Merge pull request #1926 from enjoy-digital/hyperbus_variable_latency

HyperRAM: Add variable latency and configuration support.

litex/soc/cores/hyperbus.py

@@ -1,7 +1,7 @@
 #
-# This file is part of LiteHyperBus
+# This file is part of LiteX.
 #
-# Copyright (c) 2019-2022 Florent Kermarrec <florent@enjoy-digital.fr>
+# Copyright (c) 2019-2024 Florent Kermarrec <florent@enjoy-digital.fr>
 # Copyright (c) 2019 Antti Lukats <antti.lukats@gmail.com>
 # Copyright (c) 2021 Franck Jullien <franck.jullien@collshade.fr>
 # SPDX-License-Identifier: BSD-2-Clause
@@ -11,6 +11,9 @@ from migen import *
 from litex.gen import *
 from litex.gen.genlib.misc import WaitTimer
 
+from litex.soc.interconnect.csr import *
+from litex.soc.interconnect import stream
+
 from litex.build.io import DifferentialOutput
 
 from litex.soc.interconnect import wishbone
@@ -21,25 +24,56 @@ class HyperRAM(LiteXModule):
     tCSM = 4e-6
     """HyperRAM
 
-    Provides a very simple/minimal HyperRAM core that should work with all FPGA/HyperRam chips:
+    Provides a very simple/minimal HyperRAM core with a Wishbone Interface that can work with all
-    - FPGA vendor agnostic.
+    FPGA/HyperRam chips:
-    - no setup/chip configuration (use default latency).
+    - Vendor agnostic.
+    - Fixed/Variable latency.
+    - Latency/Registers (re-)configuration.
 
-    This core favors portability and ease of use over performance.
+    Parameters:
-    """
+        pads (Record)                  : Interface to the HyperRAM connection pads.
-    def __init__(self, pads, latency=6, sys_clk_freq=None):
+        latency (int, optional)        : Initial latency setting, defaults to 6.
+        latency_mode (str, optional)   : Specifies the latency mode ('fixed' or 'variable'), defaults to 'fixed'.
+        sys_clk_freq (float, optional) : System clock frequency in Hz.
+        with_csr (bool, optional)      : Enables CSR interface for Latency/Registers configuration, defaults to True.
+
+    Attributes:
+        pads (Record)            : Platform pads of HyperRAM.
+        bus (wishbone.Interface) : Wishbone Interface.
+"""
+    def __init__(self, pads, latency=6, latency_mode="fixed", sys_clk_freq=None, with_csr=True):
         self.pads = pads
         self.bus  = bus = wishbone.Interface(data_width=32, address_width=32, addressing="word")
 
+        # Config/Reg Interface.
+        # ---------------------
+        self.conf_rst       = Signal()
+        self.conf_latency   = Signal(8, reset=latency)
+        self.reg_write      = Signal()
+        self.reg_read       = Signal()
+        self.reg_addr       = Signal(2)
+        self.reg_write_done = Signal()
+        self.reg_read_done  = Signal()
+        self.reg_write_data = Signal(16)
+        self.reg_read_data  = Signal(16)
+        if with_csr:
+            self.add_csr(default_latency=latency)
+
         # # #
 
+        # Parameters.
+        # -----------
+        assert latency_mode in ["fixed", "variable"]
+
+        # Internal Signals.
+        # -----------------
         clk       = Signal()
         clk_phase = Signal(2)
         cs        = Signal()
         ca        = Signal(48)
         ca_active = Signal()
         sr        = Signal(48)
-        sr_new    = Signal(48)
+        sr_next   = Signal(48)
         dq        = self.add_tristate(pads.dq)   if not hasattr(pads.dq,   "oe") else pads.dq
         rwds      = self.add_tristate(pads.rwds) if not hasattr(pads.rwds, "oe") else pads.rwds
         dw        = len(pads.dq)                 if not hasattr(pads.dq,   "oe") else len(pads.dq.o)
@@ -50,7 +84,7 @@ class HyperRAM(LiteXModule):
 
