i965/fs: Lower 32x32 bit multiplication on BXT.
[mesa.git] / src / mesa / drivers / dri / i965 / gen6_vs_state.c
index 905e123837ffb949f1e2640580b0c7604ab33529..35d10ef8779d31beefca1eb9f8b33fc88e77f19f 100644 (file)
@@ -58,22 +58,24 @@ gen6_upload_push_constants(struct brw_context *brw,
 {
    struct gl_context *ctx = &brw->ctx;
 
-   /* Updates the ParamaterValues[i] pointers for all parameters of the
-    * basic type of PROGRAM_STATE_VAR.
-    */
-   /* XXX: Should this happen somewhere before to get our state flag set? */
-   _mesa_load_state_parameters(ctx, prog->Parameters);
-
    if (prog_data->nr_params == 0) {
       stage_state->push_const_size = 0;
    } else {
-      float *param;
+      /* Updates the ParamaterValues[i] pointers for all parameters of the
+       * basic type of PROGRAM_STATE_VAR.
+       */
+      /* XXX: Should this happen somewhere before to get our state flag set? */
+      _mesa_load_state_parameters(ctx, prog->Parameters);
+
+      gl_constant_value *param;
       int i;
 
       param = brw_state_batch(brw, type,
-                             prog_data->nr_params * sizeof(float),
+                             prog_data->nr_params * sizeof(gl_constant_value),
                              32, &stage_state->push_const_offset);
 
+      STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
+
       /* _NEW_PROGRAM_CONSTANTS
        *
        * Also _NEW_TRANSFORM -- we may reference clip planes other than as a
@@ -91,7 +93,7 @@ gen6_upload_push_constants(struct brw_context *brw,
            if ((i & 7) == 0)
               fprintf(stderr, "g%d: ",
                        prog_data->dispatch_grf_start_reg + i / 8);
-           fprintf(stderr, "%8f ", param[i]);
+           fprintf(stderr, "%8f ", param[i].f);
            if ((i & 7) == 7)
               fprintf(stderr, "\n");
         }
@@ -128,7 +130,7 @@ gen6_upload_vs_push_constants(struct brw_context *brw)
    /* _BRW_NEW_VERTEX_PROGRAM */
    const struct brw_vertex_program *vp =
       brw_vertex_program_const(brw->vertex_program);
-   /* CACHE_NEW_VS_PROG */
+   /* BRW_NEW_VS_PROG_DATA */
    const struct brw_stage_prog_data *prog_data = &brw->vs.prog_data->base.base;
 
    gen6_upload_push_constants(brw, &vp->program.Base, prog_data,
@@ -145,11 +147,12 @@ gen6_upload_vs_push_constants(struct brw_context *brw)
 
 const struct brw_tracked_state gen6_vs_push_constants = {
    .dirty = {
-      .mesa  = _NEW_TRANSFORM | _NEW_PROGRAM_CONSTANTS,
-      .brw   = (BRW_NEW_BATCH |
-                BRW_NEW_VERTEX_PROGRAM |
-                BRW_NEW_PUSH_CONSTANT_ALLOCATION),
-      .cache = CACHE_NEW_VS_PROG,
+      .mesa  = _NEW_PROGRAM_CONSTANTS |
+               _NEW_TRANSFORM,
+      .brw   = BRW_NEW_BATCH |
+               BRW_NEW_PUSH_CONSTANT_ALLOCATION |
+               BRW_NEW_VERTEX_PROGRAM |
+               BRW_NEW_VS_PROG_DATA,
    },
    .emit = gen6_upload_vs_push_constants,
 };
@@ -157,7 +160,6 @@ const struct brw_tracked_state gen6_vs_push_constants = {
 static void
 upload_vs_state(struct brw_context *brw)
 {
-   struct gl_context *ctx = &brw->ctx;
    const struct brw_stage_state *stage_state = &brw->vs.base;
    uint32_t floating_point_mode = 0;
 
@@ -169,10 +171,9 @@ upload_vs_state(struct brw_context *brw)
     *   flush can be executed by sending a PIPE_CONTROL command with CS
     *   stall bit set and a post sync operation.
     *
-    * Although we don't disable the VS during normal drawing, BLORP sometimes
-    * disables it.  To be safe, do the flush here just in case.
+    * We've already done such a flush at the start of state upload, so we
+    * don't need to do another one here.
     */
-   intel_emit_post_sync_nonzero_flush(brw);
 
    if (stage_state->push_const_size == 0) {
       /* Disable the push constant buffers. */
@@ -199,10 +200,7 @@ upload_vs_state(struct brw_context *brw)
       ADVANCE_BATCH();
    }
 
-   /* Use ALT floating point mode for ARB vertex programs, because they
-    * require 0^0 == 1.
-    */
-   if (ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX] == NULL)
+   if (brw->vs.prog_data->base.base.use_alt_mode)
       floating_point_mode = GEN6_VS_FLOATING_POINT_MODE_ALT;
 
    BEGIN_BATCH(6);
@@ -213,10 +211,10 @@ upload_vs_state(struct brw_context *brw)
              ((brw->vs.prog_data->base.base.binding_table.size_bytes / 4) <<
               GEN6_VS_BINDING_TABLE_ENTRY_COUNT_SHIFT));
 
-   if (brw->vs.prog_data->base.total_scratch) {
+   if (brw->vs.prog_data->base.base.total_scratch) {
       OUT_RELOC(stage_state->scratch_bo,
                I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
-               ffs(brw->vs.prog_data->base.total_scratch) - 11);
+               ffs(brw->vs.prog_data->base.base.total_scratch) - 11);
    } else {
       OUT_BATCH(0);
    }
@@ -248,21 +246,21 @@ upload_vs_state(struct brw_context *brw)
     * bug reports that led to this workaround, and may be more than
     * what is strictly required to avoid the issue.
     */
-   intel_emit_post_sync_nonzero_flush(brw);
    brw_emit_pipe_control_flush(brw,
                                PIPE_CONTROL_DEPTH_STALL |
-                               PIPE_CONTROL_INSTRUCTION_FLUSH |
+                               PIPE_CONTROL_INSTRUCTION_INVALIDATE |
                                PIPE_CONTROL_STATE_CACHE_INVALIDATE);
 }
 
 const struct brw_tracked_state gen6_vs_state = {
    .dirty = {
-      .mesa  = _NEW_TRANSFORM | _NEW_PROGRAM_CONSTANTS,
-      .brw   = (BRW_NEW_CONTEXT |
-               BRW_NEW_VERTEX_PROGRAM |
-               BRW_NEW_BATCH |
-                BRW_NEW_PUSH_CONSTANT_ALLOCATION),
-      .cache = CACHE_NEW_VS_PROG
+      .mesa  = _NEW_PROGRAM_CONSTANTS |
+               _NEW_TRANSFORM,
+      .brw   = BRW_NEW_BATCH |
+               BRW_NEW_CONTEXT |
+               BRW_NEW_PUSH_CONSTANT_ALLOCATION |
+               BRW_NEW_VERTEX_PROGRAM |
+               BRW_NEW_VS_PROG_DATA,
    },
    .emit = upload_vs_state,
 };