i965/fs: Lower 32x32 bit multiplication on BXT.

[mesa.git] / src / mesa / drivers / dri / i965 / gen7_wm_state.c

diff --git a/src/mesa/drivers/dri/i965/gen7_wm_state.c b/src/mesa/drivers/dri/i965/gen7_wm_state.c
index 923414edfb8d99c16a43b885733b745ed6407405..d7be58dd0f7169b058d4e11c68013c0e10bc607d 100644 (file)
--- a/src/mesa/drivers/dri/i965/gen7_wm_state.c
+++ b/src/mesa/drivers/dri/i965/gen7_wm_state.c
@@ -30,6 +30,7 @@
 #include "program/program.h"
 #include "program/prog_parameter.h"
 #include "program/prog_statevars.h"
+#include "main/framebuffer.h"
 #include "intel_batchbuffer.h"
 
 static void
@@ -45,7 +46,7 @@ upload_wm_state(struct brw_context *brw)
    uint32_t dw1, dw2;
 
    /* _NEW_BUFFERS */
-   bool multisampled_fbo = ctx->DrawBuffer->Visual.samples > 1;
+   const bool multisampled_fbo = _mesa_geometric_samples(ctx->DrawBuffer) > 1;
 
    dw1 = dw2 = 0;
    dw1 |= GEN7_WM_STATISTICS_ENABLE;
@@ -76,6 +77,10 @@ upload_wm_state(struct brw_context *brw)
       dw1 |= GEN7_WM_KILL_ENABLE;
    }
 
+   if (_mesa_active_fragment_shader_has_atomic_ops(&brw->ctx)) {
+      dw1 |= GEN7_WM_DISPATCH_ENABLE;
+   }
+
    /* _NEW_BUFFERS | _NEW_COLOR */
    if (brw_color_buffer_write_enabled(brw) || writes_depth ||
        dw1 & GEN7_WM_KILL_ENABLE) {
@@ -123,20 +128,23 @@ const struct brw_tracked_state gen7_wm_state = {
 };
 
 static void
-upload_ps_state(struct brw_context *brw)
+gen7_upload_ps_state(struct brw_context *brw,
+                     const struct gl_fragment_program *fp,
+                     const struct brw_stage_state *stage_state,
+                     const struct brw_wm_prog_data *prog_data,
+                     bool enable_dual_src_blend, unsigned sample_mask,
+                     unsigned fast_clear_op)
 {
    struct gl_context *ctx = &brw->ctx;
    uint32_t dw2, dw4, dw5, ksp0, ksp2;
    const int max_threads_shift = brw->is_haswell ?
       HSW_PS_MAX_THREADS_SHIFT : IVB_PS_MAX_THREADS_SHIFT;
 
-   /* BRW_NEW_FS_PROG_DATA */
-   const struct brw_wm_prog_data *prog_data = brw->wm.prog_data;
-
    dw2 = dw4 = dw5 = ksp2 = 0;
 
-   dw2 |=
-      (ALIGN(brw->wm.base.sampler_count, 4) / 4) << GEN7_PS_SAMPLER_COUNT_SHIFT;
+   const unsigned sampler_count =
+      DIV_ROUND_UP(CLAMP(stage_state->sampler_count, 0, 16), 4);
+   dw2 |= SET_FIELD(sampler_count, GEN7_PS_SAMPLER_COUNT);
 
    dw2 |= ((prog_data->base.binding_table.size_bytes / 4) <<
            GEN7_PS_BINDING_TABLE_ENTRY_COUNT_SHIFT);
@@ -148,7 +156,7 @@ upload_ps_state(struct brw_context *brw)
     * in 3DSTATE_SAMPLE_MASK; the values should match. */
    /* _NEW_BUFFERS, _NEW_MULTISAMPLE */
    if (brw->is_haswell)
-      dw4 |= SET_FIELD(gen6_determine_sample_mask(brw), HSW_PS_SAMPLE_MASK);
+      dw4 |= SET_FIELD(sample_mask, HSW_PS_SAMPLE_MASK);
 
    dw4 |= (brw->max_wm_threads - 1) << max_threads_shift;
 
