compiler.c

@@ -253,6 +253,7 @@ brw_get_compiler_config_value(const struct brw_compiler *compiler)
       DEBUG_DO32,
       DEBUG_SOFT64,
       DEBUG_NO_SEND_GATHER,
+      DEBUG_NO_VRT,
    };
    for (uint32_t i = 0; i < ARRAY_SIZE(debug_bits); i++) {
       insert_u64_bit(&config, INTEL_DEBUG(debug_bits[i]));
@@ -385,3 +386,13 @@ brw_stage_prog_data_add_printf(struct brw_stage_prog_data *prog_data,
              print->arg_sizes, sizeof(print->arg_sizes[0]) *print->num_args);
    }
 }
+
+unsigned
+ptl_register_blocks(unsigned grf_used)
+{
+   if (INTEL_DEBUG(DEBUG_NO_VRT))
+      return (BRW_MAX_GRF / 32) - 1;
+
+   const unsigned n = DIV_ROUND_UP(grf_used, 32) - 1;
+   return (n < 6 ? n : 7);
+}

src/intel/compiler/brw_compiler.h

@@ -652,12 +652,7 @@ struct brw_stage_prog_data {
  * Convert a number of GRF registers used (grf_used in prog_data) into
  * a number of GRF register blocks supported by the hardware on PTL+.
  */
-static inline unsigned
-ptl_register_blocks(unsigned grf_used)
-{
-   const unsigned n = DIV_ROUND_UP(grf_used, 32) - 1;
-   return (n < 6 ? n : 7);
-}
+unsigned ptl_register_blocks(unsigned grf_used);
 
 static inline uint32_t *
 brw_stage_prog_data_add_params(struct brw_stage_prog_data *prog_data,

src/intel/compiler/brw_reg_allocate.cpp

@@ -29,6 +29,7 @@
 #include "brw_shader.h"
 #include "brw_builder.h"
 #include "brw_cfg.h"
+#include "dev/intel_debug.h"
 #include "util/set.h"
 #include "util/register_allocate.h"
 
@@ -80,7 +81,8 @@ extern "C" void
 brw_alloc_reg_sets(struct brw_compiler *compiler)
 {
    const struct intel_device_info *devinfo = compiler->devinfo;
-   int base_reg_count = (devinfo->ver >= 30 ? XE3_MAX_GRF / reg_unit(devinfo) :
+   int base_reg_count = (devinfo->ver >= 30 && !INTEL_DEBUG(DEBUG_NO_VRT) ?
+                         XE3_MAX_GRF / reg_unit(devinfo) :
                          BRW_MAX_GRF);
 
    /* The registers used to make up almost all values handled in the compiler

src/intel/dev/intel_debug.c

@@ -121,6 +121,7 @@ static const struct debug_control_bitset debug_control[] = {
    OPT1("shader-print",      DEBUG_SHADER_PRINT),
    OPT1("cl-quiet",          DEBUG_CL_QUIET),
    OPT1("no-send-gather",    DEBUG_NO_SEND_GATHER),
+   OPT1("no-vrt",            DEBUG_NO_VRT),
    OPT1("shaders-lineno",    DEBUG_SHADERS_LINENO),
    OPT1("show_shader_stage", DEBUG_SHOW_SHADER_STAGE),
    { NULL, }

src/intel/dev/intel_debug.h

@@ -92,6 +92,7 @@ enum intel_debug_flag {
    DEBUG_BVH_TLAS_IR_AS,
    DEBUG_BVH_NO_BUILD,
    DEBUG_NO_SEND_GATHER,
+   DEBUG_NO_VRT,
    DEBUG_RT_NO_TRACE,
    DEBUG_SHADERS_LINENO,
    DEBUG_SHOW_SHADER_STAGE,

src/intel/vulkan/genX_init_state.c

@@ -636,7 +636,7 @@ init_render_queue_state(struct anv_queue *queue, bool is_companion_rcs_batch)
 #if GFX_VERx10 >= 125
    anv_batch_emit(batch, GENX(STATE_COMPUTE_MODE), cm) {
 #if GFX_VER >= 30
-      cm.EnableVariableRegisterSizeAllocation = true;
+      cm.EnableVariableRegisterSizeAllocation = !INTEL_DEBUG(DEBUG_NO_VRT);
 #endif
       cm.Mask1 = 0xffff;
 #if GFX_VERx10 >= 200
@@ -774,7 +774,7 @@ init_compute_queue_state(struct anv_queue *queue)
    anv_batch_emit(batch, GENX(STATE_COMPUTE_MODE), cm) {
 #if GFX_VER >= 30
       cm.EnableVariableRegisterSizeAllocationMask = 1;
-      cm.EnableVariableRegisterSizeAllocation = true;
+      cm.EnableVariableRegisterSizeAllocation = !INTEL_DEBUG(DEBUG_NO_VRT);
 #endif
 #if GFX_VER >= 20
       cm.AsyncComputeThreadLimit = ACTL_Max8;