intel/fs: Rework KSP data to be SIMD width-based

Reviewed-by: Matt Turner <[email protected]>

3 files changed, 43 insertions, 47 deletions

diff --git a/src/intel/compiler/brw_compiler.h b/src/intel/compiler/brw_compiler.h
index a8ae243e9f5..d3d1f35d016 100644
--- a/src/intel/compiler/brw_compiler.h
+++ b/src/intel/compiler/brw_compiler.h
@@ -683,11 +683,11 @@ struct brw_wm_prog_data {
 
    GLuint num_varying_inputs;
 
-   uint8_t reg_blocks_0;
-   uint8_t reg_blocks_2;
+   uint8_t reg_blocks_8;
+   uint8_t reg_blocks_16;
 
-   uint8_t dispatch_grf_start_reg_2;
-   uint32_t prog_offset_2;
+   uint8_t dispatch_grf_start_reg_16;
+   uint32_t prog_offset_16;
 
    struct {
       /** @{
@@ -784,51 +784,48 @@ brw_fs_simd_width_for_ksp(unsigned ksp_idx, bool simd8_enabled,
 
 static inline uint32_t
 _brw_wm_prog_data_prog_offset(const struct brw_wm_prog_data *prog_data,
-                              unsigned ksp_idx)
+                              unsigned simd_width)
 {
-   switch (ksp_idx) {
-   case 0: return 0;
-   case 1: return 0;
-   case 2: return prog_data->prog_offset_2;
-   default:
-      unreachable("Invalid KSP index");
+   switch (simd_width) {
+   case 8: return 0;
+   case 16: return prog_data->prog_offset_16;
+   default: return 0;
    }
 }
 
 #define brw_wm_prog_data_prog_offset(prog_data, wm_state, ksp_idx) \
-   _brw_wm_prog_data_prog_offset(prog_data, ksp_idx)
+   _brw_wm_prog_data_prog_offset(prog_data, \
+      brw_wm_state_simd_width_for_ksp(wm_state, ksp_idx))
 
 static inline uint8_t
 _brw_wm_prog_data_dispatch_grf_start_reg(const struct brw_wm_prog_data *prog_data,
-                                         unsigned ksp_idx)
+                                         unsigned simd_width)
 {
-   switch (ksp_idx) {
-   case 0: return prog_data->base.dispatch_grf_start_reg;
-   case 1: return 0;
-   case 2: return prog_data->dispatch_grf_start_reg_2;
-   default:
-      unreachable("Invalid KSP index");
+   switch (simd_width) {
+   case 8: return prog_data->base.dispatch_grf_start_reg;
+   case 16: return prog_data->dispatch_grf_start_reg_16;
+   default: return 0;
    }
 }
 
 #define brw_wm_prog_data_dispatch_grf_start_reg(prog_data, wm_state, ksp_idx) \
-   _brw_wm_prog_data_dispatch_grf_start_reg(prog_data, ksp_idx)
+   _brw_wm_prog_data_dispatch_grf_start_reg(prog_data, \
+      brw_wm_state_simd_width_for_ksp(wm_state, ksp_idx))
 
 static inline uint8_t
 _brw_wm_prog_data_reg_blocks(const struct brw_wm_prog_data *prog_data,
-                             unsigned ksp_idx)
+                             unsigned simd_width)
 {
-   switch (ksp_idx) {
-   case 0: return prog_data->reg_blocks_0;
-   case 1: return 0;
-   case 2: return prog_data->reg_blocks_2;
-   default:
-      unreachable("Invalid KSP index");
+   switch (simd_width) {
+   case 8: return prog_data->reg_blocks_8;
+   case 16: return prog_data->reg_blocks_16;
+   default: return 0;
    }
 }
 