         # Rst.
         if hasattr(pads, "rst_n"):
-            self.comb += pads.rst_n.eq(1)
+            self.comb += pads.rst_n.eq(1 & ~self.conf_rst)
 
         # CSn.
         self.comb += pads.cs_n[0].eq(~cs)
@@ -80,16 +114,16 @@ class HyperRAM(LiteXModule):
         dqi = Signal(dw)
         self.sync += dqi.eq(dq.i) # Sample on 90° and 270°
         self.comb += [
-            sr_new.eq(Cat(dqi, sr[:-dw])),
+            sr_next.eq(Cat(dqi, sr[:-dw])),
             If(ca_active,
-                sr_new.eq(Cat(dqi[:8], sr[:-8])) # Only 8-bit during Command/Address.
+                sr_next.eq(Cat(dqi[:8], sr[:-8])) # Only 8-bit during Command/Address.
             )
         ]
-        self.sync += If(clk_phase[0] == 0, sr.eq(sr_new)) # Shift on 0° and 180°
+        self.sync += If(clk_phase[0] == 0, sr.eq(sr_next)) # Shift on 0° and 180°
 
         # Data Shift-Out Register ------------------------------------------------------------------
         self.comb += [
-            bus.dat_r.eq(sr_new),
+            bus.dat_r.eq(sr_next),
             If(dq.oe,
                 dq.o.eq(sr[-dw:]),
                 If(ca_active,
@@ -98,41 +132,73 @@ class HyperRAM(LiteXModule):
             )
         ]
 
+        # Register Access/Buffer -------------------------------------------------------------------
+        reg_write_req = Signal()
+        reg_read_req  = Signal()
+        self.reg_buf = reg_buf = stream.SyncFIFO(
+            layout = [("write", 1), ("read", 1), ("addr", 4), ("data", 16)],
+            depth  = 4,
+        )
+        reg_ep = reg_buf.source
+        self.comb += [
+            reg_buf.sink.valid.eq(self.reg_write | self.reg_read),
+            reg_buf.sink.write.eq(self.reg_write),
+            reg_buf.sink.read.eq(self.reg_read),
+            reg_buf.sink.addr.eq(self.reg_addr),
+            reg_buf.sink.data.eq(self.reg_write_data),
+            reg_write_req.eq(reg_ep.valid & reg_ep.write),
+            reg_read_req.eq( reg_ep.valid & reg_ep.read),
+        ]
+        self.sync += If(reg_buf.sink.valid,
+            self.reg_write_done.eq(0),
+            self.reg_read_done.eq(0),
+        )
+
         # Command generation -----------------------------------------------------------------------
         ashift = {8:1, 16:0}[dw]
         self.comb += [
-            ca[47].eq(~bus.we),               # R/W#
+            # Register Command Generation.
-            ca[45].eq(1),                     # Burst Type (Linear)
+            If(reg_write_req | reg_read_req,
-            ca[16:45].eq(bus.adr[3-ashift:]), # Row & Upper Column Address
+                ca[47].eq(reg_ep.read), # R/W#
-            ca[ashift:3].eq(bus.adr),         # Lower Column Address
+                ca[46].eq(1),           # Register Space.
+                ca[45].eq(1),           # Burst Type (Linear)
+                Case(reg_ep.addr, {
+                    0 : ca[0:40].eq(0x00_00_00_00_00), # Identification Register 0 (Read Only).
+                    1 : ca[0:40].eq(0x00_00_00_00_01), # Identification Register 1 (Read Only).
+                    2 : ca[0:40].eq(0x00_01_00_00_00), # Configuration Register 0.
+                    3 : ca[0:40].eq(0x00_01_00_00_01), # Configuration Register 1.
+                }),
+            # Wishbone Command Generation.
+            ).Else(
+                ca[47].eq(~bus.we),                # R/W#
+                ca[46].eq(0),                      # Memory Space.
+                ca[45].eq(1),                      # Burst Type (Linear)
+                ca[16:45].eq(bus.adr[3-ashift:]),  # Row & Upper Column Address
+                ca[ashift:3].eq(bus.adr),          # Lower Column Address
+            )
         ]
 
