squash i965: Use struct push_const_info and support cross-thread constants

The cross thread constant support appears on Haswell. It allows us to
upload a set of uniform data for all threads without duplicating it
per thread.

We also support per-thread data which allows us to store a per-thread
ID in one of the uniforms that can be used to calculate the
gl_LocalInvocationIndex and gl_LocalInvocationID variables.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>

2 files changed, 62 insertions, 47 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_defines.h b/src/mesa/drivers/dri/i965/brw_defines.h
index 31b3336772..54a7fd0244 100644
--- a/src/mesa/drivers/dri/i965/brw_defines.h
+++ b/src/mesa/drivers/dri/i965/brw_defines.h
@@ -2947,6 +2947,9 @@ enum brw_wm_barycentric_interp_mode {
 # define MEDIA_GPGPU_THREAD_COUNT_MASK          INTEL_MASK(7, 0)
 # define GEN8_MEDIA_GPGPU_THREAD_COUNT_SHIFT    0
 # define GEN8_MEDIA_GPGPU_THREAD_COUNT_MASK     INTEL_MASK(9, 0)
+/* GEN7 DW6, GEN8+ DW7 */
+# define CROSS_THREAD_READ_LENGTH_SHIFT         0
+# define CROSS_THREAD_READ_LENGTH_MASK          INTEL_MASK(7, 0)
 #define MEDIA_STATE_FLUSH                       0x7004
 #define GPGPU_WALKER                            0x7105
 /* GEN7 DW0 */
diff --git a/src/mesa/drivers/dri/i965/gen7_cs_state.c b/src/mesa/drivers/dri/i965/gen7_cs_state.c
index 04aad13aa5..11cb81ba42 100644
--- a/src/mesa/drivers/dri/i965/gen7_cs_state.c
+++ b/src/mesa/drivers/dri/i965/gen7_cs_state.c
@@ -137,15 +137,9 @@ brw_upload_cs_state(struct brw_context *brw)
                                             prog_data->binding_table.size_bytes,
                                             32, &stage_state->bind_bo_offset);
 
-   unsigned local_id_dwords = 0;
+   struct push_const_info push_info;
+   fill_push_const_info(brw, prog, cs_prog_data, &push_info);
 
-   if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID)
-      local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
-
-   unsigned push_constant_data_size =
-      (prog_data->nr_params + local_id_dwords) * sizeof(gl_constant_value);
-   unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-   unsigned push_constant_regs = reg_aligned_constant_size / 32;
    unsigned threads = get_cs_thread_count(cs_prog_data);
 
    uint32_t dwords = brw->gen < 8 ? 8 : 9;
@@ -196,7 +190,9 @@ brw_upload_cs_state(struct brw_context *brw)
     *
     * Note: The constant data is built in brw_upload_cs_push_constants below.
     */
-   const uint32_t vfe_curbe_allocation = push_constant_regs * threads;
+   const uint32_t vfe_curbe_allocation =
+      ALIGN(push_info.per_thread.regs * threads +
+            push_info.cross_thread.regs, 2);
    OUT_BATCH(SET_FIELD(vfe_urb_allocation, MEDIA_VFE_STATE_URB_ALLOC) |
              SET_FIELD(vfe_curbe_allocation, MEDIA_VFE_STATE_CURBE_ALLOC));
    OUT_BATCH(0);
@@ -204,11 +200,11 @@ brw_upload_cs_state(struct brw_context *brw)
    OUT_BATCH(0);
    ADVANCE_BATCH();
 
-   if (reg_aligned_constant_size > 0) {
+   if (push_info.total.size > 0) {
       BEGIN_BATCH(4);
       OUT_BATCH(MEDIA_CURBE_LOAD << 16 | (4 - 2));
       OUT_BATCH(0);
-      OUT_BATCH(ALIGN(reg_aligned_constant_size * threads, 64));
+      OUT_BATCH(ALIGN(push_info.total.size, 64));
       OUT_BATCH(stage_state->push_const_offset);
       ADVANCE_BATCH();
    }
@@ -227,7 +223,7 @@ brw_upload_cs_state(struct brw_context *brw)
    desc[dw++] = stage_state->sampler_offset |
       ((stage_state->sampler_count + 3) / 4);
    desc[dw++] = stage_state->bind_bo_offset;
-   desc[dw++] = SET_FIELD(push_constant_regs, MEDIA_CURBE_READ_LENGTH);
+   desc[dw++] = SET_FIELD(push_info.per_thread.regs, MEDIA_CURBE_READ_LENGTH);
    const uint32_t media_threads =
       brw->gen >= 8 ?
       SET_FIELD(threads, GEN8_MEDIA_GPGPU_THREAD_COUNT) :
@@ -249,6 +245,9 @@ brw_upload_cs_state(struct brw_context *brw)
       SET_FIELD(slm_size, MEDIA_SHARED_LOCAL_MEMORY_SIZE) |
       media_threads;
 
