diff --git a/src/gallium/drivers/radeonsi/si_compute_blit.c b/src/gallium/drivers/radeonsi/si_compute_blit.c
index 0e14a6be9e2..9142e3a0b1e 100644
--- a/src/gallium/drivers/radeonsi/si_compute_blit.c
+++ b/src/gallium/drivers/radeonsi/si_compute_blit.c
@@ -126,9 +126,6 @@ static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe
    assert(size % 4 == 0);
    unsigned size_12 = DIV_ROUND_UP(size, 12);
 
-   unsigned data[4] = {0};
-   memcpy(data, clear_value, 12);
-
    if (!(flags & SI_OP_SKIP_CACHE_INV_BEFORE))
       sctx->flags |= si_get_flush_flags(sctx, coher, SI_COMPUTE_DST_CACHE_POLICY);
 
@@ -140,16 +137,8 @@ static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe
        (1u << si_get_shaderbuf_slot(0)))
       saved_writable_mask = 1;
 
-   struct pipe_constant_buffer saved_cb = {};
-   si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &saved_cb);
-
    void *saved_cs = sctx->cs_shader_state.program;
 
-   struct pipe_constant_buffer cb = {};
-   cb.buffer_size = sizeof(data);
-   cb.user_buffer = data;
-   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, false, &cb);
-
    struct pipe_shader_buffer sb = {0};
    sb.buffer = dst;
    sb.buffer_offset = dst_offset;
@@ -157,6 +146,8 @@ static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe
 
    ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0, 1, &sb, 0x1);
 
+   memcpy(sctx->cs_user_data, clear_value, 12);
+
    struct pipe_grid_info info = {0};
 
    if (!sctx->cs_clear_12bytes_buffer)
@@ -173,8 +164,6 @@ static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe
    si_launch_grid_internal(sctx, &info, saved_cs, flags);
 
    ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0, 1, &saved_sb, saved_writable_mask);
-   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, true, &saved_cb);
-
    pipe_resource_reference(&saved_sb.buffer, NULL);
 }
 
diff --git a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
index 94d86686f2c..48fafdbb0ee 100644
--- a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
+++ b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
@@ -686,21 +686,25 @@ void *si_clear_12bytes_buffer_shader(struct pipe_context *ctx)
                               "PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
                               "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
                               "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
+                              "PROPERTY CS_USER_DATA_COMPONENTS_AMD 3\n"
                               "DCL SV[0], THREAD_ID\n"
                               "DCL SV[1], BLOCK_ID\n"
+                              "DCL SV[2], CS_USER_DATA_AMD\n"
                               "DCL BUFFER[0]\n"
-                              "DCL CONST[0][0..0]\n" // 0:xyzw
                               "DCL TEMP[0..0]\n"
                               "IMM[0] UINT32 {64, 1, 12, 0}\n"
                               "UMAD TEMP[0].x, SV[1].xyzz, IMM[0].xyyy, SV[0].xyzz\n"
                               "UMUL TEMP[0].x, TEMP[0].xyzz, IMM[0].zzzz\n" // 12 bytes
-                              "STORE BUFFER[0].xyz, TEMP[0].xxxx, CONST[0][0].xyzw\n"
+                              "STORE BUFFER[0].xyz, TEMP[0].xxxx, SV[2].xyzz%s\n"
                               "END\n";
-
+   char final_text[2048];
    struct tgsi_token tokens[1024];
    struct pipe_compute_state state = {0};
 
-   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
+   snprintf(final_text, sizeof(final_text), text,
+            SI_COMPUTE_DST_CACHE_POLICY != L2_LRU ? ", STREAM_CACHE_POLICY" : "");
+
+   if (!tgsi_text_translate(final_text, tokens, ARRAY_SIZE(tokens))) {
       assert(false);
       return NULL;
    }