-        # Latency count starts from the middle of the command (thus the -4). In fixed latency mode
-        # (default), latency is 2 x Latency count. We have 4 x sys_clk per RAM clock:
-        latency_cycles = (latency * 2 * 4) - 4
-
         # Bus Latch --------------------------------------------------------------------------------
         bus_adr   = Signal(32)
         bus_we    = Signal()
         bus_sel   = Signal(4)
         bus_latch = Signal()
         self.sync += If(bus_latch,
-            If(bus.we,
+            If(bus.we, sr.eq(Cat(Signal(16), bus.dat_w))),
-                sr.eq(Cat(Signal(16), bus.dat_w)),
-            ),
             bus_we.eq(bus.we),
             bus_sel.eq(bus.sel),
             bus_adr.eq(bus.adr)
         )
 
         # FSM (Sequencer) --------------------------------------------------------------------------
-        cycles = Signal(8)
+        cycles  = Signal(8)
-        first  = Signal()
+        first   = Signal()
+        refresh = Signal()
         self.fsm = fsm = FSM(reset_state="IDLE")
         fsm.act("IDLE",
             NextValue(first, 1),
-            If(bus.cyc & bus.stb,
+            If(clk_phase == 0,
-                If(clk_phase == 0,
+                If((bus.cyc & bus.stb) | reg_write_req | reg_read_req,
                     NextValue(sr, ca),
                     NextState("SEND-COMMAND-ADDRESS")
                 )
@@ -146,18 +212,53 @@ class HyperRAM(LiteXModule):
             dq.oe.eq(1),
             # Wait for 6*2 cycles...
             If(cycles == (6*2 - 1),
-                NextState("WAIT-LATENCY")
+                If(reg_write_req,
+                    NextValue(sr, Cat(Signal(40), self.reg_write_data[8:])),
+                    NextState("REG-WRITE-0")
+                ).Else(
+                    # Sample RWDS to know if 1X/2X Latency should be used (Refresh).
+                    NextValue(refresh, rwds.i | (latency_mode in ["fixed"])),
+                    NextState("WAIT-LATENCY")
+                )
+            )
+        )
+        fsm.act("REG-WRITE-0",
+            # Set CSn.
+            cs.eq(1),
+            # Send Reg on DQ.
+            ca_active.eq(1),
+            dq.oe.eq(1),
+            # Wait for 2 cycles...
+            If(cycles == (2 - 1),
+                NextValue(sr, Cat(Signal(40), self.reg_write_data[:8])),
+                NextState("REG-WRITE-1")
+            )
+        )
+        fsm.act("REG-WRITE-1",
+            # Set CSn.
+            cs.eq(1),
+            # Send Reg on DQ.
+            ca_active.eq(1),
+            dq.oe.eq(1),
+            # Wait for 2 cycles...
+            If(cycles == (2 - 1),
+                reg_ep.ready.eq(1),
+                NextValue(self.reg_write_done, 1),
+                NextState("IDLE")
             )
         )
         fsm.act("WAIT-LATENCY",
             # Set CSn.
             cs.eq(1),
-            # Wait for Latency cycles...
+            # Wait for 1X or 2X Latency cycles... (-4 since count start in the middle of the command).
-            If(cycles == (latency_cycles - 1),
+            If(((cycles == 2*(self.conf_latency * 4) - 4 - 1) &  refresh) | # 2X Latency (No DRAM refresh required).
+               ((cycles == 1*(self.conf_latency * 4) - 4 - 1) & ~refresh) , # 1X Latency (   DRAM refresh required).
                 # Latch Bus.
                 bus_latch.eq(1),
                 # Early Write Ack (to allow bursting).
-                bus.ack.eq(bus.we),
+                If(~reg_read_req,
+                    bus.ack.eq(bus.we),
+                ),
                 NextState("READ-WRITE-DATA0")
             )
         )
@@ -168,6 +269,7 @@ class HyperRAM(LiteXModule):
                 burst_timer.wait.eq(1),
                 # Set CSn.
                 cs.eq(1),
+                ca_active.eq(reg_read_req),