 #define brw_wm_prog_data_reg_blocks(prog_data, wm_state, ksp_idx) \
-   _brw_wm_prog_data_reg_blocks(prog_data, ksp_idx)
+   _brw_wm_prog_data_reg_blocks(prog_data, \
+      brw_wm_state_simd_width_for_ksp(wm_state, ksp_idx))
 
 struct brw_push_const_block {
    unsigned dwords;     /* Dword count, not reg aligned */
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 5361e5cccf3..ee0d1967ecc 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -7099,8 +7099,6 @@ brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
       brw_compute_barycentric_interp_modes(compiler->devinfo, shader);
 
    cfg_t *simd8_cfg = NULL, *simd16_cfg = NULL;
-   uint8_t simd8_grf_start = 0, simd16_grf_start = 0;
-   unsigned simd8_grf_used = 0, simd16_grf_used = 0;
 
    fs_visitor v8(compiler, log_data, mem_ctx, key,
                  &prog_data->base, prog, shader, 8,
@@ -7112,8 +7110,8 @@ brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
       return NULL;
    } else if (likely(!(INTEL_DEBUG & DEBUG_NO8))) {
       simd8_cfg = v8.cfg;
-      simd8_grf_start = v8.payload.num_regs;
-      simd8_grf_used = v8.grf_used;
+      prog_data->base.dispatch_grf_start_reg = v8.payload.num_regs;
+      prog_data->reg_blocks_8 = brw_register_blocks(v8.grf_used);
    }
 
    if (v8.max_dispatch_width >= 16 &&
@@ -7129,8 +7127,8 @@ brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
                                    v16.fail_msg);
       } else {
          simd16_cfg = v16.cfg;
-         simd16_grf_start = v16.payload.num_regs;
-         simd16_grf_used = v16.grf_used;
+         prog_data->dispatch_grf_start_reg_16 = v16.payload.num_regs;
+         prog_data->reg_blocks_16 = brw_register_blocks(v16.grf_used);
       }
    }
 
@@ -7146,6 +7144,16 @@ brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
    if (compiler->devinfo->gen < 5 && simd16_cfg)
       simd8_cfg = NULL;
 
+   if (compiler->devinfo->gen <= 5 && !simd8_cfg) {
+      /* Iron lake and earlier only have one Dispatch GRF start field.  Make
+       * the data available in the base prog data struct for convenience.
+       */
+      if (simd16_cfg) {
+         prog_data->base.dispatch_grf_start_reg =
+            prog_data->dispatch_grf_start_reg_16;
+      }
+   }
+
    if (prog_data->persample_dispatch) {
       /* Starting with SandyBridge (where we first get MSAA), the different
        * pixel dispatch combinations are grouped into classifications A
@@ -7184,20 +7192,11 @@ brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
    if (simd8_cfg) {
       prog_data->dispatch_8 = true;
       g.generate_code(simd8_cfg, 8);
-      prog_data->base.dispatch_grf_start_reg = simd8_grf_start;
-      prog_data->reg_blocks_0 = brw_register_blocks(simd8_grf_used);
+   }
 
-      if (simd16_cfg) {
-         prog_data->dispatch_16 = true;
-         prog_data->prog_offset_2 = g.generate_code(simd16_cfg, 16);
-         prog_data->dispatch_grf_start_reg_2 = simd16_grf_start;
-         prog_data->reg_blocks_2 = brw_register_blocks(simd16_grf_used);
-      }
-   } else if (simd16_cfg) {
+   if (simd16_cfg) {
       prog_data->dispatch_16 = true;
-      g.generate_code(simd16_cfg, 16);
-      prog_data->base.dispatch_grf_start_reg = simd16_grf_start;
-      prog_data->reg_blocks_0 = brw_register_blocks(simd16_grf_used);
+      prog_data->prog_offset_16 = g.generate_code(simd16_cfg, 16);
    }
 
    return g.get_assembly();
diff --git a/src/intel/compiler/brw_fs_visitor.cpp b/src/intel/compiler/brw_fs_visitor.cpp
index a24808eac69..d3d69135711 100644
--- a/src/intel/compiler/brw_fs_visitor.cpp
+++ b/src/intel/compiler/brw_fs_visitor.cpp
@@ -126,7 +126,7 @@ fs_visitor::emit_dummy_fs()
    stage_prog_data->nr_pull_params = 0;
    stage_prog_data->curb_read_length = 0;
    stage_prog_data->dispatch_grf_start_reg = 2;
-   wm_prog_data->dispatch_grf_start_reg_2 = 2;
+   wm_prog_data->dispatch_grf_start_reg_16 = 2;
    grf_used = 1; /* Gen4-5 don't allow zero GRF blocks */
 
    calculate_cfg();