From 674ca7d88f919f8305af59dbfd918a436c0c9889 Mon Sep 17 00:00:00 2001
From: Jens Steube <jens.steube@gmail.com>
Date: Mon, 12 Jul 2021 11:27:05 +0200
Subject: [PATCH] Add GPU threads to kernel cache checksum because it has an
 influence on HIP offline compile options Add V_ALIGNBIT_B32 inline assembly
 wrapper because HIP does not provide amd_bitalign()

---
 OpenCL/inc_platform.cl | 39 ++++++++++++++++++++++++++++++++++++---
 src/backend.c          | 13 ++++++++-----
 2 files changed, 44 insertions(+), 8 deletions(-)

diff --git a/OpenCL/inc_platform.cl b/OpenCL/inc_platform.cl
index 1a63e7c3a..806a403e4 100644
--- a/OpenCL/inc_platform.cl
+++ b/OpenCL/inc_platform.cl
@@ -246,22 +246,55 @@ DECLSPEC u32 rotr32_S (const u32 a, const int n)
 
 DECLSPEC u64x rotl64 (const u64x a, const int n)
 {
-  return ((a << n) | ((a >> (64 - n))));
+  return rotr64 (a, 64 - n);
+}
+
+DECLSPEC u32 amd_bitalign_S (const u32 a, const u32 b, const int n)
+{
+  u32 r = 0;
+
+  __asm__ ("V_ALIGNBIT_B32 %0, %1, %2, %3;" : "=v"(r): "v"(a), "v"(b), "v"(n));
+
+  return r;
 }
 
 DECLSPEC u64x rotr64 (const u64x a, const int n)
 {
+  #if VECT_SIZE == 1
+  return rotr64_S (a, n);
+  #else
   return ((a >> n) | ((a << (64 - n))));
+  #endif
 }
 
 DECLSPEC u64 rotl64_S (const u64 a, const int n)
 {
-  return ((a << n) | ((a >> (64 - n))));
+  return rotr64_S (a, 64 - n);
 }
 
 DECLSPEC u64 rotr64_S (const u64 a, const int n)
 {
-  return ((a >> n) | ((a << (64 - n))));
+  vconv64_t in;
+
+  in.v64 = a;
+
+  const u32 a0 = in.v32.a;
+  const u32 a1 = in.v32.b;
+
+  vconv64_t out;
+
+  if (n < 32)
+  {
+    out.v32.a = amd_bitalign_S (a1, a0, n);
+    out.v32.b = amd_bitalign_S (a0, a1, n);
+  }
+  else
+  {
+    out.v32.a = amd_bitalign_S (a0, a1, n - 32);
+    out.v32.b = amd_bitalign_S (a1, a0, n - 32);
+  }
+
+  return out.v64;
 }
 
 #define FIXED_THREAD_COUNT(n) __launch_bounds__((n), 0)
diff --git a/src/backend.c b/src/backend.c
index 6620d3d42..02cbbc3dd 100644
--- a/src/backend.c
+++ b/src/backend.c
@@ -10498,8 +10498,9 @@ static bool load_kernel (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_p
       // TODO HIP
       // no -offload-arch= aka --gpu-architecture because hiprtc gets native arch from hip_context
 
-      hiprtc_options[0] = "--gpu-max-threads-per-block=64";
-      hiprtc_options[1] = "";
+      hc_asprintf (&hiprtc_options[0], "--gpu-max-threads-per-block=%u", device_param->kernel_threads);
+
+      hiprtc_options[1] = "-O3";
       hiprtc_options[2] = "";
       hiprtc_options[3] = "";
 
@@ -11588,7 +11589,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
     char device_name_chksum_amp_mp[HCBUFSIZ_TINY] = { 0 };
 
-    const size_t dnclen_amp_mp = snprintf (device_name_chksum_amp_mp, HCBUFSIZ_TINY, "%d-%d-%d-%d-%u-%s-%s-%s",
+    const size_t dnclen_amp_mp = snprintf (device_name_chksum_amp_mp, HCBUFSIZ_TINY, "%d-%d-%d-%d-%u-%s-%s-%s-%d",
       backend_ctx->comptime,
       backend_ctx->cuda_driver_version,
       backend_ctx->hip_driver_version,
@@ -11596,7 +11597,8 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
       device_param->opencl_platform_vendor_id,
       device_param->device_name,
       device_param->opencl_device_version,
-      device_param->opencl_driver_version);
+      device_param->opencl_driver_version,
+      device_param->kernel_threads);
 
     md5_ctx_t md5_ctx;
 
@@ -11887,7 +11889,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
 
       const u32 extra_value = (user_options->attack_mode == ATTACK_MODE_ASSOCIATION) ? ATTACK_MODE_ASSOCIATION : ATTACK_MODE_NONE;
 
-      const size_t dnclen = snprintf (device_name_chksum, HCBUFSIZ_TINY, "%d-%d-%d-%d-%u-%s-%s-%s-%d-%u-%u-%s",
+      const size_t dnclen = snprintf (device_name_chksum, HCBUFSIZ_TINY, "%d-%d-%d-%d-%u-%s-%s-%s-%d-%u-%d-%u-%s",
         backend_ctx->comptime,
         backend_ctx->cuda_driver_version,
         backend_ctx->hip_driver_version,
@@ -11897,6 +11899,7 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
         device_param->opencl_device_version,
         device_param->opencl_driver_version,
         device_param->vector_width,
+        device_param->kernel_threads,
         hashconfig->kern_type,
         extra_value,
         build_options_module_buf);