                 # Send Data on DQ/RWDS (for write).
                 If(bus_we,
                     dq.oe.eq(1),
@@ -182,7 +284,7 @@ class HyperRAM(LiteXModule):
                     If(n == (states - 1),
                         NextValue(first, 0),
                         # Continue burst when a consecutive access is ready.
-                        If(bus.stb & bus.cyc & (bus.we == bus_we) & (bus.adr == (bus_adr + 1)) & (~burst_timer.done),
+                        If(~reg_read_req & bus.stb & bus.cyc & (bus.we == bus_we) & (bus.adr == (bus_adr + 1)) & (~burst_timer.done),
                             # Latch Bus.
                             bus_latch.eq(1),
                             # Early Write Ack (to allow bursting).
@@ -194,7 +296,14 @@ class HyperRAM(LiteXModule):
                     ),
                     # Read Ack (when dat_r ready).
                     If((n == 0) & ~first,
-                        bus.ack.eq(~bus_we),
+                        If(reg_read_req,
+                            reg_ep.ready.eq(1),
+                            NextValue(self.reg_read_done, 1),
+                            NextValue(self.reg_read_data, bus.dat_r),
+                            NextState("IDLE"),
+                        ).Else(
+                            bus.ack.eq(~bus_we),
+                        )
                     )
                 )
             )
@@ -206,3 +315,49 @@ class HyperRAM(LiteXModule):
         t = TSTriple(len(pad))
         self.specials += t.get_tristate(pad)
         return t
+
+    def add_csr(self, default_latency=6):
+        # Config Interface.
+        # -----------------
+        self.config = CSRStorage(fields=[
+            CSRField("rst",     offset=0, size=1, pulse=True, description="HyperRAM Rst."),
+            CSRField("latency", offset=8, size=8,             description="HyperRAM Latency (X1).", reset=default_latency),
+        ])
+        self.comb += [
+            self.conf_rst.eq(    self.config.fields.rst),
+            self.conf_latency.eq(self.config.fields.latency),
+        ]
+
+        # Reg Interface.
+        # --------------
+        self.reg_control = CSRStorage(fields=[
+            CSRField("write", offset=0, size=1, pulse=True, description="Issue Register Write."),
+            CSRField("read",  offset=1, size=1, pulse=True, description="Issue Register Read."),
+            CSRField("addr",  offset=8, size=4, values=[
+                ("``0b00``", "Identification Register 0 (Read Only)."),
+                ("``0b01``", "Identification Register 1 (Read Only)."),
+                ("``0b10``", "Configuration Register 0."),
+                ("``0b11``", "Configuration Register 1."),
+            ]),
+        ])
+        self.reg_status = CSRStatus(fields=[
+            CSRField("write_done", offset=0, size=1, description="Register Write Done."),
+            CSRField("read_done",  offset=1, size=1, description="Register Read Done."),
+        ])
+        self.reg_wdata = CSRStorage(16, description="Register Write Data.")
+        self.reg_rdata = CSRStatus( 16, description="Register Read Data.")
+
+        self.comb += [
+            # Control.
+            self.reg_write.eq(self.reg_control.fields.write),
+            self.reg_read.eq( self.reg_control.fields.read),
+            self.reg_addr.eq( self.reg_control.fields.addr),
+
+            # Status.
+            self.reg_status.fields.write_done.eq(self.reg_write_done),
+            self.reg_status.fields.read_done.eq( self.reg_read_done),
+
+            # Data.
+            self.reg_write_data.eq(self.reg_wdata.storage),
+            self.reg_rdata.status.eq(self.reg_read_data),
+        ]

litex/soc/software/bios/main.c

@@ -175,6 +175,87 @@ __attribute__((__used__)) int main(int i, char **c)
 
         sdr_ok = 1;
 