+   desc[dw++] =
+      SET_FIELD(push_info.cross_thread.regs, CROSS_THREAD_READ_LENGTH);
+
    BEGIN_BATCH(4);
    OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD << 16 | (4 - 2));
    OUT_BATCH(0);
@@ -291,10 +290,9 @@ brw_upload_cs_push_constants(struct brw_context *brw,
    struct gl_context *ctx = &brw->ctx;
    const struct brw_stage_prog_data *prog_data =
       (struct brw_stage_prog_data*) cs_prog_data;
-   unsigned local_id_dwords = 0;
 
-   if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID)
-      local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
+   struct push_const_info push_info;
+   fill_push_const_info(brw, prog, cs_prog_data, &push_info);
 
    /* Updates the ParamaterValues[i] pointers for all parameters of the
     * basic type of PROGRAM_STATE_VAR.
@@ -302,42 +300,56 @@ brw_upload_cs_push_constants(struct brw_context *brw,
    /* 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 && local_id_dwords == 0) {
+   if (push_info.total.size == 0) {
       stage_state->push_const_size = 0;
-   } else {
-      gl_constant_value *param;
-      unsigned i, t;
-
-      const unsigned push_constant_data_size =
-         (local_id_dwords + prog_data->nr_params) * sizeof(gl_constant_value);
-      const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
-      const unsigned param_aligned_count =
-         reg_aligned_constant_size / sizeof(*param);
-
-      unsigned threads = get_cs_thread_count(cs_prog_data);
-
-      param = (gl_constant_value*)
-         brw_state_batch(brw, type,
-                         ALIGN(reg_aligned_constant_size * threads, 64),
-                         64, &stage_state->push_const_offset);
-      assert(param);
-
-      STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
-
-      brw_cs_fill_local_id_payload(cs_prog_data, param, threads,
-                                   reg_aligned_constant_size);
-
-      /* _NEW_PROGRAM_CONSTANTS */
-      for (t = 0; t < threads; t++) {
-         gl_constant_value *next_param =
-            &param[t * param_aligned_count + local_id_dwords];
-         for (i = 0; i < prog_data->nr_params; i++) {
-            next_param[i] = *prog_data->param[i];
-         }
+      return;
+   }
+
+
+   gl_constant_value *param;
+   param = (gl_constant_value*)
+      brw_state_batch(brw, type,
+                      ALIGN(push_info.total.size, 64),
+                      64, &stage_state->push_const_offset);
+   assert(param);
+
+   STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
+
+   unsigned threads = get_cs_thread_count(cs_prog_data);
+
+   /* Cross thread data must fill registers */
+   assert(push_info.cross_thread.dwords % 8 == 0 ||
+          push_info.per_thread.size == 0);
+   if (push_info.cross_thread.size > 0) {
+      gl_constant_value *param_copy = param;
+      assert(cs_prog_data->thread_local_id_index < 0 ||
+             cs_prog_data->thread_local_id_index >=
+                push_info.cross_thread.dwords);
+      for (unsigned i = 0; i < push_info.cross_thread.dwords; i++) {
+         param_copy[i] = *prog_data->param[i];
       }
+   }
 
-      stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
+   gl_constant_value thread_id;
+   if (push_info.per_thread.size > 0) {
+      for (unsigned t = 0; t < threads; t++) {
+         gl_constant_value *param_copy =
+            &param[8 * t + push_info.cross_thread.dwords];
+         assert(push_info.cross_thread.dwords < prog_data->nr_params);
+         for (unsigned si = push_info.cross_thread.dwords, di = 0;
+              si < prog_data->nr_params;
+              si++, di++) {
+            if (si != cs_prog_data->thread_local_id_index)
+               param_copy[di] = *prog_data->param[si];
+            else {
+               thread_id.u = t * cs_prog_data->simd_size;
+               param_copy[di] = thread_id;
+            }
+         }
+      }
    }
+
+   stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
 }