@@ -181,16 +189,11 @@ upload_ps_state(struct brw_context *brw)
    else
       dw4 |= GEN7_PS_POSOFFSET_NONE;
 
-   /* BRW_NEW_FS_PROG_DATA | _NEW_COLOR
-    *
-    * The hardware wedges if you have this bit set but don't turn on any dual
+   /* The hardware wedges if you have this bit set but don't turn on any dual
     * source blend factors.
     */
-   if (prog_data->dual_src_blend &&
-       (ctx->Color.BlendEnabled & 1) &&
-       ctx->Color.Blend[0]._UsesDualSrc) {
+   if (enable_dual_src_blend)
       dw4 |= GEN7_PS_DUAL_SOURCE_BLEND_ENABLE;
-   }
 
    /* BRW_NEW_FS_PROG_DATA */
    if (prog_data->num_varying_inputs != 0)
@@ -202,7 +205,7 @@ upload_ps_state(struct brw_context *brw)
     * better performance than 'SIMD8 only' dispatch.
     */
    int min_inv_per_frag =
-      _mesa_get_min_invocations_per_fragment(ctx, brw->fragment_program, false);
+      _mesa_get_min_invocations_per_fragment(ctx, fp, false);
    assert(min_inv_per_frag >= 1);
 
    if (prog_data->prog_offset_16 || prog_data->no_8) {
@@ -213,22 +216,22 @@ upload_ps_state(struct brw_context *brw)
                  GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
          dw5 |= (prog_data->dispatch_grf_start_reg_16 <<
                  GEN7_PS_DISPATCH_START_GRF_SHIFT_2);
-         ksp0 = brw->wm.base.prog_offset;
-         ksp2 = brw->wm.base.prog_offset + prog_data->prog_offset_16;
+         ksp0 = stage_state->prog_offset;
+         ksp2 = stage_state->prog_offset + prog_data->prog_offset_16;
       } else {
          dw5 |= (prog_data->dispatch_grf_start_reg_16 <<
                  GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
-         ksp0 = brw->wm.base.prog_offset + prog_data->prog_offset_16;
+         ksp0 = stage_state->prog_offset + prog_data->prog_offset_16;
       }
    }
    else {
       dw4 |= GEN7_PS_8_DISPATCH_ENABLE;
       dw5 |= (prog_data->base.dispatch_grf_start_reg <<
               GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
-      ksp0 = brw->wm.base.prog_offset;
+      ksp0 = stage_state->prog_offset;
    }
 
-   dw4 |= brw->wm.fast_clear_op;
+   dw4 |= fast_clear_op;
 
    BEGIN_BATCH(8);
    OUT_BATCH(_3DSTATE_PS << 16 | (8 - 2));
@@ -248,6 +251,25 @@ upload_ps_state(struct brw_context *brw)
    ADVANCE_BATCH();
 }
 
+static void
+upload_ps_state(struct brw_context *brw)
+{
+   /* BRW_NEW_FS_PROG_DATA */
+   const struct brw_wm_prog_data *prog_data = brw->wm.prog_data;
+   const struct gl_context *ctx = &brw->ctx;
+   /* BRW_NEW_FS_PROG_DATA | _NEW_COLOR */
+   const bool enable_dual_src_blend = prog_data->dual_src_blend &&
+                                      (ctx->Color.BlendEnabled & 1) &&
+                                      ctx->Color.Blend[0]._UsesDualSrc;
+   /* _NEW_BUFFERS, _NEW_MULTISAMPLE */
+   const unsigned sample_mask =
+      brw->is_haswell ? gen6_determine_sample_mask(brw) : 0;
+
+   gen7_upload_ps_state(brw, brw->fragment_program, &brw->wm.base, prog_data,
+                        enable_dual_src_blend, sample_mask,
+                        brw->wm.fast_clear_op);
+}
+
 const struct brw_tracked_state gen7_ps_state = {
    .dirty = {
       .mesa  = _NEW_BUFFERS |