+#ifdef CSR_HYPERRAM_BASE /* FIXME: Move to libbase/hyperram.h/c? */
+    /* Helper Functions */
+
+    printf("HyperRAM init...\n");
+    void hyperram_write_reg(uint16_t reg_addr, uint16_t data) {
+        /* Write data to the register */
+        hyperram_reg_wdata_write(data);
+        hyperram_reg_control_write(
+            1        << CSR_HYPERRAM_REG_CONTROL_WRITE_OFFSET |
+            0        << CSR_HYPERRAM_REG_CONTROL_READ_OFFSET  |
+            reg_addr << CSR_HYPERRAM_REG_CONTROL_ADDR_OFFSET
+        );
+        /* Wait for write to complete */
+        while ((hyperram_reg_status_read() & (1 << CSR_HYPERRAM_REG_STATUS_WRITE_DONE_OFFSET)) == 0);
+    }
+
+    uint16_t hyperram_read_reg(uint16_t reg_addr) {
+        /* Read data from the register */
+        hyperram_reg_control_write(
+            0        << CSR_HYPERRAM_REG_CONTROL_WRITE_OFFSET |
+            1        << CSR_HYPERRAM_REG_CONTROL_READ_OFFSET  |
+            reg_addr << CSR_HYPERRAM_REG_CONTROL_ADDR_OFFSET
+        );
+        /* Wait for read to complete */
+        while ((hyperram_reg_status_read() & (1 << CSR_HYPERRAM_REG_STATUS_READ_DONE_OFFSET)) == 0);
+        return hyperram_reg_rdata_read();
+    }
+
+    /* Configuration and Utility Functions */
+
+    uint16_t hyperram_get_core_latency_setting(uint32_t clk_freq) {
+        /* Raw clock latency settings for the HyperRAM core */
+        if (clk_freq <=  85000000) return 3; /* 3 Clock Latency */
+        if (clk_freq <= 104000000) return 4; /* 4 Clock Latency */
+        if (clk_freq <= 133000000) return 5; /* 5 Clock Latency */
+        if (clk_freq <= 166000000) return 6; /* 6 Clock Latency */
+        if (clk_freq <= 250000000) return 7; /* 7 Clock Latency */
+        return 7; /* Default to highest latency for safety */
+    }
+
+    uint16_t hyperram_get_chip_latency_setting(uint32_t clk_freq) {
+        /* LUT/Translated settings for the HyperRAM chip */
+        if (clk_freq <=  85000000) return 0b1110; /* 3 Clock Latency */
+        if (clk_freq <= 104000000) return 0b1111; /* 4 Clock Latency */
+        if (clk_freq <= 133000000) return 0b0000; /* 5 Clock Latency */
+        if (clk_freq <= 166000000) return 0b0001; /* 6 Clock Latency */
+        if (clk_freq <= 250000000) return 0b0010; /* 7 Clock Latency */
+        return 0b0010; /* Default to highest latency for safety */
+    }
+
+    void hyperram_configure_latency(void) {
+        uint16_t config_reg_0 = 0x8f2f;
+        uint16_t core_latency_setting;
+        uint16_t chip_latency_setting;
+
+        /* Compute Latency settings */
+        core_latency_setting = hyperram_get_core_latency_setting(CONFIG_CLOCK_FREQUENCY/4);
+        chip_latency_setting = hyperram_get_chip_latency_setting(CONFIG_CLOCK_FREQUENCY/4);
+
+        /* Write Latency to HyperRAM Core */
+        printf("HyperRAM Core Latency: %d CK (X1).\n", core_latency_setting);
+        hyperram_config_write(core_latency_setting << CSR_HYPERRAM_CONFIG_LATENCY_OFFSET);
+
+        /* Enable Variable Latency on HyperRAM Chip */
+        config_reg_0 &= ~(0b1 << 3); /* Enable Variable Latency */
+
+        /* Update Latency on HyperRAM Chip */
+        config_reg_0 &= ~(0b1111 << 4);
+        config_reg_0 |= chip_latency_setting << 4;
+
+        /* Write Configuration Register 0 to HyperRAM Chip */
+        hyperram_write_reg(2, config_reg_0);
+
+        /* Read current configuration */
+        config_reg_0 = hyperram_read_reg(2);
+        printf("HyperRAM Configuration Register 0: %08x\n", config_reg_0);
+    }
+    hyperram_configure_latency();
+    printf("\n");
+#endif
+
 #if defined(CSR_ETHMAC_BASE) || defined(MAIN_RAM_BASE) || defined(CSR_SPIFLASH_CORE_BASE)
     printf("--========== \e[1mInitialization\e[0m ============--\n");
 #ifdef CSR_ETHMAC_BASE

test/test_hyperbus.py

@@ -33,7 +33,33 @@ class TestHyperBus(unittest.TestCase):
         pads = Record([("clk_p", 1), ("clk_n", 1), ("cs_n", 1), ("dq", 8), ("rwds", 1)])
         hyperram = HyperRAM(pads)
 
-    def test_hyperram_write(self):
+    def test_hyperram_write_latency_5_2x(self):
+        def fpga_gen(dut):
+            yield from dut.bus.write(0x1234, 0xdeadbeef, sel=0b1001)
+            yield
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--__--__--__--_______"
+            cs_n    = "--________________________________________________________------"
+            dq_oe   = "__------------____________________________________--------______"
+            dq_o    = "002000048d0000000000000000000000000000000000000000deadbeef000000"
+            rwds_oe = "__________________________________________________--------______"
+            rwds_o  = "____________________________________________________----________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                self.assertEqual(c2bool(rwds_o[i]), (yield dut.pads.rwds.o))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=5)
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_write_latency_6_2x(self):
         def fpga_gen(dut):
             yield from dut.bus.write(0x1234, 0xdeadbeef, sel=0b1001)
             yield
@@ -56,10 +82,90 @@ class TestHyperBus(unittest.TestCase):
                 self.assertEqual(c2bool(rwds_o[i]), (yield dut.pads.rwds.o))
                 yield
 
-        dut = HyperRAM(HyperRamPads())
+        dut = HyperRAM(HyperRamPads(), latency=6)
         run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
 
-    def test_hyperram_read(self):
+    def test_hyperram_write_latency_7_2x(self):
+        def fpga_gen(dut):
+            yield from dut.bus.write(0x1234, 0xdeadbeef, sel=0b1001)
+            yield
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--_______"
+            cs_n    = "--________________________________________________________________________------"
+            dq_oe   = "__------------____________________________________________________--------______"
+            dq_o    = "002000048d00000000000000000000000000000000000000000000000000000000deadbeef000000"
+            rwds_oe = "__________________________________________________________________--------______"
+            rwds_o  = "____________________________________________________________________----________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                self.assertEqual(c2bool(rwds_o[i]), (yield dut.pads.rwds.o))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=7)
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_write_latency_7_1x(self):
+        def fpga_gen(dut):
+            yield from dut.bus.write(0x1234, 0xdeadbeef, sel=0b1001)
+            yield
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--_______"
+            cs_n    = "--____________________________________________------"
+            dq_oe   = "__------------________________________--------______"
+            dq_o    = "002000048d0000000000000000000000000000deadbeef000000"
+            rwds_oe = "______________________________________--------______"
+            rwds_o  = "________________________________________----________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                self.assertEqual(c2bool(rwds_o[i]), (yield dut.pads.rwds.o))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=7, latency_mode="variable")
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_read_latency_5_2x(self):
+        def fpga_gen(dut):
+            dat = yield from dut.bus.read(0x1234)
+            self.assertEqual(dat, 0xdeadbeef)
+            dat = yield from dut.bus.read(0x1235)
+            self.assertEqual(dat, 0xcafefade)
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--_"
+            cs_n    = "--________________________________________________________________________"
+            dq_oe   = "__------------____________________________________________________________"
+            dq_o    = "00a000048d0000000000000000000000000000000000000000000000000000000000000000"
+            dq_i    = "00000000000000000000000000000000000000000000000000deadbeefcafefade00000000"
+            rwds_oe = "__________________________________________________________________________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                yield dut.pads.dq.i.eq(int(dq_i[2*(i//2):2*(i//2)+2], 16))
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=5)
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_read_latency_6_2x(self):
         def fpga_gen(dut):
             dat = yield from dut.bus.read(0x1234)
             self.assertEqual(dat, 0xdeadbeef)
@@ -84,5 +190,93 @@ class TestHyperBus(unittest.TestCase):
                 self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
                 yield
 
-        dut = HyperRAM(HyperRamPads())
+        dut = HyperRAM(HyperRamPads(), latency=6)
         run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_read_latency_7_2x(self):
+        def fpga_gen(dut):
+            dat = yield from dut.bus.read(0x1234)
+            self.assertEqual(dat, 0xdeadbeef)
+            dat = yield from dut.bus.read(0x1235)
+            self.assertEqual(dat, 0xcafefade)
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--__--_"
+            cs_n    = "--________________________________________________________________________________________"
+            dq_oe   = "__------------____________________________________________________________________________"
+            dq_o    = "00a000048d00000000000000000000000000000000000000000000000000000000000000000000000000000000"
+            dq_i    = "000000000000000000000000000000000000000000000000000000000000000000deadbeefcafefade00000000"
+            rwds_oe = "__________________________________________________________________________________________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                yield dut.pads.dq.i.eq(int(dq_i[2*(i//2):2*(i//2)+2], 16))
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=7)
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_read_latency_7_1x(self):
+        def fpga_gen(dut):
+            dat = yield from dut.bus.read(0x1234)
+            self.assertEqual(dat, 0xdeadbeef)
+            dat = yield from dut.bus.read(0x1235)
+            self.assertEqual(dat, 0xcafefade)
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--__--__--__--__--__--__--__--__--__--__--__--_"
+            cs_n    = "--____________________________________________________________"
+            dq_oe   = "__------------________________________________________________"
+            dq_o    = "00a000048d0000000000000000000000000000000000000000000000000000"
+            dq_i    = "00000000000000000000000000000000000000deadbeefcafefade00000000"
+            rwds_oe = "______________________________________________________________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                yield dut.pads.dq.i.eq(int(dq_i[2*(i//2):2*(i//2)+2], 16))
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), latency=7, latency_mode="variable")
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")
+
+    def test_hyperram_reg_write(self):
+        def fpga_gen(dut):
+            yield dut.reg_addr.eq(2)
+            yield dut.reg_write_data.eq(0x1234)
+            yield
+            yield dut.reg_write.eq(1)
+            yield
+            yield dut.reg_write.eq(0)
+            for i in range(128):
+                yield
+
+        def hyperram_gen(dut):
+            clk     = "___--__--__--__--___________"
+            cs_n    = "--________________----------"
+            dq_oe   = "__----------------__________"
+            dq_o    = "0060000100000012340000000000"
+            rwds_oe = "____________________________"
+            rwds_o  = "____________________________"
+            for i in range(3):
+                yield
+            for i in range(len(clk)):
+                self.assertEqual(c2bool(clk[i]), (yield dut.pads.clk))
+                self.assertEqual(c2bool(cs_n[i]), (yield dut.pads.cs_n))
+                self.assertEqual(c2bool(dq_oe[i]), (yield dut.pads.dq.oe))
+                self.assertEqual(int(dq_o[2*(i//2):2*(i//2)+2], 16), (yield dut.pads.dq.o))
+                self.assertEqual(c2bool(rwds_oe[i]), (yield dut.pads.rwds.oe))
+                self.assertEqual(c2bool(rwds_o[i]), (yield dut.pads.rwds.o))
+                yield
+
+        dut = HyperRAM(HyperRamPads(), with_csr=False)
+        run_simulation(dut, [fpga_gen(dut), hyperram_gen(dut)], vcd_name="sim.vcd")